Datasets:

Daoze
/

ReviewRebuttal

Dataset card Data Studio Files Files and versions Community

Split (2)

train · 71.9k rows

paper_id stringlengths 9 13	conference_year_track stringclasses 139 values	reviewer_id stringlengths 13 13 ⌀	messages listlengths 5 21 ⌀
9ZE9L4tzO7D	AutoML 2023 Conference	Reviewer_MKuD	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_MKuD, and the following are the paper you need to review:\n```<<9ZE9L4tzO7D>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper proposes a novel method for adjusting the hyperparameters of a loss function to account for heterogeneity across classes in (multi-class) classification tasks. The method works by defining a set of per-class attributes, applying a set of transformations to the attributes and then learning a linear mapping from the transformed attributes to the hyperparameters. They consider two strategies: one that involves a bilevel optimization and one that performs post-hoc optimization using a second function, g, applied to the trained model, f. The authors compare their method against previously proposed methods for addressing class heterogeneity and find that their method performs favorably in terms of minimizing balanced error (as well as and other metrics) using common multi-class image classification benchmark datasets. \noverall_review: Positives\n- Motivation: this work addresses an interesting and relevant problem. I think with some (substantial) improvement, a version of this work would be well-received by the community. \n- Unifying framework: I enjoyed the fact that the authors demonstrated how existing methods can be viewed as instantiations of their more general method. In this regard, I think this work presents a valuable lens with which we can understand how existing methods relate to one another and how we can continue to iterate on them. \n\nNegatives\n- Writing: as noted previously, I found this work to be very poorly written and would encourage the authors to carefully check their grammar/spelling in future iterations of this work. \n- Limited exploration of attribute space: throughout the work, the authors list multiple kinds of class-level attributes that could be incorporated into their method. As such it was very disappointing to see that their experiments only considered three relatively simple attributes.\n- Poorly justified design choices: related to my point above, the choice of attributes considered is not well-motivated. In addition, other aspects of their proposed method are not well justified/explored e.g., the nonlinear feature map,\n- Empirical evaluation: as previously noted, it is unclear which, if any, of the empirical results are statistically significant; Appendix F seems to indicate that the majority of experiments are only repeated 3 times, which is quite low in my opinion. \nreview_summary: My recommendation is to reject this work in its current form due to the weak/unclear empirical evaluation, insufficient detail in the description of the method, and poor writing.\nactions_required_to_increase_overall_recommendation: I am willing to increase my score, potentially to an accept-level score, if the clarity issues are addressed and the questions raised in the \"Technical Quality and Correctness\" section are answered.\nRating: 2 reject, significant issues present\nConfidence: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper\n", "role": "assistant" }, { "content": "Title: Response to Reviewer MKuD\n(1) Re: Why and how to design the nonlinear function map? How to determine $\\beta$ value?\nThese functions are chosen to be polylogarithms or polynomials inspired by [25, 38]. In Appendix E (Appendix D in the old version), we also provide theoretical justification for the nonlinear map. Also, thanks to the adaptability of the method, any potential functions could be used.\n\nThroughout all our experiments, including those conducted on various datasets, utilizing two optimization methods (bilevel and post-hoc), and targeting distinct fairness objectives, we employed a fixed $\\beta$ value and a nonlinear map. Our results demonstrated promising outcomes, suggesting that the result is not sensitive to $\\beta$ value and the nonlinear map. \n\n[25] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314, 2020\n\n[38] Han-Jia Ye, Hong-You Chen, De-Chuan Zhan, and Wei-Lun Chao. Identifying and compensating for feature deviation in imbalanced deep learning. arXiv preprint arXiv:2001.01385, 2020. \n\n\n(2) Clarfication of algorithm and how parameters are optimized\n\n A detailed explanation is presented in ”Response about clarity and algorithm details”. We appreciate your suggestion. As per your recommendation, we have added Appendix B, which contains Figure 4, to demonstrate our entire training process methods and clarify the algorithm.\n\n(3) More runs of experiments. Thank you for your suggestion. We will add more runs for each experiment for the final manuscript.\n\n(3) Re: If the $\\omega_k^{test}$ values normalized after sampling?\n\nYes. It needs to be normalized. We clarify it in Appendix G (Appendix F in the old version).\n\n(4) Re: Computation gains. First, we would like to emphasize that our big O notation focuses on the comparison of the search space with AutoBalance. AutoBalance searches in an $\\mathcal{O}(K)$ space, where each class has its individual hyper-parameter to optimize. Even with the \"group aggregation\" strategy, which shares the same hyper-parameter for classes with a similar number of examples, the search space is only reduced by a constant, still yielding an $\\mathcal{O}(K)$ search space. In contrast, within the CAP framework, $M$ is only related to the number of attributes $n$ and $\|\\mathcal{F}\|$, making it a constant. Therefore, we state in the introduction that its order of complexity belongs to $\\mathcal{O}(1)$.\n\nFurthermore, we would like to clarify that the advantages of our method are not solely attributable to the reduction in the number of parameters. Instead of optimizing the parameters individually, which can result in fragility for tail classes with limited data, our approach facilitates the linking of all classes during training through weight-sharing. Consequently, our approach improves training stability and avoids overfitting compared to other methods.\n\nAdditionally, we extend our algorithm to post-hoc functions and multiple objectives, providing a flexible and efficient solution that enables rapid adaptation to various objectives.\n\n(5) Re: typos/grammatical errors/structural issues. Thanks for your valuable suggestions, we modified them in our new manuscript.\n Global/general Response:\n```\nWe sincerely appreciate all reviewers for taking the time to review our paper and providing valuable reviews. Firstly, we would like to address the concerns regarding the clarity of the paper and the algorithm details. We have made significant revisions to section 3 and added Figure 2 and Figure 4 to provide further clarification.\n\nAs shown in Figure 2, CAP is the overall framework proposed in our paper, with A2H being the core algorithm. \\textbf{A2H} is a meta-strategy that transforms the class-attribute prior knowledge into hyper-parameter ${\\mathbfcal{S}}$ for each class through a trainable matrix $\\mathbf{W}$, forming a training strategy that satisfies the desired fairness objective. The left half of Figure 2 specifically illustrates how our algorithm calculates and trains the weights. In the first stage, we collect class-related information and construct an attribute table of $n \\times K$ dimension. This is a general prior, which is related to the distribution of training data, the training difficulty of each class, and other factors.\n\nSubsequently, by utilizing the constructed attribute table $\\mathbfcal{A}$, the first step of \\textbf{A2H} is to compute a $K\\times M$ Feature Dictionary $\\mathbfcal{D}=\\mathbfcal{F}(\\mathbfcal{A})$ by applying a set of functions $\\mathbfcal{F}$. It is important to emphasize that $M << K$, and in AutoBalance[1], the number of parameters to be trained is directly related to the number of classes $K$. Even with the group aggregation, the scale remains at the $\\mathcal{O}(K)$. However, in \\textbf{CAP}, $M$ is only related to the number of attributes $n$ and $\|\\mathbfcal{F}\|$, making it a constant. Therefore, we state in the introduction that its order of complexity belongs to $\\mathcal{O}(1)$. Then, in the second step, the weight matrix $\\mathbf{W}$ is trained through bilevel or post-hoc methods to construct the hyperparameter $\\mathbfcal{S}$. We further demonstrate our entire training process in Figure 4.\n\n[1] Mingchen Li, Xuechen Zhang, Christos Thrampoulidis, Jiasi Chen, and Samet Oymak. Autobalance: Optimized loss functions for imbalanced data. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. \n\n```\n", "role": "user" }, { "content": "Title: Thank you\nI thank the authors for their response above. Unfortunately, some of my concerns were either unaddressed (e.g., statistical significance) or inadequately addressed (e.g., a promise of additional repetitions) so I am holding my score fixed. While I do believe that this idea has merits and I would be excited to review subsequent iterations of this work, in its current form, I do not believe it is a strong candidate for acceptance. ", "role": "assistant" } ]
9ZE9L4tzO7D	AutoML 2023 Conference	Reviewer_mnow	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_mnow, and the following are the paper you need to review:\n```<<9ZE9L4tzO7D>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper focuses on the supervised classification problem where the number of classes are relatively high, and the classes are significantly imbalanced, and the goal is to enable learning of a fair classifier with explicit focus on the performance of the classifier on the smaller \"tail classes\". To this end, the paper focuses on the use of class-specific hyperparameters that allow for fair classification, and proposes the use of \"class attributes\" -- class-specific data meta-features -- to learn a mapping from the class-attributes to class-specific hyperparameters. This setup is well motivated since the number of hyperparameters to be optimized in the vanilla case can be high for large number of classes, and the use of class-attributes and a corresponding mapping significantly reduces the number of parameters to be estimated. This is especially useful when considering classes with very few examples. The paper demonstrates how some of the existing schemes to address this problem can be obtained as special cases of the proposed class-attribute based scheme. The empirical evaluation demonstrates that the proposed use of class-attributes leads to significant performance improvement on tail classes across multiple datasets, fairness objectives, and class heterogeneity, relative to existing schemes.\n\noverall_review: To the best of my understanding, this paper studies an important problem of fair classification, and presents a fundamental framework for class-fair learning with significant practical impact on tail classes. The paper is well written, with the contributions well articulated, the proposed approach well positioned against existing work. The empirical evaluation clearly motivates each experiment, highlighting different properties of the proposed scheme, and explicitly discusses the conclusion from each of them. It is also relevant to AutoML in the sense that it is trying to configure hyperparameters (such as the hyperparameters of the loss function) and making use of data meta-features. \n\n\nI am unable to identify any significant negative aspect of this paper. There are some minor clarification issues (as I have discussed in the Clarity section, but nothing significant that cannot be easily fixed.\n\nreview_summary: I think this paper studies an important problem and presents a well-motivated intuitive approach with strong empirical performance against existing baselines. I do not recognize any significant negative aspect of this paper. Hence, I am recommending this paper be accepted.\n\nactions_required_to_increase_overall_recommendation: I am already scoring this paper fairly high. I will increase my score if the authors can improve my understanding of this paper and highlight the challenges, if any, in learning the A2H mapping and how they mitigate them.\nRating: 9 strong accept, excellent work\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer mnow \n(1) details of the \"transfer\" between \"bilevel loss function design\" and \"post-hoc optimization\".\n\nFirst, intuitively, in convex optimization, as there is only a global optimum, when we offset the logits in the loss function by $\\Delta$ and $l$, the new logits will also be scaled and offset by $\\frac{1}{\\Delta}$ and $-l$ in comparison to the original logits. In this case, the transferability between loss function design and post-hoc optimization will always be true in convex problems. However, in non-convex optimization, the loss function design and post-hoc adjustment may lead to different local minima. Despite this, our experiments have demonstrated that transferability still exists between loss function design and post-hoc optimization even in non-convex scenarios.\n\nOur main idea of \"transfer\" between \"bilevel loss function design\" and \"post-hoc optimization\" is inspired by logit-adjustment[25], where they provide detailed explanation and justification section 5.1. The logit adjusted loss directly enforce the class prior offset while learning the logits, rather than doing this post-hoc. Modified loss with $\\iota$ is equivalent to using a scorer of the form $g_y(x)=\\Delta f_y(x)+\\iota$. We thus have $argmax_{y \\in[K]} f_y(x)=argmax_{y \\in[K]} \\frac{1}{\\Delta}(g_y(x)-\\iota)$. Consequently, one can equivalently view learning with this loss as learning a standard scorer g(x), and post-hoc adjusting its logits to make a prediction. Therefore, we utilize opposite signs.\n\n[25] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314, 2020\n\n(2) Detail of generating attributes like \"class conditional error\"\n\nWe have included Fig.~2 to provide a detailed explanation of the algorithm. The class attributes are calculated before training. For instance, test-time class weights are determined based on the target fairness objectives and do not require any optimization. The class-conditional error, utilized as $\\mathcal{A}_{DIFF}$, can be computed using the classification outcome of any pre-trained model. In our experiments, considering the computational cost, we computed the class-conditional error after pre-training for post-hoc optimization and at the end of the warm-up phase for bilevel training.\n\n(3) How the proposed framework can work with different fairness objectives?\n\nFirst, in the bilevel formulation, $\\mathcal{L}_{val}$ corresponds to the fairness objectives. In our current experiments, we only optimized different fairness objectives using post-hoc optimization, with the loss function serving as the fairness objectives. This is because post-hoc optimization is a more flexible and efficient solution that enables rapid adaptation to various objectives, which aligns with our goal for quick computation. The hyper-parameters obtained from post-hoc optimization also have the potential for transferring to loss function design. We are happy to run bilevel optimization with multiple objectives if necessary. In this case, the validation loss function should be set to different fairness objectives.\n\nWe further add Fig. 4 in Appendix B to demonstrate our entire training process.\n Global/general Response:\n```\nWe sincerely appreciate all reviewers for taking the time to review our paper and providing valuable reviews. Firstly, we would like to address the concerns regarding the clarity of the paper and the algorithm details. We have made significant revisions to section 3 and added Figure 2 and Figure 4 to provide further clarification.\n\nAs shown in Figure 2, CAP is the overall framework proposed in our paper, with A2H being the core algorithm. \\textbf{A2H} is a meta-strategy that transforms the class-attribute prior knowledge into hyper-parameter ${\\mathbfcal{S}}$ for each class through a trainable matrix $\\mathbf{W}$, forming a training strategy that satisfies the desired fairness objective. The left half of Figure 2 specifically illustrates how our algorithm calculates and trains the weights. In the first stage, we collect class-related information and construct an attribute table of $n \\times K$ dimension. This is a general prior, which is related to the distribution of training data, the training difficulty of each class, and other factors.\n\nSubsequently, by utilizing the constructed attribute table $\\mathbfcal{A}$, the first step of \\textbf{A2H} is to compute a $K\\times M$ Feature Dictionary $\\mathbfcal{D}=\\mathbfcal{F}(\\mathbfcal{A})$ by applying a set of functions $\\mathbfcal{F}$. It is important to emphasize that $M << K$, and in AutoBalance[1], the number of parameters to be trained is directly related to the number of classes $K$. Even with the group aggregation, the scale remains at the $\\mathcal{O}(K)$. However, in \\textbf{CAP}, $M$ is only related to the number of attributes $n$ and $\|\\mathbfcal{F}\|$, making it a constant. Therefore, we state in the introduction that its order of complexity belongs to $\\mathcal{O}(1)$. Then, in the second step, the weight matrix $\\mathbf{W}$ is trained through bilevel or post-hoc methods to construct the hyperparameter $\\mathbfcal{S}$. We further demonstrate our entire training process in Figure 4.\n\n[1] Mingchen Li, Xuechen Zhang, Christos Thrampoulidis, Jiasi Chen, and Samet Oymak. Autobalance: Optimized loss functions for imbalanced data. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. \n\n```\n", "role": "user" }, { "content": "I appreciate the authors' reply. The response addresses most of my concerns. I continue to recommend this work for acceptance.", "role": "assistant" } ]
9ZE9L4tzO7D	AutoML 2023 Conference	Reviewer_2WMX	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_2WMX, and the following are the paper you need to review:\n```<<9ZE9L4tzO7D>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper proposed new meta-strategy class-attribute priors (CAP) that support improving fairness objectives while adapting to the different heterogeneities of the individual classes. Results show that using CAP can lead to an improvement in fairness objectives performance. \noverall_review: The paper presents a compelling argument to use the CAP meta-strategy to address differences in heterogeneities across classes when optimizing for fairness objectives. \n\nThere are minor gaps that can help strengthen the paper further. For example, In section 1.1, the authors discuss how the bilevel problem is time-consuming. However, it needs to be clarified if the proposed method is time-consuming. If yes, by what scale? It is vital to connect the benefits of CAP back to the problem with prior techniques.\n\nreview_summary: Class label attributes can influence the model performance and fairness objective when working with real-world data. While there are several methods out there in the field, there is no one method that solves them all. Different techniques can help in different instances. So I believe this CAP strategy can bolster this filed of research.\nactions_required_to_increase_overall_recommendation: A couple of statements attempt to connect the benefits of CAP back to the problem with prior techniques.\nRating: 8 accept, good paper\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer 2WMX\n(1) Re: Validation set distribution questions.\nThe validation set was chosen in accordance with prior literature. During the search phase for bilevel CAP and post-hoc transferability experiments (Sec. 4.1), we split the training set into 80\\% training and 20\\% validation to obtain the optimal loss function design. For all other post-hoc experiments (Sec. 4.2 and 4.3), we follow the setup of [11,25] by training a model on the entire training dataset as the pre-train model and optimizing a logit adjustment on a balanced validation dataset. The test dataset was left untouched as it would be impractical to optimize model parameters using this dataset. Since all experiments are conducted on the same validation set, our comparison is fair across all baseline methods and CAP. We further provided post-hoc experiments with imbalanced validation in Table 6.\n\n[11] Moritz Hardt, Eric Price, and Nathan Srebro. Equality of opportunity in supervised learning. arXiv preprint arXiv:1610.02413, 2016.\n[25] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314, 2020.\n\n(2) Re: If the proposed method is also time-consuming?\nBilevel optimization can be time-consuming, primarily due to the computation of the Hessian-vector-product. However, this can be bypassed by employing post-hoc logit adjustment. In this work, we focus on the development of a loss function through bilevel optimization and post-hoc logit adjustment. Additionally, we explore the interconnection between bilevel CAP and post-hoc optimization by transferring obtained hyper-parameters. Employing post-hoc logit adjustment or the construction of a loss function via post-hoc optimization has the potential to significantly enhance the efficiency of the process by approximately 5x. The details of transferability are shown in Appendix C.\nAs reported in reference [23], bilevel optimization can be unstable and difficult to optimize, particularly when optimizing the inner loss function. The authors propose a lengthy warm-up phase, which can be time-consuming. In Appendix B, we discuss how the use of Class-Attribute Prioritization (CAP) can improve training stability and eliminate the need for a prolonged warm-up phase. Our proposed $\\text{CAP}_{\\text{Bilevel}}$ method, which incorporates attributes, not only leverages prior knowledge but also improves optimization stability by simultaneously updating weights for all classes through the dictionary design. CAP with bilevel optimization achieves a classification error rate of 20.16 on CIFAR10-LT and 56.55 on CIFAR100-LT with only five epochs of warm-up, representing an improvement in both computational efficiency and test performance.\n\n[23] Mingchen Li, Xuechen Zhang, Christos Thrampoulidis, Jiasi Chen, and Samet Oymak. Autobalance: Optimized loss functions for imbalanced data. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. \n Global/general Response:\n```\nWe sincerely appreciate all reviewers for taking the time to review our paper and providing valuable reviews. Firstly, we would like to address the concerns regarding the clarity of the paper and the algorithm details. We have made significant revisions to section 3 and added Figure 2 and Figure 4 to provide further clarification.\n\nAs shown in Figure 2, CAP is the overall framework proposed in our paper, with A2H being the core algorithm. \\textbf{A2H} is a meta-strategy that transforms the class-attribute prior knowledge into hyper-parameter ${\\mathbfcal{S}}$ for each class through a trainable matrix $\\mathbf{W}$, forming a training strategy that satisfies the desired fairness objective. The left half of Figure 2 specifically illustrates how our algorithm calculates and trains the weights. In the first stage, we collect class-related information and construct an attribute table of $n \\times K$ dimension. This is a general prior, which is related to the distribution of training data, the training difficulty of each class, and other factors.\n\nSubsequently, by utilizing the constructed attribute table $\\mathbfcal{A}$, the first step of \\textbf{A2H} is to compute a $K\\times M$ Feature Dictionary $\\mathbfcal{D}=\\mathbfcal{F}(\\mathbfcal{A})$ by applying a set of functions $\\mathbfcal{F}$. It is important to emphasize that $M << K$, and in AutoBalance[1], the number of parameters to be trained is directly related to the number of classes $K$. Even with the group aggregation, the scale remains at the $\\mathcal{O}(K)$. However, in \\textbf{CAP}, $M$ is only related to the number of attributes $n$ and $\|\\mathbfcal{F}\|$, making it a constant. Therefore, we state in the introduction that its order of complexity belongs to $\\mathcal{O}(1)$. Then, in the second step, the weight matrix $\\mathbf{W}$ is trained through bilevel or post-hoc methods to construct the hyperparameter $\\mathbfcal{S}$. We further demonstrate our entire training process in Figure 4.\n\n[1] Mingchen Li, Xuechen Zhang, Christos Thrampoulidis, Jiasi Chen, and Samet Oymak. Autobalance: Optimized loss functions for imbalanced data. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. \n\n```\n", "role": "user" }, { "content": "I appreciate the authors' efforts in addressing my concerns. The response covers nearly all of the issues I raised. I will continue to recommend accepting this work.", "role": "assistant" } ]
A9FUg5vrw7Y6	AutoML 2023 Conference	Reviewer_gM84	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_gM84, and the following are the paper you need to review:\n```<<A9FUg5vrw7Y6>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions:  \n\nThe authors propose a novel Poisson process-based scheme for rank-based Bayesian optimization and investigate its performance on HPO and NAS tasks.\n\n \n\noverall_review:  \n\n__MAJOR POINTS__\n\n \n\n1. In terms of the MLP used to approximate the scaled rate parameter of the inhomogeneous Poisson process, a crucial question is whether the performance of PoPBO is robust to the training scheme/architecture of the MLP. Do the authors believe that a single fixed MLP architecture and training regime will remain performant across a broad range of problems or will the architecture have to be tuned? If so how will this tuning occur on each iteration of Bayesian optimization?\n\n2. It does not seem as though HEBO is run on 2-D Branin in Figure 2a?\n\n3. Is there a reason why HEBO and PPBO are not run on the HPO-Bench and NAS-Bench benchmarks?\n\n \n\n__MINOR POINTS__\n\n \n\n1. In section 1, when discussing the literature on preference BO, it may be worth referencing the recently introduced BOPE algorithm [2].\n\n2. It would be great if ADAM was cited [3] given that it is used although one might think the paper has enough citations already :)\n\n3. For the empirical evaluation, the standard error could be plotted in place of the standard deviation, this may make the trend in the empirical results visually clearer due to non-overlapping error bars.\n\n4. In Figure 3, what are the units of time cost? Are the reported values wall-clock times or CPU times?\n\n5. In Table 2, it might be misleading to imply that random search was invented in 2012!\n\n6. In Figure 10, the plots do not seem to verify the claims made by the authors that PoPBO outperforms GP-BO?\n\n7. In the algorithmic description the authors say that they optimize the MLP with SGD, however the experimental settings say that they use ADAM?\n\n8. Line 256, it would be great if the benchmarks on which BOHB and BOHAMIANN are state-of-the-art is provided if they are described as such. (This is just a matter of taste relating to the literal definition of state-of-the-art!)\n\n9. Where are the error bands for Figure 5?\n\n \n\n__REFERENCES__\n\n \n\n[1] Cowen-Rivers et al., [HEBO: Pushing the Limits of Sample-Efficient Hyperparameter Optimization](https://www.jair.org/index.php/jair/article/view/13643), Journal of Artificial Intelligence Research, 2022.\n\n[2] Lin et al., [Bayesian Optimization with Preference Exploration](https://proceedings.mlr.press/v151/jerry-lin22a.html), AISTATS, 2022.\n\n[3] Kingma and Ba, [ADAM: An Adaptive Optimizer](https://arxiv.org/pdf/1412.6980.pdf), ICLR 2014.\n\n \nreview_summary:  \n\nThe authors introduce a novel, rank-based Bayesian optimization scheme. The research arc is an important one for AutoML and the empirical evaluation (modulo the points/questions raised above) is solid. Furthermore, the paper is well-written. As such, I recommend acceptance with the scope for further increasing my score to strong accept if the outstanding points are addressed.\n\n \nactions_required_to_increase_overall_recommendation:  \n\nIf the authors can address the points raised with regard to the empirical evaluation i.e. offer their thoughts on schemes for stabilizing MLP training as components of Bayesian optimization loops, supply a working link to the codebase, and explain why the PPBO and HEBO baselines are absent in some of the experiments I will be happy to increase my score from weak accept to strong accept.\n\n \nRating: 8 accept, good paper\nConfidence: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\n", "role": "assistant" }, { "content": "Title: Response to Reviewer gM84\nThanks for your valuable comments. We answer your questions as follows in the hope of addressing your concerns.\n\nQ1: Lines 279/280, \"However, there exist real scenarios where the noise is too large to disrupt the ranking of observations, and we would like to leave it as our future work\" Do the authors mean, \"large enough to disrupt the ranking of observations\"?\n\nA1: Yes, we will leave the situations where the noise is large enough to disrupt the ranking of observations as our future work.\n\nQ2: Whether the performance of PoPBO is robust to the training scheme/architecture of the MLP. Do the authors believe that a single fixed MLP architecture and training regime will remain performant across a\nbroad range of problems or will the architecture have to be tuned? If so how will this tuning occur on each\niteration of Bayesian optimization?\n\nA2: \nWe adopt an MLP with 3 hidden layers and 128 nodes, which is commonly used in the prior works [shenlaoshi].\nMoreover, we use the same settings among all experiments displayed in our paper and run each experiment repeatedly to eliminate the effects of initialization. The experimental results indicate that the MLP performs well on both simulated and real-world benchmarks, including HPO and NAS. The tuning of the MLP occurs on each iteration.\n\nQ3: It does not seem as though HEBO is run on 2-D Branin in Figure 2a?\n\nA1: Thanks for pointing this out, and we will add the results of HEBO to Fig. 2a in the new version.\n\nQ4: Is there a reason why HEBO and PPBO are not run on the HPO-Bench and NAS-Bench benchmarks?\n\nA4: We implemented GP-BO, PoPBO, SMAC, and TPE in our own framework and built the NAS-Bench-201 and HPO-Bench in our AutoML system's knowledge base. However, we have not integrated the codes of HEBO and PPBO into our framework, and thus conducting the experiment of HEBO and PPBO on HPO-Bnech and NAS-Bench can be difficult. The results of HEBO and PPBO are obtained by running their open-sourced codes. We will integrate the codes into our new version as soon as possible.\n\nQ5: For the empirical evaluation, the standard error could be plotted in place of the standard deviation, this may make the trend in the empirical results visually clearer due to non-overlapping error bars.\n\nA5: Thanks for your kind suggestion, and we will refine it in our new version.\n\n\n\nQ6: In Figure 3, what are the units of time cost? Are the reported values wall-clock times or CPU times?\n\nA6: The units are wall-clock times. Nevertheless, in our experiment, the evaluation of candidates can be negligible compared with the training of the surrogate model and the optimization of acquisition. Therefore, the wall-clock time has little difference from the CPU time.\n\n\n\nQ7: In Figure 10, the plots do not seem to verify the claims made by the authors that PoPBO outperforms GP-BO?\n\nA7: Our PopBO has better ultimate performance under large noises. Specifically, our PoPBO can be more robust and performs better on the Hartmann-6d function under large noises (as shown in Fig. 10c). Moreover, PoPBO performs consistently better on the Rosenbrock-6d function in all settings (as shown in Fig. 10g-Fig. 10i). As for Branin-2d function, which is a relatively simple optimization benchmark, PoPBO shows a slower convergence but a better ultimate performance under large noises (as shown in Fig. 10f).\n\n\n\nQ8: In the algorithmic description the authors say that they optimize the MLP with SGD, however, the experimental settings say that they use ADAM.\n\nA8: Thanks for the reminder, we use ADAM to train our MLP model and we have refined it in our new version.\n\n\n\nQ9: Where are the error bands for Figure 5?\n\n A9: We find that the standard deviations overlap too much on Rosenbrock-6d under the log scale, which motivates us to remove it in Figure 5.\n\n\n\nQ10: For the typo and other suggestions.**\n\nA10: Thanks for these kind suggestions. We will refine these tips in the new version.\n\n\n Global/general Response:\n```\nWe thank the valuable comments of all the reviewers, and we refine the submission according to the feedback. The modified parts are colored (magenta for Reviewer gM84, red for Reviewer j7Na, cyan for Reviewer 22Se, blue for Reviewer 46kD, and green for Reviewer itsg). \n\n```\n", "role": "user" }, { "content": "Title: Thanks for the Response and Changes\nMany thanks to the authors for their response and changes. I am happy to raise my score following the additional baselines implemented and clarifications. \n\nIn terms of training the MLP on each iteration of BO do the authors have a stopping criterion for training?", "role": "assistant" }, { "content": "Title: Response to Reviewer gM84\nThanks for your positive response. We train the MLP on each iteration of BO for 100 epochs without an criterion of early stop due to the insufficient observations in the BO context.", "role": "user" }, { "content": "I am delighted to have engaged in the above conversation with the author, which has added considerable value to this work.", "role": "assistant" } ]
A9FUg5vrw7Y6	AutoML 2023 Conference	Reviewer_j7Na	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_j7Na, and the following are the paper you need to review:\n```<<A9FUg5vrw7Y6>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The authors present a novel approach towards preferential Bayesian Optimization dubbed Poisson Process Bayesian Optimization (PoPBO). It is applicable for \"relative\" response surfaces, i.e. situations in which preferetial comparisons between candidates are available but not necissarily the concrete numerical values of the objective function under consideration. It relies on the assumption of a non-homogenous Poisson process. The authors introduce two acquisition functions, namely the Rectified Lower Confidence Bound (R-LCB) and Expected Ranking Improvement (ERI).\noverall_review: Positive:\n- The paper addresses an interesting problem and proposes a novel approach towards Bayesian Optimization with preference feedback. This potentially has a broad impact on AutoML as well as BO in general.\n- The approach is empirically evaluated both on simulated and real-world benchmarks, including HPO and NAS. It shows competitive performance in terms of minimum regret with the considered baselines. The experiments seem to be conducted correctly and the results back the authors claims.\n- In addition to the competitive performance, the Poisson process based BO surrogate model procedure only has a computational complexity of O(N²) for fitting, unlike the GP which has cubic runtime. \n\nNegative:\n- The paper is very convoluted and packed with formal details. It is very difficult to comprehend it.\nreview_summary: The paper addresses an interesting problem, i.e. Bayesian Optimization when only preference feedback can be acquired. While it is difficult to comprehend the paper due to it's convoluted formal explanations, the derived approach seems sound to me. The experimental evaluation shows competitive performance both on simulated as well as real-world benchmarks.\nactions_required_to_increase_overall_recommendation: - The first sentence of the abstract should not list SMAC as a surrogate model but rather the random forest used by SMAC\n- There are some linguistical problems:\n - Capitalization, e.g. \"Random Forest\" vs. \"random forests\"\n - Punctuation\n - Missing use of indefinite and definite articles\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper\n", "role": "assistant" }, { "content": "Title: Response to Reviewer j7Na\nThanks for your valuable comments. We answer your questions as follows in the hope of addressing your concerns.\n\n\nQ1: The first sentence of the abstract should not list SMAC as a surrogate model but rather the random forest used by SMAC. Some typos\n\nA1: Thanks for the suggestion, we will refine it in our next version.\n\n Global/general Response:\n```\nWe thank the valuable comments of all the reviewers, and we refine the submission according to the feedback. The modified parts are colored (magenta for Reviewer gM84, red for Reviewer j7Na, cyan for Reviewer 22Se, blue for Reviewer 46kD, and green for Reviewer itsg). \n\n```\n", "role": "user" }, { "content": "I value the authors' response. The response covers nearly all of the issues I raised. I continue to recommend this work positively.", "role": "assistant" } ]
A9FUg5vrw7Y6	AutoML 2023 Conference	Reviewer_22Se	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_22Se, and the following are the paper you need to review:\n```<<A9FUg5vrw7Y6>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper falls into the scope of preferential Bayesian optimization and it attempts to capture the relative performances instead of absolute performances. It describes a new Bayesian optimization framework called Poisson Process Bayesian Optimization (PoPBO). Authors proposed to directly fit the ranking based response surface based on Poisson process. This is unlike existing work, which tries to fit the absolute response surface first and then obtain the relative response surface. The proposed method is robust to noise. Besides, this paper also proposed two acquisition functions named R-LCB and ERI. These acquisition functions are derived from the vanilla LCB and EI for better exploitation-exploration trade-off. The computational complexity of PoPBO is much lower than that of GP-BO. The method outperforms many prior BO methods on both simulated functions and real-world benchmarks, including HPO and NAS.\n\n\noverall_review: Strengths:\n\n- The paper addresses an important problem: black-box optimization with only relative rankings of candidates. Absolute values can be hard to obtain in some cases. Besides, building absolute surface is not robust of noise. \n- This paper discuss limitations of existing work in detail and proposed a framework to mitigate these limitations. \n\nWeaknesses:\n- The effect of parameter q in eq(7) is not discussed in detail. From fig.5, the q parameter has a huge influence to the final performance. In practice, we cannot try a list of different q parameters as the underline functions is expensive. That means, we will never be able to inspect the influence of parameter q. How could one set it correctly? \n- Continuing the above point. The influence of q is only discussed for a single synthetic function. Given this parameter is so important, without discussing it extensively, the reviewer will not be able to increase the score. \nreview_summary: This paper addresses an important problem: black-box optimization with only relative rankings of candidates. The proposed method outperforms many prior BO methods on both simulated functions and real-world benchmarks, including HPO and NAS. This means that the PoPBO framework can be used to optimize hyperparameters and neural architecture search in AutoML more efficiently than previous methods. \n\nHowever the authors did not discuss the parameter q in R-LCB, this makes it difficult to justify the method's influence as it seems the performance is very sensitive to this parameter. \nactions_required_to_increase_overall_recommendation: 1. perform statistical test for the experimental results. It is hard to tell whether the proposed method is significantly better than other methods through mean plot. I would also suggest using median plot. \n2. discuss influence of parameter q for different functions.\nRating: 5 marginally below the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer 22Se\nThanks for your valuable comments. We answer your questions as follows in the hope of addressing your concerns.\n\n\nQ1: How is the MLP trained? What are the inputs and what are the outputs?\n\nA1: We train the MLP model with Adam optimizer and the joint Poisson likelihood in Eq. 4. The configuration $x$ is the input of the MLP, and its average ranking rate $\\lambda_\\xi(x)$ is the output.\n\n\n\nQ2: The effect of parameter q in eq(7) is not discussed in detail. From Fig. 5, the q parameter has a huge influence on the final performance. In practice, we cannot try a list of different q parameters as the underline functions are expensive. That means we will never be able to inspect the influence of parameter q. How could one set it correctly?\n\nA2: Actually, we keep the same set of q across all experiments according to our ablation study. The experimental results show that this setting can perform well on a variety of tasks and we think it needs no more adjustment. Moreover, following the same ablation settings on Rosenbrock, we add a supplementary experiment on NASBench201 to solve your concerns. The results of this supplementary experiment have been published in Appendix C.7 in the new version. \n\n\nQ3: The influence of q is only discussed for a single synthetic function.\n\nA3: As mentioned above, we also conduct the ablation study on NASBench. The results show that our settings of q (0.4 for ERI and 0.6 for R-LCB) perform still well on the real-world benchmark.\n\n Global/general Response:\n```\nWe thank the valuable comments of all the reviewers, and we refine the submission according to the feedback. The modified parts are colored (magenta for Reviewer gM84, red for Reviewer j7Na, cyan for Reviewer 22Se, blue for Reviewer 46kD, and green for Reviewer itsg). \n\n```\n", "role": "user" }, { "content": "I acknowledge the authors' response. They have addressed most of the issues I raised, and I will likely raise the rating.", "role": "assistant" } ]
A9FUg5vrw7Y6	AutoML 2023 Conference	Reviewer_46kD	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_46kD, and the following are the paper you need to review:\n```<<A9FUg5vrw7Y6>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This work suggests a ranking-based surrogate model using Poisson processes. Instead of Gaussian processes or other surrogate models, it determines relative ordering of historical queries and make use of such ordering in next query selection. The authors provide a new form of acquisition functions for the ranking-based surrogate model. Finally the authors demonstrate experimental results with baseline methods as well as the proposed method.\noverall_review: I think that this paper is quite interesting and novel. However, I want to leave some comments.\n\n* Could you elaborate on the contents in Section 3.2? How does it estimate ranking and how does it work?\n\n* Following the above questions, the authors should revise Section 3 of the paper.\n\n* I am curious why multi-layer perceptron is used to estimate $\\lambda_{\\xi}$. Did you try other models?\n\n* Is there any guarantee that machine learning models are capable of estimating $\\lambda_{\\xi}$?\n\n* Moreover, I would like to hear the implication of Equation (3), (4), and (5).\n\n* Could you add a description of why $\\sigma(x) = \\sqrt{\\mu(x)}$?\n\n* I think that you missed the recent studies of the line of TPEs (i.e., likelihood-free Bayesian optimization). You need to refer to [1, 2].\n\n* I think you need to be consistent for a scale of y-axis. Some figures have a logarithmic scale, but the others are not.\n\n[1] Tiao, Louis C., et al. \"BORE: Bayesian optimization by density-ratio estimation.\" International Conference on Machine Learning, 2021.\n\n[2] Song, Jiaming, et al. \"A general recipe for likelihood-free Bayesian optimization.\" International Conference on Machine Learning, 2022.\nreview_summary: Please see the text boxes above.\nactions_required_to_increase_overall_recommendation: I am now leaning towards a positive score. But, if my questions will be resolved, I will increase my score more.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer 46kD\nThanks for your valuable comments. We answer your questions as follows in the hope of addressing your concerns.\n\nQ1: How does it estimate ranking and how does it work?\n\nA1: We would like to capture the ranking of each candidate $x$ over a domain $S$. However, $S$ usually contains a large number of candidates making it intractable to obtain the true ranking. Therefore, we sample a set $\\hat{S}$ from the search space $X$ and attempt to capture the ranking of $x$ over $S\\cap\\hat{S}$, denoted as $\\hat{R}_x(S)$, by a truncated non-homogeneous Poisson process. \nTherefore, the probability of a configuration $x$ ranking as $k$ over $\\hat{S}$ can be estimated by Eq. 3.\n\nQ2: Implication of Equations (3), (4), and (5).\n\nA2: Eq. 3 is the probability density function of a truncated non-homogeneous Poisson distribution, representing the probability of $x$ ranking as $k$ over the set $\\hat{S}$.\n\nEq.4 is the log-likelihood of the ranking of a set of samples. Given a set $hat{S}$, we can get the ranking of each sample over the whole set. Suppose the ranking of $x_j$ is $\\hat{k}_{x_j}$, then the log-likelihood of the ranking of all samples can be approximated as Eq. 4, i.e.:\n\n$$\n\\begin{align}\nlog L(\\hat{K}\|\\hat{S}; \\theta) &= log \\prod_{j=1}^N P(\\hat{R}_{x_j}=\\hat{k}_{x_j}\|x, \\hat{S}).\n\\end{align}\n$$\n\nEq. 5 is the predictive distribution of a new sample $x^$ in the inference phase. \n\n\n\nQ3: Why multi-layer perceptron is used to estimate $\\lambda_\\xi$. Did you try other models? Is there any guarantee that machine learning models are capable of estimating $\\lambda_\\xi$?\n\nA3: We do not try other models. In this work, we would like to learn the function $\\lambda_\\xi$ and utilize MLP because many prior works demonstrate that MLP can fit complex functions. We would like to point out that MLP is not the only way to estimate $\\lambda_\\xi$, and our method has no strong dependency on MLP. Therefore, demonstrating that MLPs are capable of estimating $\\lambda_\\xi$ is out of the scope of this work.\n\n\nQ4: why $\\sigma(x) = \\sqrt(\\mu(x))$?\n\nA3: It is the property of Poisson distribution, i.e., the variance equals the expectation.\n\nQ5: Lack of recent studies of the line of TPEs.\n%(i.e., likelihood-free Bayesian optimization)\n\nA5: Thanks for kindly suggesting the two TPE methods, we will add them to our related work. \n\nQ6: Inconsistency for a scale of the y-axis.*\n\nA6: We adjust the scale of the y-axis for a clear illustration. In Fig.2a (Branin) and Fig.2b (Hartmann), the ranges of the regrets are more suitable to the normal scale, while others are suitable to the log scale.\n\n Global/general Response:\n```\nWe thank the valuable comments of all the reviewers, and we refine the submission according to the feedback. The modified parts are colored (magenta for Reviewer gM84, red for Reviewer j7Na, cyan for Reviewer 22Se, blue for Reviewer 46kD, and green for Reviewer itsg). \n\n```\n", "role": "user" }, { "content": "Thank you for the detailed response from the authors. It satisfactorily addresses the major issues I had. I keep my recommendation for this paper to be accepted.", "role": "assistant" } ]
A9FUg5vrw7Y6	AutoML 2023 Conference	Reviewer_itsg	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_itsg, and the following are the paper you need to review:\n```<<A9FUg5vrw7Y6>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: - A ranking-based surrogate model in the form of the Poisson processes, to enable BO using only the relative ranking of observations, as opposed to their (possibly unattainable) values. The Poisson processes model is approximated through an MLP \n- The adaptation of UCB and EI, and their gradients, to the proposed ranking-based model\n- A resulting BO framework that is computationally cheaper $(\\mathcal{O}(n^2)$ than the traditional $(\\mathcal{O}(n^3)$\n- Empirical results\noverall_review: _Since the concerns are outlined in greater detail above, this section will be shorter._\n\n\n### Pros\n- __Model is novel:__ Poisson processes are a new type of model that inherently supports ranking-based data\n- __The approach is straight-forward__: The approach is relatively simple and generally well-communicated.\n\n### Cons\n- __Acquisition functions have multiple hyperparameters__ though these should debatably be out-of-the-box (and thus not a significant contribution). Adding hyperparameters is seldom a good thing, and EI/UCB should not need additional ones given their principled foundations.\n- __Introducing an MLP as a novel model type is not very novel.__ The nuances of the model training is not further explored in the paper.\n- __The quality of the proposed model is not thoroughly assessed__ in terms of prediction ability, e.g. validation set log-likelihood. The first plot indicates that it predicts well, but that is all there is to assess the quality of the model's predictive capabilities. \n- __The noise evaluation is lackluster__ and does not benchmark against the proper methods. GP with sufficient noise modelling (what are the noise assumptions/prior in the GP?) and Noisy EI.\n- __Preferential BO literature is not benchmarked against__, at all. This should be the most relevant comparison of all.\n\nAs such, this paper should not be accepted in its current format. Some premises (i.e. that GPs are not robust to noise) are unconventional, and something I personally disagree with. Discounting the rectified acquisition functions, the novelty is not very high. Moreover, there are too many outstanding questions with regard to the performance of the approach relative to its competition across several relevant dimensions, as outlined above.\n\nreview_summary: The paper presents Poisson processes, a novel tool for ranking-based Bayesian optimization. \n\n### Pros\n- Poisson processes are a new type of model and a good idea\n- The approach is straight-forward and generally well-communicated\n\n### Cons\n- Acquisition functions should be out-of-the-box\n- Model is not thoroughly assessed\n- The noise robustness theory and evaluation is lackluster \n- It does not benchmark against the relevant Preferential BO literature\n- Introducing an MLP as a novel model type is, arguably, not very novel.\n\nThis paper should not be accepted in its current format. The novelty is not very high and there are severe outstanding questions regarding performance.\nactions_required_to_increase_overall_recommendation: I will consider raising my scores if the concerns regarding the experimental setup and lack of relevant benchmarks are addressed. These have all been outlined above.\nRating: 2 reject, significant issues present\nConfidence: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\n", "role": "assistant" }, { "content": "Title: Response to Reviewer itsg\nThanks for your valuable comments. We answer your questions as follows in the hope of addressing your concerns.\n\n\nQ1: The robustness analysis is rudimentary, and is not unique to the Poisson Process - the exact same argument can be made for the GP.\n\nA1: We are afraid that it is a misunderstanding of our work. \nOur point is that relative responses can be more robust to noises than absolute responses, which is also shown in many prior works[1,2]. Motivated by this, we attempt to utilize the Poisson process to directly capture the ranking-based metric, which is one of the commonly used relative responses. Moreover, we would like to emphasize that the robustness of our PoPBO comes from the robust ranking-based metric, rather than the Poisson process.\n\n[1] Nguyen, Q. P., Tay, S., Low, B. K. H., and Jaillet, P. (2021). Top-k ranking Bayesian optimization. In AAAI\n\n[2] Salinas, D., Shen, H., and Perrone, V. (2020). A quantile-based approach for hyperparameter transfer learning. In ICML.\n\n\nQ2: I disagree with the notion that the GP is not robust to noise. I do not believe that the first figure and robustness analysis is sufficient evidence to claim that Poisson processes are more noise-robust. This is expanded on below and in the clarity section.\n\nA2: We never claim that PP is more robust to noise than GP, instead, Fig.1 aims to show that ranking-based metric is more robust to noise than absolute responses. We choose GP to capture the absolute response surface which is widely used and well-performed in BO. \nThe reason we introduce the Poisson process as a BO surrogate model lies in its natural property to directly capture the ranking-based response surface, rather than its robustness to noise.\n\n\n\nQ3: The initialization scheme is unconventional and unmotivated large. it seems like the proposed method needs more data than other methods (since MLP surrogates are known to be data-hungry).\n\nA3: It should be noticed that all methods have the same initialization scheme, i.e., other methods and our PoPBO all have 30 initial samples in the Rosenbrock-6d simulated benchmark. We also conduct a supplementary experiment (in appendix C.9) that sets only 7 initial points for the 6-d Rosenbrock benchmark to address your concerns. The results show that our PoPBO still outperforms other methods.\n\nQ4: Lack of comparison with PBO\n\nA4: We can not find the open-sourced code about PBO during the period of our experiments. We have implemented a version of PBO according to the paper. However, the version does not work on these benchmarks due to a lack of implementation details. We have sent a mail to the author of PBO and we will finish this experiment and add the results to our new version as soon as we receive the response.\n\nQ5: Why LCB and EI should be rectified\n\nA5: We analyze that LCB and EI are designed for the absolute value-based surrogate models, and it can be inappropriate to directly transfer them to the PoPBO, which is analyzed in the manuscript (Lines 183-185). To this end, we propose simple but effective strategies, i.e., the rectified LCB and EI, to accommodate the PoPBO. The experimental results demonstrate the effectiveness.\n\nQ6: Why EI makes sense in a ranking-based context?\n\nA6: As same as in an absolute-based context, ERI demonstrates the expected improvement of the ranking of a configuration over the worst tolerable ranking. Suppose the worst tolerable ranking is $K_m$, then ERI represents the expected improvement of a configuration ranking ahead of $K_m$.\n\nQ7: Noisy EI should be compared.\n\nA7: Thanks for your kind suggestion. We will try to conduct the experiments and update the results in our future version. However, we are afraid that comparing with Noisy EI is out of the scope of this work. As mentioned in A1 and A2, our motivation is that the relative response surface is more robust to noise and thus we introduce PoPBO to directly capture the relative response surface. We don't consider any information about noise during the acquisition procedure.\n\nQ8: Lack of runtime analysis.\n\nA8: The runtime analysis is detailed in Sec. 5.1 (the paragraph Computational Cost) and Fig. 3.\n\nQ9: The setting of GP in the experiments in Fig. 1.\n\nA9: We fit GP with MLE by slice sampling. The noise prior of GP is a Gaussian distribution with 0 expectation and 0.001 variance and the mean function is set to be a scalar which equals zero, which is a common setting.\n Global/general Response:\n```\nWe thank the valuable comments of all the reviewers, and we refine the submission according to the feedback. The modified parts are colored (magenta for Reviewer gM84, red for Reviewer j7Na, cyan for Reviewer 22Se, blue for Reviewer 46kD, and green for Reviewer itsg). \n\n```\n", "role": "user" }, { "content": "Title: Response to authors\nThank you for the response, and I apologize for the delay in mine. I appreciate the runtime analysis and the DoE experiments. However, many of my concerns regarding the paper remain.\n\n__Q1_: The robustness analysis is rudimentary, and is not unique to the Poisson Process - the exact same argument can be made for the GP.__\nI was referring to the fact that the analysis is simplistic. For example, what makes $97.72$% a particularly large or small number in the given context? This is particularly true when the percentage is not put in relation to anything.\n\n_Q2_: I disagree with the notion that the GP is not robust to noise. I do not believe that the first figure and robustness analysis is sufficient...\n\nThere is analysis and a subsequent intermediary conclusion in line 150-152 which strongly suggests this. \n\nQ4: Lack of comparison with PBO\nProposing a method intended for preferential BO without benchmarking against preferential BO alternatives is a rather large flaw. It is unfortunate that code is not available, but this does not warrant its exclusion. \n\nQ9: The setting of GP in the experiments in Fig. 1.\nMLE and Slice Sampling is rather contradictory. Which one is it? Fixing an incorrect noise variance level of 0.0001 invites misranking - the noise level of the GP should _clearly_ be inferred along with other hyperparameters in this case for a fair comparison. This is a rather concerning error in my view, and makes the results untrustworthy. Moreover, a noisy problem setting warrants noise-adapted acquisition function, so I disagree that Noisy EI is outside the scope of the work.\n\nAs I have substantial outstanding concerns regarding the correctness of the work, I will not increase my score. \n\n\n", "role": "assistant" } ]
AwL9ZZOPVlN	AutoML 2023 Conference	Reviewer_iLB7	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_iLB7, and the following are the paper you need to review:\n```<<AwL9ZZOPVlN>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper proposes a novel framework for NAS that simultaneously optimizes for performance and “introspectability,” and a multi-objective evolutionary NAS algorithm called XNAS that optimizes over this pareto frontier. The main improvement introduced with this framework is the increased robustness of this kind of multi-objective optimization over traditional accuracy-only optimization. The authors conducted experiments on the multi-objective performance of XNAS on 3 image classification tasks, and performed analysis on the network motifs and model introspectability, trustworthiness, and debuggability.\noverall_review: Strengths:\n\nThis paper includes a well-thought out experimental design that showcases the effectiveness of using the bi-level optimization problem consisting of the model accuracy and class-wise disentanglement using an average cosine distance for all pairwise comparisons (Section 3.2). XNAS utilizes this performance measure to search for architectures that are prone to separation between classes through the NSGA-II NAS algorithm. This framework for multi-objective optimization is shown to be more robust to identifying mispredictions, improving model explainability, and improving hypervolume in this multi-objective task. \n\nWeaknesses:\n\nShowcasing the model’s performance is good, but using the hypervolume as a comparison against the single objective optimization (figure 2) is somewhat misleading. In this case, it might come across as a means of proof for the improvement of the model when all the figure is showing is that the new framework optimizes over a new objective.\nAnother weakness in this paper is that it does not describe in detail the thought process behind the design of this interpretability objective and its expected use. There is a mention (lines 261-262) that discuss the motivation behind DNN trustworthiness and disentanglement, but there is no clearly described motivation elsewhere.\nFigure 3 is very small, and dense. Subsequently, it is difficult to grasp information from it.\nOne minor point is on the specificity of modality. By focusing on image classification tasks, it showcases a good general use case scenario, but this model would probably aim to see production into tasks where the higher prediction confidence or trustworthiness is useful (i.e. higher risk tasks) or the better model debuggability is useful, which is not covered in the 3 tasks.\n\nreview_summary: Overall, this paper includes thorough experimentation on image classification using the new framework for multi-objective NAS. While there are still some considerations about the overall motivational design behind this research, it clearly identifies a path to create debuggable models through multi-objective NAS. \nactions_required_to_increase_overall_recommendation: For figure 2, it might be better to showcase the difference between only the interpretability of the two optimization tasks rather than the hypervolume. This would showcase the models.\n\nExtending the experiments to cover more applicable tasks might include facial recognition tasks, or other “higher-risk” classification tasks.\n\nFor the lack of clear motivation (not in the field of research but in the actual workflow for this research), putting in a description of the effectiveness of leveraging the new explainability outputs over the single objective would be beneficial.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Author Response to Reviewer iLB7\nThank you for your support of our work! Based on your great feedback, we have made some changes to improve the quality of the paper. We address individual concerns in the following:\n\n> Showcasing the model’s performance is good, but using the hypervolume as a comparison against the single objective optimization (figure 2) is somewhat misleading. In this case, it might come across as a means of proof for the improvement of the model when all the figure is showing is that the new framework optimizes over a new objective. [...] For figure 2, it might be better to showcase the difference between only the interpretability of the two optimization tasks rather than the hypervolume.\n\nOur intent here was to demonstrate that our approach performs better than random (not optimizing for introspectability). To quantify how effective our approach is, we computed the hypervolume and observed measurable improvements. However, we can see how this can be misleading. We have updated Figure 2 to more clearly demonstrate that we care about the differences in introspectability mostly, while also showing that the achieved accuracies are still in the same range.\n\n> [The paper] does not describe in detail the thought process behind the design of this interpretability objective and its expected use. There is a mention (lines 261-262) that discuss the motivation behind DNN trustworthiness and disentanglement, but there is no clearly described motivation elsewhere. [...] For the lack of clear motivation (not in the field of research but in the actual workflow for this research), putting in a description of the effectiveness of leveraging the new explainability outputs over the single objective would be beneficial.\n\nWe appreciate you pointing this out as the motivation could certainly be made more clear. We have added a new paragraph in the \"Introspectability\" section within Section 3.2 that better motivates the use of introspectability. Please let us know if this clears it up and how we can further improve it.\n\n> Figure 3 is very small, and dense. Subsequently, it is difficult to grasp information from it.\n\nThanks to the increase in the page limit, we were able to increase the size of this figure by 25%. It is much more readable now.\n\n> One minor point is on the specificity of modality. By focusing on image classification tasks, it showcases a good general use case scenario, but this model would probably aim to see production into tasks where the higher prediction confidence or trustworthiness is useful (i.e. higher risk tasks) or the better model debuggability is useful, which is not covered in the 3 tasks. [...] Extending the experiments to cover more applicable tasks might include facial recognition tasks, or other “higher-risk” classification tasks.\n\nWe agree that this would be a highly appropriate experiment for the proposed framework. However, given the limited rebuttal timeline, we cannot run this experiment in time to include in the paper and thus leave it to future work.\n\nHaving addressed your concerns, we hope that you will consider improving your rating. :)", "role": "user" }, { "content": "Thank you for the detailed response from the authors. It resolves the majority of my concerns. I still support my favorable recommendation for this manuscript.", "role": "assistant" } ]
AwL9ZZOPVlN	AutoML 2023 Conference	Reviewer_XTyx	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_XTyx, and the following are the paper you need to review:\n```<<AwL9ZZOPVlN>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The authors propose multi-objective NAS that optimizes both performance and interpretability. The main insight presented by the authors is quantifying the introspectability as the disentanglement between latent representations for different data classes. Experiments show good trade-off can be achieved between accuracy and interospectability.\n\noverall_review: Positives:\n- The paper is well-written and easy to follow.\n- The paper addresses a timely problem, and the high-level idea is promising.\n\nNegatives:\n- Technical novelty is limited since the main technical contribution is defining a metric via disentanglement.\n- Not clear how feasible it is to extend the metric to beyond classification use cases.\n- Evaluation is done on a limited scope.\nreview_summary: Overall, I believe the authors are exploring the right direction, but the detailed approaches as well as the evaluation can be further enhanced.\nactions_required_to_increase_overall_recommendation: Suggest the authors to further discuss the insight behind the metric definition, compare with other existing approaches, and elaborate on how the approach can be applied to other tasks.\nRating: 2 reject, significant issues present\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Author Response to Reviewer XTyx\nThank you for your review. Based on your feedback, we have made some changes to improve the quality of the paper. We address individual concerns in the following:\n\n> However, the authors do not fully convince me why this disentanglement of dataset classes is the right approach.\n\nWe have added a new paragraph in the \"Introspectability\" section within Section 3.2 that better motivates introspectability. Therein, we discuss how introspectability enables us to have a prediction process that reflects the uncertainty of an instantiated model, a means of probing why decisions were made, and a means of identifying or correcting mispredictions. We then, in our experiments section, demonstrate how this can be accomplished as a result of the disentanglement of classes. This is especially convenient as it requires minimal expert domain knowledge. Please let us know if this clears it up and how we can further improve it.\n\n> Technical novelty is limited since the main technical contribution is defining a metric via disentanglement.\n\nWe would also like to say that our contributions also include how to optimize for the Pareto front with existing approaches, rigorous analysis of the architectures spanning the Pareto front (trade-offs in debuggability vs. accuracy, architectural motifs, etc.), and demonstration of how introspectability can be used to produce more trustworthy models in several experiments. Furthermore, the implementation of our framework is nontrivial, which we highlight in the supplemental material PDF and code.\n\n> Not clear how feasible it is to extend the metric to beyond classification use cases. the authors do not discuss how this metric can be extended beyond image classification use cases.\n\nWhile the evaluation on non-classification tasks is outside the scope of the paper, we absolutely agree that this is an important consideration. We have added discussion of how this can be done (with domain knowledge or additional data-driven statistics) in the \"Introspectability\" section within Section 3.2 (last paragraph) - introspectability is applicable to non-classification tasks (e.g., regression) as well.\n\n> compare with other existing approaches\n\nWe think that there is not a straightforward way of comparing the introspectability approach to other approaches out there, at least quantitatively. However, if this comment is concerning NAS methods, we consider the comparison of XNAS with other multi-objective population-based NAS algorithms to be out of scope for this work. While we would like to explore incorporating other NAS methods into our framework for evaluation, we consider this work to be a standalone proof of concept.\n\nHaving addressed your concerns, we hope that you will consider improving your rating. :)", "role": "user" }, { "content": "I'm grateful for the authors' response. It addresses most of my concerns, and I am inclined to revise the rating upwards.", "role": "assistant" } ]
AwL9ZZOPVlN	AutoML 2023 Conference	Reviewer_mmgM	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_mmgM, and the following are the paper you need to review:\n```<<AwL9ZZOPVlN>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The motivation of this paper is to optimize task performance and interpretability jointly. This paper proposed an interesting metric, introspectability, to qualify the model's interpretability. Introspectability measures cosine similarity between activation vectors. It is not clear why introspectability can be used to qualify interpretability. The well-known multiobjective evolutionary algorithm NSGA-II is used to optimize two objectives. The algorithm is demonstrated in NAS-Bench-201 on MNIST, CIFAR10, and ImageNet-16-120. \noverall_review: This paper provides a new application of multi-objective NAS. But the algorithm contribution is weak. The model's interpretability metric, introspectability, does not make sense to me. I hope the authors can clarify it. \nreview_summary: This paper proposed a new application of multi-objective NAS. Algorithms are not novel. How to qualify the model's interpretability is an interesting point. But I have concerns about it as I mentioned. So, my initial rating is borderline reject. I keep some space to adjust the score if the authors can solve my concern. Because the multi-objective NAS part is not new. I expect strong contributions from other sides, ie. model interpretability. \nactions_required_to_increase_overall_recommendation: I raised concerns about introspectability. If the authors can better explain it and provide some toy examples, I'd like to increase the score.\nRating: 5 marginally below the acceptance threshold\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Author Response to Reviewer mmgM\nThank you for your support of our work! Based on your great feedback, we have made some changes to improve the quality of the paper. We address individual concerns in the following:\n\n> It is not clear why introspectability can be used to qualify interpretability. I did not understand why the cosine distance between activation vectors can be used to quantify the model's interpretability. I did not figure out the connections between introspectability and interpretability. Can you provide a toy example to show the connection between introspectability and interoperability? [...] I raised concerns about introspectability. If the authors can better explain it and provide some toy examples, I'd like to increase the score.\n\nWe have added new paragraphs in the \"Introspectability\" section within Section 3.2 that better motivates introspectability and explains how it can be used. Introspectability enables us to have a prediction process that reflects the uncertainty of an instantiated model, a means of probing why decisions were made, and a means of identifying or correcting mispredictions. The mean activations per class can be thought of as the centroids that live in latent space, and the distance from each centroid can be thought of as the likelihood that a prediction is correct (with lower values being more likely). This enables us to identify mispredictions and mislabeled data, to identify why mispredictions happen (e.g., due to similar latent representations between classes), and to correct models under-performing on certain classes (e.g., by explicitly emphasizing latent separation between two confounded classes).\n\nToy example: Say that we have a model that predicts whether the weather will be sunny (0), rainy (1), or snowy (2). In a model that perfectly maximizes introspectability, the \"centroids\" corresponding to each class would be pointing in opposite directions. However, since this is not possible for multiple classes, the distances just need to be sufficiently far. For instance, for the toy centroids $a$=[1 0 0], $b$=[0 1 0], and $c$=[0 0 1] we have pairwise distances $ab$=0, $ac$=1, and $bc$=1. These distances inform us how well-separated the classes are in the model, potential interdependence between classes, and enable a variety of debugging. For a queried point, let's say it has the latent representation $d$=[0.5 0.1 0.1] in the model. This has distances of 0.04 to $a$, 0.81 to $b$, and 0.81 to $c$. There are several things we can do with this information. To measure confidence, we can find which centroid that $d$ is closest to ($a$ in this case). The distance to this centroid can be interpreted as inverse likelihood, which we discuss in the experiments section. Alternatively, we can say that since $d$ is significantly close to $a$ (sunny), and say that its ground truth label is rainy, then we can say that the data point is likely mislabeled (with the likelihood being proportional to the distance). Further technical details are given in the experiments section of this example and additional debugging applications. Please let us know if this helps!\n\n> The algorithm of multi-objective NAS is not novel\n\nWe apologize if it seemed that we claimed that the underlying NAS algorithm is novel. We use NSGA-II as part of our pipeline and for simplicity we refer to it with the introspectability metric as \"XNAS.\" We have updated the paper, specifically the contributions enumerated in the introduction, to make it clear that the novelty comes from the contributed metric, its use, and subsequent analyses.\n\n> On the other hand, can you provide FLOPs of your searched models? I want to check if introspectability will prefer smaller networks.\n\nYou are actually on to something here - introspectability prefers networks with more pooling layers than convolutional layers, which can be confirmed in the main text in Figure 4. As pooling layers are favorable for the metric, the number of parameters is reduced and the number of FLOPs is reduced. However, it is not a direct relationship between FLOPs and introspectability. We have just added figures showing the relationship between FLOPs and the two objectives (accuracy and introspectability) in Appendix P. We also recommend looking at the discovered motifs in Appendices F, G, and Q and how they relate to the two objectives.\n\n(see reply for continued response)\n> In experiments, classification accuracy is not good. That does not make sense to me. Multi-objective optimization will not prevent you from finding extreme solutions, namely high-accuracy solutions.\n\nThis is a good observation, however, the limited accuracy is not reflective of our approach. Rather, it is limited by the possible architectures within the NAS-Bench-201 search space. These accuracies are actually in the ballpark for the search space - see the reported NAS-Bench-201 accuracies from the original paper for both ImageNet-16-120 and CIFAR-10. Our reported maximum accuracies should be considered here as there is a trade-off with introspectability, as discussed in the paper. For instance, taking a look at ImageNet-16-120, our achieved performance is right at the optimal solutions.\n\n> The definition of domination has a problem. Not all items of a are necessarily bigger than b.\n\nThat is true - not all items of $\\mathbf{a}$ need to be bigger than those of $\\mathbf{b}$ for a solution to be non-dominated. However, this is the definition of strict domination, which we use to help define the set of non-dominated solutions. If we look at the definition again, $\\\\{\\mathbf{y}^{\\prime} \\in Y \\mid \\\\{\\mathbf{y} \\in Y \\mid \\mathbf{y} \\succ \\mathbf{y}^{\\prime}\\\\} = \\emptyset \\\\}$, it actually says we want the set of all solutions that are _not_ strictly dominated by another solution. This leaves us with just the non-dominated solutions. Let us know if you are still unclear.\n\n> In the literature review of NAS, mentioned works are a little bit old. I hope you can revise this part and add more recent works.\n\nWe have updated the references in the NAS background sections to include more recent papers, including:\n\n- Lopes, Vasco, et al. \"Efficient guided evolution for neural architecture search.\" Proceedings of the Genetic and Evolutionary Computation Conference Companion. 2022.\n- Baymurzina, Dilyara, et al. \"A review of neural architecture search.\" Neurocomputing 474 (2022): 82-93.\n- Tan, Mingxing, et al. \"Mnasnet: Platform-aware neural architecture search for mobile.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.\n- Zheng, Xinyue, et al. \"Disentangled neural architecture search.\" 2022 International Joint Conference on Neural Networks (IJCNN). IEEE, 2022.\n\n> The authors use random sampling to initialize vectors. I suggest they use the Latin Hypercube Sampling (LHS) sampling method.\n\nThank you for suggesting this, it looks quite promising. We have not heard of this to initialize the population in NAS before and will definitely experiment with this in the near future!\n\n> DAGs in NAS-Bench-201 are represented as vectors. Each element represents an operator. But I am not sure how you represent connections between nodes. Can you explain it?\n\nAbsolutely - each DAG is represented as a 6-dimensional vector, which determines the operations within a searched cell. The searched cell is shown in detail in Figure A1 in Appendix A. The connections are based on a fixed structure with four nodes and six edges. Each operation in the vector corresponds to an edge in the cell depending on its position. The text in Appendix A gives further details - let us know if this clears things up!\n\nHaving addressed your concerns, we hope that you will consider improving your rating. :)", "role": "user" }, { "content": "Thank you for the authors' reply. They have addressed most of the issues I raised, and I am considering increasing the score.", "role": "assistant" } ]
CF7FzuUkUck	AutoML 2023 Conference	Reviewer_cDaz	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_cDaz, and the following are the paper you need to review:\n```<<CF7FzuUkUck>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The main contribution of the paper is a multicriteria oversampling method that addresses the problems of bias and class imbalance.\n\noverall_review: Although the multicriteria oversampling method that addresses the problems of bias and class imbalance is interesting, it isn't clearly motivated and the relevance for AutoML is arguable.\n\nAdditionally, minor issues:\n-the statement in p1/l25 is arguable.\n-shouldn't the threshold in p5/l163 be a parameter of the evaluation function?\n\nreview_summary: Besides lack of clarity and many technical issues, the most important reason for my evaluation is that this is not really an AutoML paper.\n\nactions_required_to_increase_overall_recommendation: 1. improve motivation\n2. improve clarity of experimental setup and discussion of results\n3. complete comparison with state-of-the-art methods\nRating: 2 reject, significant issues present\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Author response to Reviewer cDaz\nRegarding your three \"Actions Required To Increase Overall Recommendation\":\n\n> 1. improve motivation\n\nWe have revised Section 2 to make the motivation more explicit. Please see the portions highlighted in blue.\n\n> 2. improve clarity of experimental setup and discussion of results\n\nYou were concerned that the 80% threshold is arbitrary. We have made two changes. First, we generalized the definitions of the threshold-based objectives to use \\tau instead of a hard-coded value. Second, we added an explanation that the 80% threshold is based on a rule by the US Equal Opportunity Commission.\n\n> 3. complete comparison with state-of-the-art methods\n\nWe already compare against 3 imbalance mitigators and 9 bias mitigators. That said, for the final version of the paper, we will add experiments comparing to two additional methods, FOS and Undersampling-multivariate.\n", "role": "user" }, { "content": "I appreciate the authors' feedback. They have addressed most of the issues I raised, and I am inclined to revise the rating upwards.", "role": "assistant" } ]
CF7FzuUkUck	AutoML 2023 Conference	Reviewer_3CZ2	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_3CZ2, and the following are the paper you need to review:\n```<<CF7FzuUkUck>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper proposes a simple approach, called ORBIT, that aims to address two types of bias: 1) class imbalance (where some classes have more/fewer examples than other classes); and 2) group bias, which causes examples with certain protected attribute values to be classified into unfavorable outcomes. Class imbalance can reduce the accuracy of a model, while group bias can negatively impact fairness. While oversampling can be used to mitigate both of these types of bias, the level of oversampling is a hyperparameter that should be tuned to balance between achieving good accuracy and fairness---simply reducing class imbalance may increase group bias, and vice versa.\n\nThe ORBIT algorithm extends SMOTE (the Synthetic Minority Oversampling Technique). ORBIT has two hyperparameters that can be tuned to focus more on reducing class imbalance or group bias. The paper evaluates ORBIT on 12 datasets, and compares it to three other approaches to mitigate class imbalance and nine other approaches to mitigate group bias. The authors also investigate five different ways to combine metrics for accuracy and fairness, to use single-objective hyperparameter optimization to tune the ORBIT hyperparameters.\noverall_review: Pros\n\n* The paper studies an interesting problem, aiming to control the degree to which both class imbalance and group bias are mitigated, by selectively oversampling the data.\n\n* The paper is mostly well-written except for a few important issues discussed in the \"Cons\" section below.\n\n* The authors provided source code with their submission, with files to reproduce specific figures in the paper. I have not run it, but I believe that the results are reproducible.\n\n* The experimental results (Figures 1, 4, 5, 6) provide error bars, which is nice to see.\n\n* The ORBIT algorithm is very simple; the core of the algorithm simply computes the number of samples to draw for each condition { (Class 0, Group 0), (Class 0, Group 1), (Class 1, Group 0), (Class 1, Group 1)} and then oversamples the data accordingly, using SMOTE.\n\n* Figure 2 (while cluttered and insufficiently annotated) is nice, and does a good job illustrating how oversampling to address only class imbalance may harm group bias, and vice versa. It also shows clearly how ORBIT can repair both class imbalance and group bias, or how it can selectively repair either one without making the other worse.\n\n* The paper performs a decent empirical evaluation, comparing ORBIT to three other approaches for imbalance mitigation and nine other approaches to bias mitigation, across 12 datasets.\n\n* Because there may be a trade-off between accuracy and fairness when mitigating bias, teh authors investigate five metrics using different approaches for combining fairness and accuracy into one objective, that can be used for single-objective (as opposed to multi-objective) hyperparameter optimization.\n\n\nCons\n\n* The paper deals with only binary classes and binary groups. Is ORBIT only applicable to such binary settings? All of the 12 datasets used in the paper are binary classification tasks.\n\n* The presentation of the results is problematic. In many of the plots, the results overlap so much that the labels are basically unreadable. I think that the authors should consider alternative formats to convey the results, that make it easy to parse each method without needing to guess what the labels are. For example, the results in the Fair-SMOTE paper, presented in tabular form, are relatively easier to parse.\n\n* The caption of Figure 1 is unclear; it does not give enough information to understand the plot (e.g., the metrics used, what each subplot represents, etc.) The titles of all subplots in Figure 1 are identical, which is confusing; these titles would be better changed to \"Logistic Regression,\" \"Support Vector Machine,\" \"Gradient Boosting,\" and \"Neural Network\" to be more informative about what each subplot shows.\n\n* Section 2 is poorly structured. While the section title is \"Problem Statement,\" it starts with an experimental result rather than describing the problem; it starts with \"Figure 1 shows results for ORBIT\" before the ORBIT method has been described. In addition, Figure 1 is not clearly described: disparate impact should be defined, and the way in which it is made symmetric should be clarified. The motivation for Figure 1 is to show that \"the effect of imbalance and bias repair on estimators is unpredictable,\" but this is only mentioned at the end of the second paragraph in the section. Either the Figure 1 example should be given after describing the problem (which starts in the third paragraph), or the motivation for Figure 1 should be given before describing the results. For example, it would be better to write an intro sentence like: \"To motivate the need to tune the bias repair factor, we first provide an example problem where the effect of bias repair is hard to predict a priori, when training different architectures (logistic regression, SVM, gradient boosting, or a neural network).\"\n\n* It seems that only one metric is used to measure fairness: (symmetric) disparate impact (DI). Because the tasks considered for ORBIT are all binary, other metrics are easily applicable, including recall, false alarm, accuracy, precision, and F1 scores---these can be computed from the confusion matrix of binary classification.\n\n* Around Eq. 1, it would be good to clarify whether \"disparate impact\" is the same as group bias.\n\n* In the paragraph at Line 67, it would be good to state what the optimal values are for the measures (e.g., the optimal class imbalance $\\sigma_{\\text{ci}}$ and group bias $\\sigma_{\\text{di}}$.\n\n* For Figure 2, the caption does not explain anything about how to interpret the figure. The numbers on the x- and y-axes are not explained (e.g., 400/600 and 0.333/0.714, etc.) I was able to parse the figure after reading it over for a while, but ideally each component should be described. While I think this figure looks nice, it is quite cluttered; I don't think it is necessary to write the \"0_\", \"1_\", \"_0\", and \"_1\" labels, because these look confusing without context, and do not add to the figure.\n\n* In theory, can ORBIT always perfectly eliminate imbalance and bias? Are there assumptions under which it can or cannot work?\n\n* In Line 81, it would be good to state the ranges for the hyperparameters $\\delta_{\\text{imbalance}}$ and $\\delta_{\\text{bias}}$. That is, are they both in $[0, 1]$?\n\n* Conceptually, it might be good to discuss tasks for which one may only want to eliminate group bias or class imbalance alone, as shown in Figure 2(e,f). If ORBIT can eliminate both types of bias, under what conditions would we only want to eliminate one of them?\n\n* I believe that $n_{\\text{ci}}$ and $n_{\\text{di}}$ are only defined on Line 82, after they were used in the previous paragraph.\n\n* Define \"diaeresis labels\" when using the term.\n\n* In Step 5 (Lines 141-142), what is the inverse of the get_class function? This step should be explained in more detail.\n\n* In Line 148, instead of \"followed by one of the classifiers listed in the caption of Figure 1,\" it would be clearer to list the classifiers here (logistic regression, SVM, gradient boosting, neural network).\n\n* In Lines 211-213, why does Hyperopt not have the option to select the neural network classifier (only the other three are options)?\n\n* Why is ORBIT not compared to FOS?\n\n* Figure 3 seems overly complicated, especially the \"test\" branch (why is there a separate node for \"X-y split\" when this is a dummy operation, as the inputs and targets are separate to start with?). Also, Figure 3 in this paper is quite similar in structure to Figure 4 in Fair-SMOTE, and it is not clear whether this is intentional or not.\n\n* The paper should provide more discussion of the hyperparameters for baseline methods, and how those hyperparameters were tuned.\n\n* The discussion of related work is limited, and could be expanded. I think the paper should give a better overview of different methods to address class imbalance, including undersampling the majority class [1, 2, 3, 4] and oversampling the minority class [1, 5]. These are just a few of the references missing from the paper:\n\n[1] Buda et al., \"A systematic study of the class imbalance problem in convolutional neural networks,\" Neural Networks 2018.\n\n[2] He & Garcia. \"Learning from imbalanced data,\" IEEE 2009.\n\n[3] Japkowicz. \"The class imbalance problem: Significance and strategies,\" AI 2000.\n\n[4] Ouyang et al., \"Factors in finetuning deep model for object detection with long-tail distribution,\" CVPR 2016.\n\n[5] Byrd & Lipton. \"What is the effect of importance weighting in deep learning?\" ICML 2019.\n\n\n\n\nMinor\n\n* In the abstract, it is stated that \"This paper demonstrates how to use a Bayesian optimizer to tune ORBIT\" --> Why is using BayesOpt here considered a contribution?\n\n* Also in the abstract, it may be slightly misleading to say that \"this paper introduces a new bias mitigator along with a methodology for applying automated machine learning for fairness,\" because this may imply that the notion of using AutoML for fairness is also new, which is not the case.\n\n* Figure 2 in this paper seems similar to Figure 3 in the Fair-SMOTE paper; is this just a different way to visualize a similar setup?\n\n* In Line 33, it is unclear what is meant by \"repair the data.\"\n\n* In Line 34, it should be more clearly explained why the \"effect of different mitigation levels is unpredictable.\" Shouldn't there typically be a predictable effect of the mitigation level with respect to reducing class imbalance at the expense of group bias, or vice versa?\n\n* I think it would be good to use consistent terminology throughout the paper: is \"mitigation level\" the same as \"repair level\"?\n\n* In the last paragraph of the introduction, it would be good to clarify which hyperparameters need to be tuned, when discussing AutoML applied to ORBIT.\n\n* I don't think the expanded form of the acronym \"ORBIT\" is ever written out. I inferred from the title that it is \"Oversampling to Repair Bias and Imbalance Together (ORBIT),\" but this should be clearly stated in the paper to avoid confusion.\n\n* L75: \"highest repair for\" --> \"highest repair coefficient for\"?\nreview_summary: ORBIT is a simple method that computes new dataset sizes, and uses SMOTE to oversample accordingly. ORBIT provides two hyperparameters that control the degree of mitigation of class imbalance and group bias, respectively. The authors evaluate ORBIT on several binary classification datasets, and compare to several other imbalance and bias mitigators. Unfortunately, the paper suffers from several clarity issues, both in the organization of the writing and in the presentation of results (in particular, the result plots are very hard to read because many points---and their labels---overlap). In addition, the related work is not well covered, and only one fairness metric is reported.\nactions_required_to_increase_overall_recommendation: I think that the results need to be presented more clearly, and the diagrams (e.g., Figure 3) should be revised to improve clarity. The writing also needs to be reorganized in places to improve clarity, for example in Section 2. Regarding the evaluation, it would be good to have more fairness metrics reported, as was done in prior work such as Fair-SMOTE. The discussion of related work also needs to be substantially improved.\nRating: 2 reject, significant issues present\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Author response to Reviewer 3CZ2\nThank you for your detailed constructive feedback! We have made extensive changes to the writing of the paper to address as many of your points as we could. Below, we focus on the main points under \"Actions Required To Increase Overall Recommendation\":\n\n> the results need to be presented more clearly\n\nWe have changed all the scatter plots to have non-overlapping data labels.\n\n> the diagrams (e.g., Figure 3) should be revised to improve clarity\n\nWe have applied your feedback to improve clarity in the hand-drawn diagrams.\n\n> The writing also needs to be reorganized in places to improve clarity, for example in Section 2.\n\nWe have revised Section 2 based on your feedback. Please see the portions highlighted in blue.\n\n> it would be good to have more fairness metrics reported\n\nWe can report more fairness metrics in the final version of this paper.\n\n> The discussion of related work also needs to be substantially improved.\n\nWe have added a new paragraph at the beginning of the related work citing and discussing general literature on class imbalance correction.\n", "role": "user" }, { "content": "Thank you for the authors' detailed response. They have adequately addressed the majority of my concerns, and I may consider adjusting the rating upwards.", "role": "assistant" } ]
CF7FzuUkUck	AutoML 2023 Conference	Reviewer_6LDz	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_6LDz, and the following are the paper you need to review:\n```<<CF7FzuUkUck>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This study introduces a method called ORBIT to mitigate the imbalance and bias at the same time. And they show that the best mitigation always does not lead to the best results, and there should be tuning to achieve the best results.\noverall_review: Positive:\n- Using an extensive number of data sets and models that help to generalize the idea\n- well-presented and written\n\nNegative:\nJust two minor issues:\n\n- Why the authors used only binary classification data sets? It would be interesting to see how it generalizes to multi-classification problems\n- In Fig 4, 5, 6. for some cases, it is hard to check the results due to the overlapping of texts. Maybe try another way of demonstration or use tables instead?\nreview_summary: Based on their contribution to making tuning between bias and imbalance and the insights they provided, and also considering the soundness of the paper, I believe having such papers in the community will help the community greatly.\nactions_required_to_increase_overall_recommendation: I think the figures should be changed in a way to be more readable.\nRating: 8 accept, good paper\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Author response to Reviewer 6LDz\nThank you for your positive review! Based on your feedback, we have changed all the scatter plots to have non-overlapping data labels.", "role": "user" }, { "content": "I acknowledge the authors' response. The reply answers almost all of my questions. I continue to recommend this work for acceptance.", "role": "assistant" } ]
CF7FzuUkUck	AutoML 2023 Conference	Reviewer_cM3T	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_cM3T, and the following are the paper you need to review:\n```<<CF7FzuUkUck>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper introduces a new algorithm, ORBIT, which extends SMOTE as an oversampling technique that is focused on repairing group bias and class imbalance at the same time with tunable parameters. Experiments show that ORBIT successfully repaired performance across many different models and datasets compared to the other techniques.\n\noverall_review: This a very well-written paper. I have two significant concerns that need to be addressed.\n\n- Add a limitation section to discuss examples/use cases where ORBIT fails.\n- Address the difference in reported performance in Figure 4 and Table 4 in [8].\n\nreview_summary: The ORBIT algorithm can significantly impact AutoML systems to help mitigate the negative effects of group bias and class imbalance simultaneously. The paper is well written and has the potential to be great, considering the authors highlight some of the significant impacts of the algorithm compared to SMOTE, FOS, or other existing techniques.\nactions_required_to_increase_overall_recommendation: Adding a limitation section and expanding on the difference in performance will help. Adding some information on the ethical effect of repairing can further strengthen the paper.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Author response to Reviewer cM3T\nThank you for your positive review! Regarding your points under \"Actions Required To Increase Overall Recommendation:\"\n\n> Adding a limitation section\n\nWe have added a new limitation section with a discussion of the points you suggested, please see the updated PDF.\n\n> expanding on the difference in performance\n\nYou asked why SMOTE on the Adult dataset yielded a different result in the FOS paper than in our paper. We found it difficult to track down the reason. The FOS paper assumes that all features are numeric, but does not discuss how they encode categorical features. We use an ordinal encoder, so it is possible they used a different encoder that yielded better results on this dataset. Furthermore, the FOS paper does not report error bars, so it is even possible that their result simply arose from a lucky split.\n\n> Adding some information on the ethical effect of repairing\n\nYou wrote \"it would be a bonus to see how the bias mitigator helped repair some ethical issues about the dataset\". We ran out of time during the author response period to dig deeper into this. One thought would be to look at individual fairness for samples whose predictions changed. Note that the aggregate effect of such cases is already reflected in the measured disparate impact scores.\n", "role": "user" }, { "content": "Thanks for the authors' response. The response covers nearly all of the issues I raised. I still support my favorable recommendation for this manuscript.", "role": "assistant" } ]
EMys3eIDJ2	AutoML 2023 Conference	Reviewer_73uc	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_73uc, and the following are the paper you need to review:\n```<<EMys3eIDJ2>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This work investigates how the architecture of a neural network impacts three metrics: The “expressivity” meaning how well the next work can approximate the training data, the “convergence” meaning how fast the network learns, and the “generalization” meaning how well the network performs on unseen data. Given a fixed budget of parameters, a “no free lunch” behavior can be observed, meaning that no model from a NAS space can achieve optimal performance in all three metrics, given a fixed amount of parameters. The authors derive the claims from a theoretical point of view and perform experimental evaluations on Nas-bench-102 to verify the claims. \n\noverall_review: Positives:\n\n- The results are interesting from a deep learning theory perspective. \n- The formal results section as well as the experiments are extensive.\n- The three metrics are an interesting way to characterize the performance of a neural network.\n\nNegatives:\n\n- The authors claim in the abstract that the default measure of performance in AutoML (the test accuracy) is closely related to the three aspects (expressivity/convergence/generalization). However, this is only explained for the aspect of ‘generalization’ in Section 4.4, but not for the concepts of expressivity and convergence. As far as I can see, this is a major flaw, as it limits the conclusion that insights can be derived for the AutoML community. \n- The “Architecture bias in AutoML” section (lines 192-196) is shallow, given that here the “no free lunch” behavior is put in context to applications in AutoML. A more detailed explanation is necessary. \nreview_summary: Overall, the paper presents interesting results and supports them with theoretical derivations and particle experiments. My main concern is that the impact on the field of AutoML is limited. \n\nactions_required_to_increase_overall_recommendation: More explanations on the impact in the field of AutoML.\nRating: 5 marginally below the acceptance threshold\nConfidence: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper\n", "role": "assistant" }, { "content": "Title: Response to Reviewer 73uc and Follow-up\n### 1. Paper Structure\nThank you for your valuable feedback! We have taken your concerns into consideration and have made revisions to our draft by establishing more connections between Section 3 and Section 4, as well as including red lines and subtitles to better illustrate these connections.\n\nWe have chosen to defer formal results to Section 4 in order to present our core findings and implications to wider AutoML audiences without being overwhelmed with theoretical details.\n\nWe are open to further revisions based on your feedback! As confirmed by reviewer YQ71, our draft is “well-written and clearly explains complex technical concepts in a way that is accessible to a broad audience” and by reviewer 7Zmj that our draft is “very clear.”\n\n### 2. Explanation: why test accuracy is closely related to three aspects (expressivity/convergence/generalization)\nWe appreciate your question. It is important to note that the (decomposition of) test loss is directly linked to expressivity, convergence, and generalization.\n\nThis is because: related generalization bounds [1, 2], including ours (Theorm 4.5), are based on the assumption that wide networks can achieve small training errors. See Section 4 in [1] “predicts linear convergence of GD to 0 loss” and Remark 3.4 in [2] “Theorem 3.3 suggests that if the data can be classiﬁed … with a small training error, the over-parameterized ReLU network learned by Algorithm 1 will have a small generalization error.”\n\nTherefore, it is crucial to ensure that the network can achieve a low training error (i.e., high expressivity), and we hope to achieve this with fewer steps (i.e. faster convergence). That means, the training error and the convergence rate are not explicitly characterized in Theorem 4.5, which makes our analysis of expressivity and convergence necessary.\n\nWe will make sure to include this explanation in our camera ready.\n\n\n### 3. “Architecture bias in AutoML” is shallow\nThank you for your input. We have revised this point to provide more precise context, focusing specifically on differentiable NAS, and have included additional explanations.\n\n[1] “Fine-Grained Analysis of Optimization and Generalization for Overparameterized Two-Layer Neural Networks” Arora et al. 2019\n\n[2] “Generalization Bounds of Stochastic Gradient Descent for Wide and Deep Neural Networks” Cao et al. 2019\n\nDear reviewer 73uc,\n\nSince the author-reviewer discussion period will end soon, we will highly appreciate it if you could check our updated draft and our response to your review comments at your earliest convenience. This way, if you have further questions and comments, we can still have time to reply.\n\nIf our response resolves your concerns, we kindly ask you to consider raising the rating of our work.\n\nThank you very much for your time and efforts!\n\nSincerely,\n\nAuthors of submission 59\n Global/general Response:\n```\nWe truly thank all questions and suggestions from the three reviewers. We are happy that reviewers YQ71 and 7Zmj acknowledge that we work on an important problem for the AutoML community, with well-conducted experiments, and also acknowledge our clear paper writing. We also appreciate reviewer D5bQ’s confirmation of our reproducibility.\n\nWe address all questions below.\n```\n\n Global/general Response:\n```\nDear all reviewers,\n\nSince the author-reviewer discussion period has started for a few days, we will appreciate it if you could check our updated draft and our response to your review comments soon. This way, if you have further questions and comments, we can still reply before the author-reviewer discussion period ends.\n\nIf our response resolves your concerns, we kindly ask you to consider raising the rating of our work.\n\nThank you very much for your time and efforts!\n\nSincerely,\n\nAuthors of submission 59\n```\n", "role": "user" }, { "content": "Title: Response to Authors\nI thank the authors for their response. \n\n1. Regarding the paper structure, I acknowledge that the current structure was chosen to present the results in the first part of the manuscript and defer the theoretical explanations to the second part which is a valid reason. \n\n2. I also appreciate the explanations on the decomposition of test accuracy metric. \n\n3. However, I still find the section on the \"Architecture bias in (differentiable) NAS\" to not provide enough context to explain the impact of the findings on applications in AutoML. As the focus of this venue is specifically on AutoML, I see this section as one of the most important parts of the paper that would warrant an in-depth explanation (which should also be feasible given that there is still roughly one more page left to the page limit). ", "role": "assistant" }, { "content": "Title: Thanks for your feedback!\nDear reviewer 73uc,\n\nWe are deeply grateful for your prompt and insightful feedback on our manuscript.\n\nIn response to your suggestions, we have revised and expanded Section 3.2.5 to provide a more comprehensive context and clearer explanations of our impact and findings regarding AutoML. We acknowledge the significance of this section for the AutoML community and have made the necessary improvements.\n\nWe kindly request that you review the updated Section 3.2.5. We are committed to making additional revisions if needed to ensure the highest quality of our work.\n\nSincerely,\nAuthors of submission 59", "role": "user" }, { "content": "Title: Explanation on AutoML Impact\nI thank the authors for extending the section. I remain skeptical, especially as the explanation in line 213 \"Therefore, the differentiable search tends to favor networks that can minimize training loss as quickly as possible.\" sounds more like an overfitting problem to me, which would be obvious in a network with not enough learnable parameters. I am therefore raising my score only from weak reject to borderline. ", "role": "assistant" }, { "content": "Title: Thanks for your feedback and look forward to further discussions\nDear reviewer 73uc,\n\nWe appreciate your continued engagement in the discussion!\n\nWe would like to elaborate on the following points:\n1. This collapse issue happens at the early stage of DARTS optimization (that is why \"early stopping\" is proposed to mitigate this problem [1]). \nAt this stage, network parameters have not been fully trained and the training loss remains high. Consequently, the DARTS supernet and its subnetworks are not in an overfitting state.\n2. \"networks with not enough learnable parameters\" usually tend to underfit, not overfit.\n\nWe hope this clarification addresses your concerns, and we are happy to further discuss before the rebuttal period ends.\n\nSincerely,\n\nAuthors of submission 59\n\n[1] \"DARTS+: Improved Differentiable Architecture Search with Early Stopping\" Liang et al. 2020\nDear reviewer 73uc,\n\nPer the AutoML official policy, we can still read and respond to reviews/reviewer comments. Thus, we would highly appreciate it if you could check our latest response to your comment. This way, if you have further questions or comments, we can still reply in a timely manner.\n\nIf our response resolves your concerns, we kindly ask you to consider raising the rating of our work.\n\nThank you very much for your time and efforts!\n\nSincerely,\n\nAuthors of submission 59", "role": "user" }, { "content": "I appreciate the fruitful discussion with the author, which really makes this work stronger.", "role": "assistant" } ]
EMys3eIDJ2	AutoML 2023 Conference	Reviewer_YQ71	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_YQ71, and the following are the paper you need to review:\n```<<EMys3eIDJ2>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper main contributions are as follows:\n• Theoretical analysis of the dependence of a deep network’s manifold complexity, convergence rate, and generalization gap on its graph topology.\n• Discovery of the “no free lunch” behaviour: given a fixed budget on the number of parameters pre-defined in an architecture space, no such a network can achieve optimal expressivity, convergence and generalization at the same time. \n• Analysis explaining a wide range of observations in AutoML and NAS practices. Experiments on popular datasets confirm the theoretical analysis. \noverall_review: \nPros\n1. sheds light on the limitations and challenges of optimizing neural networks in an architecture space\n2. The paper is well-written and clearly explains complex technical concepts in a way that is accessible to a broad audience.\n3. The paper demonstrates a comprehensive understanding of the technical literature in the field, and builds on it in a novel and innovative way\n3. \n\nCons\n1. To generalise the results obtained from this study, different network and datasets should be included.\n2. different networks may have different optimal learning rate. Therefore, it may be deemed unfair to compare them under identical training protocols\n\n\n\nreview_summary: \nThe paper facilitates the explanation of the architecture bias in AutoML and NAS applications through jointly analysing a network’s expressivity, trainability, and generalization, and how they are influenced by the architecture’s graph topology. Given a fixed budget of the number of parameters, it has been shown that the expressivity favors networks of deep and narrow graph topologies, whereas both\nthe trainability and generalization prefer wide and shallow ones. The paper is the first to discover that these inductive biases lead to a “no free lunch” behavior in deep network architectures. The paper is well-written and easy to follow. \n\nHowever To generliaze the results obtained from this study, different network should be studied. Additionally, different networks may have different optimal learning rate. Therefore, it may be deemed unfair to compare them under identical training protocols. \n\nactions_required_to_increase_overall_recommendation: To generliaze the results obtained from this study, different network should be studied. Additionally, different networks may have different optimal learning rate. Therefore, it may be deemed unfair to compare them under identical training protocols.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer YQ71\n### 1. Different networks and datasets\nThank you for your feedback. Our experiments were designed to focus on the graph-based architecture space for image recognition, primarily due to the fact that most highly cited NAS benchmarks (NAS-Bench-101, NAS-Bench-201) are graph-based and used for image recognition. We followed this convention to align with the broader AutoML community.\n\nMoving forward, we aim to include a wider range of architectures and datasets in our research.\n\n### 2. Different networks may have different optimal learning rates.\nWe really appreciate this question!\n\nIt is worth noting that current NAS benchmarks (e.g., NAS-Bench-101, NAS-Bench-201) employ fixed training hyperparameters across all networks, which may not be optimal. In this work, we have opted to follow the convention established by these benchmarks to ensure compatibility with the broader AutoML community.\n\nMeanwhile, we are currently undertaking further research to analyze the impact of architecture-aware training hyperparameters on network performance.\n\n Global/general Response:\n```\nWe truly thank all questions and suggestions from the three reviewers. We are happy that reviewers YQ71 and 7Zmj acknowledge that we work on an important problem for the AutoML community, with well-conducted experiments, and also acknowledge our clear paper writing. We also appreciate reviewer D5bQ’s confirmation of our reproducibility.\n\nWe address all questions below.\n```\n\n Global/general Response:\n```\nDear all reviewers,\n\nSince the author-reviewer discussion period has started for a few days, we will appreciate it if you could check our updated draft and our response to your review comments soon. This way, if you have further questions and comments, we can still reply before the author-reviewer discussion period ends.\n\nIf our response resolves your concerns, we kindly ask you to consider raising the rating of our work.\n\nThank you very much for your time and efforts!\n\nSincerely,\n\nAuthors of submission 59\n```\n", "role": "user" }, { "content": "Thanks for the authors' response. The response addresses most of my concerns. I remain supportive of my positive recommendation for this paper.", "role": "assistant" } ]
EMys3eIDJ2	AutoML 2023 Conference	Reviewer_7Zmj	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_7Zmj, and the following are the paper you need to review:\n```<<EMys3eIDJ2>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper formalizes expressivity, convergence rate and generalization gap of a neural network based on its graph topology, and discover that given such representation and a fixed parameter budget, it is not possible to achieve optimal expressivity, convergence and generalization. This is because given the graph formulation of architectures, where Ph stands for number of end-to-end unique paths, and dp being the number of linear transformation operations on the p-th path, expressivity requies smaller Ph and larger dp, whereas convergence rate and generalization require larger Ph and lower dp. This is further empirically justified in a fixed setting in Figure 3 and Figure 4.\noverall_review: This is a well written paper, with limited but effective empirical testing. Their formalization of expressivity, generalization and convergence is well justified and supported with conclusions verified by Figure 3 and Figure 4. The limitations are discussed in the conclusion, which would motivate future work. \n\nI would like the authors views on including discussions regarding the implications of the structure of the linear transformations e.g. instead of an MLP, would convolutions/transformer based architectures and other aspects of the operations themselves (stride/pooling etc.) change the conclusions in the paper?\n\nDiscussion on whether their toy search space is sufficient to support their conclusions would also help.\n\n\nreview_summary: I am somewhat confident that this paper should be accepted, it is well written with proofs and empirical evidence on a small search space of 729 neural networks to justify their observations derived from their formalization of expressivity, generalization and convergence. It is possible that I did not understand some parts of the submission (generalizability theorems) due to lack of familiarity with prior work.\nactions_required_to_increase_overall_recommendation: Comments on whether their toy search space is sufficiently large/diverse for verifying their claims would help.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper\n", "role": "assistant" }, { "content": "Title: Response to Reviewer 7Zmj\n### 1. MSE loss in Theorem 4.4\n\nAlthough in Theorem 4.4 we used MSE loss, in our experiments we generalize this conclusion to the cross-entropy loss, which removes the possible limitation.\n\n### 2. More layer types\n\nThanks for this question! We will try to expand our operation set to include convolution, attention, and other layers.\n\nMeanwhile, as the implication of our work on the convolutional network is beyond Linear-ReLU/Skip/Zero operations (listed in Section 3.2.5), our core conclusion, the “No Free Lunch” behavior remains unchanged. The primary difference may lie in the requirements on the number of channels (for example, the requirement is multiplied by poly(p) from Theorem 6.1 to Theorem 7.1 in [1]).\n\nIt is worth noting that our experiments were rigorous despite using only three operations. We trained 729 architectures on three datasets, with each architecture trained for 3000 epochs and tested across three levels of randomness.\n\n[1] “Gradient Descent Finds Global Minima of Deep Neural Networks” Du et al. 2018\n\n Global/general Response:\n```\nWe truly thank all questions and suggestions from the three reviewers. We are happy that reviewers YQ71 and 7Zmj acknowledge that we work on an important problem for the AutoML community, with well-conducted experiments, and also acknowledge our clear paper writing. We also appreciate reviewer D5bQ’s confirmation of our reproducibility.\n\nWe address all questions below.\n```\n\n Global/general Response:\n```\nDear all reviewers,\n\nSince the author-reviewer discussion period has started for a few days, we will appreciate it if you could check our updated draft and our response to your review comments soon. This way, if you have further questions and comments, we can still reply before the author-reviewer discussion period ends.\n\nIf our response resolves your concerns, we kindly ask you to consider raising the rating of our work.\n\nThank you very much for your time and efforts!\n\nSincerely,\n\nAuthors of submission 59\n```\n", "role": "user" }, { "content": "Title: Response To Authors\nThank you for your response. \n\nI would like to keep my score and maintain my current confidence in my evaluation. \nI strongly feel that a more in-depth empirical evaluation may be very valuable (in terms of the architecture design space). I will look at the discussions with the other reviewers and re-asses with higher confidence if possible once those discussions conclude.", "role": "assistant" } ]
EMys3eIDJ2	AutoML 2023 Conference	Reviewer_VWUy	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_VWUy, and the following are the paper you need to review:\n```<<EMys3eIDJ2>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper discusses the factors that define the \"goodness\" of neural architectures. It demonstrates that there there is no single architecture that best optimizes all 3 key properties, namely $\\textbf{expressivity}$ (the complexity of functions a network can approximate), $\\textbf{convergence}$ (training speed under gradient descent), and $\\textbf{generalization}$ (performance on unseen data).\noverall_review: The paper gives a clear introduction to the topic and provides a clear structure of all findings. The experiments are limited in the sense that the width and depth of a neural architecture might be misinterpreted and the the relatively small operation set might influence the expressiveness of the architecture space. Also, some architectural topologies have been evaluated more intensively than others and there might be an unequal advantage from the pre-defined hyperparameter setting. The formal results offer valuable insights; however, it would be beneficial to present them as concisely as possible, emphasizing the essential conclusions only.\nreview_summary: Considering the technical inconsistencies and the limited impact, I am inclined to reject the paper in its present form. Nonetheless, if the identified shortcomings are adequately addressed and resolved, the work has the potential to provide valuable insights for the community.\nactions_required_to_increase_overall_recommendation: (already addressed in depth above)\nRating: 4 possibly reject, but has redeeming facets\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer VWUy and Follow-up\n### 1. Depth and width of the ResNet block in Figure 2\nThis ResNet block is interpreted as two parallel paths ($P_H$ = 2), instead of two layers. The skip-connection path has a depth of 0, while the other path has a depth of 3 (since there are three Linear-ReLU operations).\n\n### 2. Total number of architectures\nWhen H = 3, there are 6 edges in the graph, with each edge being free to choose from 3 available operations (Linear-ReLU, Skip-connection, Zero). We thus have in total $3^6 = 729$ architectures. This is also true for our experiments: we trained all 729 architectures. Furthermore, we trained each network from this set three times with different randomness.\n\n### 3. “comparison under the same training protocol may be unfair”\nWe really appreciate this question!\n\nIt is worth noting that current NAS benchmarks (e.g., NAS-Bench-101, NAS-Bench-201) employ fixed training hyperparameters across all networks, which may not be optimal. In this work, we have opted to follow the convention established by these benchmarks to ensure compatibility with the broader AutoML community.\n\nMeanwhile, we are currently undertaking further research to analyze the impact of architecture-aware training hyperparameters on network performance.\n\n### 4. “networks of deep and narrow topology have higher expressivity”\nOur findings indicate that, given a fixed budget of the number of parameters, as illustrated in Figure 3 Left, the converged training loss is positively correlated ($R(P) = 0.11$) with the network’s topological width, indicating that wider topology will have higher training loss. Our training loss is small because we train networks for a long time (3000 epochs) to ensure that they are converged.\n\n### 5. “wide and shallow ones show faster convergence”\nMost existing works focus on network’s channel width (i.e. number of neurons) of single layers, in the context of feedforward networks. Our work generalizes this result to complicated graph topologies, instead of the channel width of each layer.\n\n### 6. “wide and shallow ones show […] smaller generalization gap”\nWe verify this conclusion on three datasets: CIFAR-10, CIFAR-100 (Figure 5), and Tiny Imagenet (Figure 3). We calculate the correlation between the generalization gap vs. topological depth ($R(d_p)$) and width ($R(P)$). From the right columns of both Figure 3 and 5, we find a positive correlation with depth and a negative correlation with width. This confirms our conclusion that wide and shallow networks show smaller generalization gap.\n\n### 7. Domain of image recognition only\nOur experiments are on image recognition mainly because most highly cited NAS benchmarks (NAS-Bench-101, NAS-Bench-201) are for image recognition, and we follow this convention to meet the trend of the broad AutoML community. Meanwhile, our theory does not assume any data modality.\n\n### 8. Other operations\nThanks for this question! We will try to expand our operation set to include convolution, attention, and other layers.\n\nIt is worth noting that our experiments were rigorous despite using only three operations. We trained 729 architectures on three datasets, with each architecture trained for 3000 epochs and tested across three levels of randomness.\n\nMeanwhile, the implication of our work on the convolutional network is beyond Linear-ReLU/Skip/Zero operations, as listed in Section 3.2.5.\n\n### 9. “skip-connection can improve the number of neurons from exponential to polynomial of the network depth.”\nThis means skip-connection can improve (reduce) the requirement on the channel number from exponential to polynomial of the network depth. In other words, the convergence result will hold without requiring the network to be exponentially wide. We have revised our draft with a better explanation.\n\n### 10. Better elaboration of Section 4.2\nThank you! We followed your suggestion and have moved Lemma 4.1 to the appendix and added more explanation for Theorem 4.1 in red lines.\nDear reviewer VWUy,\n\nSince the author-reviewer discussion period will end soon, we will highly appreciate it if you could check our updated draft and our response to your review comments at your earliest convenience. This way, if you have further questions and comments, we can still have time to reply.\n\nIf our response resolves your concerns, we kindly ask you to consider raising the rating of our work.\n\nThank you very much for your time and efforts!\n\nSincerely,\n\nAuthors of submission 59\n Global/general Response:\n```\nWe truly thank all questions and suggestions from the three reviewers. We are happy that reviewers YQ71 and 7Zmj acknowledge that we work on an important problem for the AutoML community, with well-conducted experiments, and also acknowledge our clear paper writing. We also appreciate reviewer D5bQ’s confirmation of our reproducibility.\n\nWe address all questions below.\n```\n\n Global/general Response:\n```\nDear all reviewers,\n\nSince the author-reviewer discussion period has started for a few days, we will appreciate it if you could check our updated draft and our response to your review comments soon. This way, if you have further questions and comments, we can still reply before the author-reviewer discussion period ends.\n\nIf our response resolves your concerns, we kindly ask you to consider raising the rating of our work.\n\nThank you very much for your time and efforts!\n\nSincerely,\n\nAuthors of submission 59\n```\n", "role": "user" }, { "content": "Title: Response to Authors\nThank you for the response! While certain aspects are more comprehensible now and the overall clarity of the paper could be improved, the two most critical points remain unresolved: \n\n1. Since you are defining the depth of a neural network as the average over all paths, the residual block has an average depth of (0 + 3) / 2 = 1.5, while the same architecture without the skip-connection has an average depth of 3 / 1 = 3. A justification for this has yet to be provided.\n\n2. I understand that the architecture space of H = 3 includes 729 architectures. However, architectures with a width of 1 (single path) can only have operations of type \"Linear-ReLU\" or \"Zero\", since \"Skip-connection\"does not make sense for a single path. Thus, there is only a single architecture with width of 1 and depth of 3 (all nodes are connected with \"Linear-ReLU\") [o -> o -> o -> o], 4 architectures with a width of 1 and a depth of 2 (nodes that are connected with \"Linear-ReLU\": {0,1,2}, {0,1,3}, {0,2,3}, {1,2,3}) [o -> o -> o], and 6 architectures with a width of 1 and a depth of 1 (nodes that are connected with \"Linear-ReLU\": {0,1}, {0,2}, {0,3}, {1,2}, {1,3}, {2,3}) [o -> o]. This influences the expressivness of your results.\n\nAlso, a correlation coefficient of 0.11 only suggests a weak positive correlation (point 4). Given that the hyperparameters have not been optimized for the different architectures (point 3), the implications of this results are limited and should be interpreted with caution.\nLast but not least, point 7 is indeed well-founded, it was just not effectively conveyed in the main paper.\n\nIn light of the provided information, I will maintain my current score.", "role": "assistant" }, { "content": "Title: Further Response to Reviewer VWUy\nDear reviewer VWUy,\n\nThank you very much for your important questions!\n\nIt took us some extra time to run more experiments and better prepare this response.\n\n1. Yes, your point of view is correct, which is why in Figure 2 we wrote $d_0 = 3$ for MLP and $d_0 = 0, d_1 = 3$ for ResNet Block (i.e. an average depth of 1.5).\n\n2. It is true that there is only one architecture with a width of 1 and a depth of 3. However, there are 25 architectures with a width of 1 and a depth of 2. For each list below, we have an input node and three hidden nodes, and integers indicate operations connecting with previous nodes. \"0\" indicates \"Zero\" (broken edge), \"1\" indicates \"Skip-connect\", \"2\" indicates \"Linear-ReLU\". For example, \"[[0], [2, 0], [0, 0, 2]]\" means \"Zero\" b/w input and node 1; \"Linear-ReLU\" b/w input and node 2; \"Zero\" b/w node 1 and node 2; \"Zero\" b/w input and node 3; \"Zero\" b/w node 1 and node 3; \"Linear-ReLU\" b/w node 2 and node 3. Certainly, the \"Zero\" operator probably makes some architectures identical eventually, but they are still valid computational graphs in the architecture space, and they contribute to randomness of dots in our plots.\n* [[0], [2, 0], [0, 0, 2]]\n* [[0], [2, 0], [0, 1, 2]]\n* [[0], [2, 0], [0, 2, 2]]\n* [[0], [2, 1], [0, 0, 2]]\n* [[0], [2, 1], [0, 1, 2]]\n* [[0], [2, 1], [0, 2, 2]]\n* [[0], [2, 2], [0, 0, 2]]\n* [[0], [2, 2], [0, 1, 2]]\n* [[0], [2, 2], [0, 2, 2]]\n* [[1], [0, 2], [0, 0, 2]]\n* [[1], [2, 0], [0, 0, 2]]\n* [[2], [0, 0], [0, 2, 0]]\n* [[2], [0, 0], [0, 2, 1]]\n* [[2], [0, 0], [0, 2, 2]]\n* [[2], [0, 1], [0, 0, 2]]\n* [[2], [0, 1], [0, 2, 0]]\n* [[2], [0, 2], [0, 0, 1]]\n* [[2], [0, 2], [0, 2, 0]]\n* [[2], [1, 0], [0, 2, 0]]\n* [[2], [1, 1], [0, 2, 0]]\n* [[2], [1, 2], [0, 2, 0]]\n* [[2], [2, 0], [0, 0, 2]]\n* [[2], [2, 0], [0, 2, 0]]\n* [[2], [2, 1], [0, 2, 0]]\n* [[2], [2, 2], [0, 2, 0]]\n\nSimilarly, there are 114 architectures with a width of 1 and a depth of 1.\n\n3. We follow your recommendation, and prepare a comprehensive experiment with tailored learning rates for each architecture.\nWe study the empirical expressivity, trainability, and generalization by training networks with their optimal learning rates tuned by grid search.\n\nWe attach our revised draft in this anonymous repo: https://anonymous.4open.science/r/anonymous_AutoML2023_59-7058/AutoML2023_59.pdf\n\nAs shown in Figure 3 in our Section C.3, even with architecture-wise optimal learning rates, our conclusion still holds: in a complete architecture space, given a fixed budget of network parameters, deep narrow networks have higher expressivity, while wide shallow networks converge faster and generalize better.\n\nWe attach the optimal learning rates we find for each architecture in the anonymous repo (\"lrs.txt\"): https://anonymous.4open.science/r/anonymous_AutoML2023_59-7058/lrs.txt", "role": "user" }, { "content": "Title: Further Response to Authors\nThank you once again for the clarification.\n\nConcerning points 1 and 3, although you effectively eliminate one potential cause of misconception by using tailored learning rates, the current definition of 'depth' could still affect the conclusions drawn. For instance, the observed results could primarily be due to the presence or absence of skip connections rather than being attributed to the depth or shallowness of the network (correlation vs causation). An ablation study might provide valuable insights in this regard.\n\nConcerning point 2, while including \"broken connections\" indeed increases the number of evaluated architectures, the distribution remains uneven, ranging from a single (!) evaluated architecture to 114 evaluated architectures for a width of 1. As initially suggested, a more balanced approach to the distribution of evaluated architectures could enhance comparability.", "role": "assistant" }, { "content": "Title: Further Response to Reviewer VWUy\nDear reviewer VWUy,\n\nWe appreciate your timely response and discussion!\n\n1. Shu et al. [1] have a definition of depth that shares similar motivation with us, i.e., to characterize the graph topology of networks in NAS, and we also share similar architecture spaces (which include skip connections). Their observations are accepted by the AutoML community. Studying an architecture space with the skip connection is also common in the AutoML community [2].\n\nMeanwhile, would you suggest any concrete ablation study that could clear your concern?\n\n2. We would like to point out:\n\ni) The motivation for projecting the high-dimensional architecture space to two principal directions (topological depth and width) is not to pursue a balanced distribution of architectures in this 2D space. Instead, it is because we find that the topological depth and width can well-distinguish the behavior of networks (w.r.t. expressivity/convergence/generalization). As an analogy, PCA never projects data to be a balanced distribution, but to preserve feature variances.\n\nii) Again, numerous NAS papers work on \"unbalanced\" architecture spaces, including [1] and [2].\n\n\n[1] \"Understanding Architectures Learnt by Cell-based Neural Architecture Search\" Shu et al. 2019.\n\n[2] \"NAS evaluation is frustratingly hard\" Yang et al. 2020", "role": "user" }, { "content": "I appreciate the fruitful discussion with the author, which has positively impacted the strength of this work.", "role": "assistant" } ]
Ec09TcV_HKq	AutoML 2023 Conference	Reviewer_W8u8	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_W8u8, and the following are the paper you need to review:\n```<<Ec09TcV_HKq>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper showcases an analysis of Hyperparameter Optimization for Automated Reinforcement Learning. Within the analysis of this process, the authors showcased the need for dynamic configurations for hyperparameters in AutoRL. Using IGPR and discretized modality plots showcased the dynamic hyperparameter landscape for DQN and SAC.\n\noverall_review: Strengths:\n\nThe authors introduce a new pipeline to showcase hyperparameter landscapes for dynamic hyperparameters with discrete training steps, or more specifically for AutoRL problems. The paper also showcases definite proof for the use of dynamic hyperparameter configurations for AutoRL algorithms. \n\nWeaknesses:\n\nOne weakness is the lack of analysis of the dynamic hyperparameter landscapes. The follow-on shown in the paper for this behavior was that this proves that dynamic hyperparameter landscapes should be prioritized, not how the hyperparameters change or how they can be predicted. This more important aspect to the field here is the characterization of the hyperparameter landscape. Discussing potential ways to analyze these landscapes would be beneficial. Another potential way to provide more information is to run more experiments to show different hyperparameter landscapes. \nOne other potential weakness in this paper is the use of Sobol sampling. Grid sampling can be non-optimal, but in a landscape analysis study, grid search can provide a good baseline to study the complexities of the landscape.\n\nreview_summary: This paper conducted several experiments into the hyperparameter landscape of AutoRL algorithms in order to prove the need for dynamic HP configurations, and to identify a novel way to track and analyze these HP landscapes. While interesting, this paper lacks some substantial contribution.\nactions_required_to_increase_overall_recommendation: My main suggestion is to provide more information on these hyperparameter landscapes within this paper. Specifically by plotting more scenarios with different algorithms. The use of a small proof for the need for dynamic hyperparameter configurations is interesting, but not impactful enough to warrant acceptance.\n\nPage 8, figure 6 description: Disretized -> Discretized\nRating: 4 possibly reject, but has redeeming facets\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Initial Response\nThank you for your review. Please find answers to your questions below.\n\n## More Scenarios\n\nWe added PPO on Bipedal-Walker as a new scenario in our revision. With our selection of diverse RL algorithms, encompassing policy-gradient based, value-based and entropy-based methods, we believe we are able to showcase the general dynamicity of the hyperparameter landscapes.\n\n## Analysis of Landscapes\n\nWe agree that our approach opens up many possibilities for further studies and insights in understanding AutoRL and thus the design of better AutoRL approaches. We strongly believe that already the analyses we provide in the paper are able to highlight some important points that can be corroborated by the literature. For example, the HP sensitivity of PPO can be very well validated with the multiple papers that talk about the importance of every design decision in PPO [1]. In contrast, we see a more stable landscape for SAC on a problem of similar complexity, signifying and corroborating the stability of entropy-based methods. With the data we collected for this study, the community can build upon it and for example further examine the interplay between the learning dynamics, exploration, and credit assignment.\n\n[1] Engstrom, Logan, et al. \"Implementation matters in deep policy gradients: A case study on ppo and trpo.\" arXiv preprint arXiv:2005.12729 (2020).\n\n\n## Sampling on the Landscape\n\nWe specifically choose Sobol sampling to use the best of both worlds, random sampling and grid sampling. Like grid search, Sobol sampling evenly covers the space and at the same time introduces more variations along one dimension. This is especially important in the case of low-effective dimensionality where we have unimportant hyperparameters, which we do not know beforehand [Bergstra & Bengio, 2012].\n\n## Potential Impact\n\nAnalyzing hyperparameter landscapes may allow for the prediction of optimized hyperparameter trajectories later on. This requires a sound foundation and the proposed work shows how the analysis of a landscape could be done. Due to the focus of this paper, predicting trajectories is not part of it, but will be considered for future work.\n\n Global/general Response:\n```\nWe thank all the reviewers for their helpful feedback. We are glad that the reviewers consider our approach to be relevant to the AutoRL community by filling the gap through empirical evidence to commonly held assumptions (Reviewer 9o8e), and potentially instrumental in analyzing hyperparameters and finding better configurations (Reviewer 7fjf). We also appreciate the reviewers acknowledging the solidity of our experimental procedure (Reviewer W8u8), noticing the benefit of our methodical study of HP landscapes, and considering it a crucial step in creating a practical guide for HPO for DeepRL applications (Reviewer FvT5).\n\nWe would like to reiterate that the rationale behind our paper was to show the importance of landscape analyses and their importance to the field of AutoRL. This study is a first step in that direction, and our hope is that with this, we can bring the attention of larger parts of the AutoML community, a community known for adding data-driven principles to the design of ML pipelines, to work on further properties of AutoRL. We believe our methodology can be built upon by future work to explore a more nuanced interplay between hyperparameters, exploration, and credit assignment in RL. \n\nWhile we will address individual points as separate replies, we would like to address some general points in this comment. \n## Additional Experiments\nWe additionally evaluated PPO for three phases on the Bipedal-Walker-v2 environment, which can be found in the results section. For this experiment, we varied the learning rate, discount factor, and the generalized advantage estimate factor (gae_lambda). The plots show that PPO is able to achieve higher returns in very specific hyperparameter regions as opposed to SAC which achieves a generally stable performance across a wide range. This adds to previous arguments about PPO being highly sensitive to code-level optimizations [1,2].\n\nWe believe that, along the same lines as DQN and SAC, these results demonstrate that our methodology adds principled analyses of RL algorithms that can be potentially applied to understand the nuances between dynamic hyperparameter changes and the learning dynamics of the algorithms that they impact. \n\nOur general comments regarding multi-fidelity optimization still hold. However, we would additionally like to point out that these results, the generated meta-data and the approach itself are meant as a gateway to more interesting analyses that can be possible using our methodology. Along this line, we believe work in understanding learning dynamics and exploration behavior of the algorithms is a promising and crucial future direction. \n\n## Sampling Strategy\n\nWhile grid search is potentially more interpretable, we don't know whether dimensions are correlated [3]. This creates issues since a simple grid would be inefficient in effectively covering the search space and potentially hampers the scalability of our approach. \n\nOur rationale for choosing Sobol sequences, as a low-discrepancy sequence, was to exploit its pseudo-random nature and its ability to cover more values along one dimension of the search space, addressing the potential low effective dimensionality. Thus, we see Sobol sequences as a way to get best of both worlds of grid search and random search. We further highlight that Sobol sequences are a well-known strategy for initial designs in optimizers such as Bayesian Optimization.\n\n\n[1] Engstrom, Logan, et al. \"Implementation matters in deep policy gradients: A case study on ppo and trpo.\" arXiv preprint arXiv:2005.12729 (2020).\n[2] Huang, Shengyi, et al. \"The 37 implementation details of proximal policy optimization.\" The ICLR Blog Track 2023 (2022).\n[3] Bergstra, James, and Yoshua Bengio. \"Random search for hyper-parameter optimization.\" Journal of machine learning research 13.2 (2012).\n\n## Changelog\nAs a result of the feedback received we made several updates to the paper:\n\n- Inclusion of more experimental results: PPO on Bipedal Walker\n- Updated Figure 1 and its description\n- Redesigned Figure 2 (data collection) and added a more comprehensive caption\n\nWe marked the updated text with green color.\n\n```\n", "role": "user" }, { "content": "Thanks for the authors' clarification. They have addressed most of the issues I raised, and I may consider adjusting the rating upwards.", "role": "assistant" } ]
Ec09TcV_HKq	AutoML 2023 Conference	Reviewer_NcQo	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_NcQo, and the following are the paper you need to review:\n```<<Ec09TcV_HKq>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper performs an empirical investigation of hyperparameter loss landscapes for two RL tasks (Cartpole and the Hopper environment from MuJoCo) using two algorithms (DQN and SAC). The authors focus on the dependence of each method/environment on hyperparameters at different points during training, motivated by prior empirical results showing the dynamically modifying hyperparameters during training can be beneficial compared to using static hyperparameters.\n\nThe authors visualized the hyperparameter loss landscapes at iterations 50k, 100k, and 150k for DQN, and at iterations 125k, 250k, 375k, and 500k for SAC. They found that the landscapes vary significantly at these points in training, supporting the notion of using hyperparameter schedules. They also plotted a measure of multimodality of the returns produced by any given hyperparameter configuration, across the hyperparameter space, and found that configurations become more multimodal as training progresses.\noverall_review: Pros\n\n* The paper studies an interesting problem, trying to understand the impact of hyperparameters over the course of training for RL tasks. Insights into the effects of these hyperparameters may be useful to understand why practical approaches that dynamically update the hyperparameters work well.\n\n* The results showing that the hyperparameter loss landscapes change substantially during training support the idea that hyperparameter schedules are important in RL.\n\n\nCons\n\n* Although I appreciate that running a sweep over different hyperparameters is not cheap, the empirical evaluation is still fairly limited, only investigating two algorithms (DQN and SAC) on two tasks (Cartpole and Hopper). These tasks are small-scale, so it is not clear whether the hyperparameter loss surfaces for larger-scale tasks would follow a similar pattern (e.g., perhaps the trends would be different, or they could be more or less pronounced).\n\n* The paper does not provide practical takeaways or actionable recommendations about which schedules to use in practical situations.\n\n* Overall, the clarity of the presentation could be improved. Figure 1 is not very well-designed and does not add much to the paper. The reader can only guess that the symbols in the \"data collection\" box are related to the data collection procedure from Figure 2 after having read several pages further in the paper. Also, Figure 1 seems too generic to be useful, because the steps of 1) collecting data for a performance dataset, 2) using that data to construct hyperparameter landscapes; and 3) analyzing that landscape, are very broad, and not specific to RL; they could just as well be applied to arbitrary ML tasks. From Figure 1, it is not clear what the \"search space\" on the left-hand side refers to. In the caption of Figure 1, it is not clear what \"modeled with landscape functions\" mean?\n\n* The organization of the writing is awkward. There is a lot of space devoted to explaining the basics of RL (Section 3.1), that could likely be summarized in one paragraph. At the same time, important details like the sampling procedure for hyperparameters are not expanded on clearly. The paper could use another pass over the writing to reduce wordiness and improve clarity.\n\n* In Section 4, it states that \"we imitate an optimal AutoRL optimizer that would always go for the best possible hyperparameter configuration schedule.\" The way this is described, it seems like a chicken-and-egg problem: how does one know what the best hyperparameter configuration schedule is without first knowing how the hyperparameter loss landscapes look under different conditions?\n\n* Why are the hyperparameter configurations sampled using the complicated scrambled Sobol sampling strategy? As this paper only considers three hyperparameters for each algorithm, it should be possible to run a simple grid search, which would be more interpretable.\n\n* Figure 2 is hard to parse, and not clearly described in the main text. In addition, the caption is completely uninformative. The font size is too small, and the arrows between blocks are not easy to interpret (owing to the lack of expanation in the caption). Why does it look like Step 5 in Figure 2 is determining the best configuration based on $\\mathcal{D}\\_{ls(1)}$? In the \"Training and Evaluation\" paragraph, it states that policies are selected based on their final performance, implying that in each stage the selection is based on $\\mathcal{D}\\_{\\text{final}(i)}$, not $\\mathcal{D}\\_{ls(i)}$.\n\n* It is not made clear what the \"middle surfaces\" are in Figure 3.\n\n* What do the white triangles denote in Figures 3 and 4? Are these the hyperparameters that were evaluated? They appear to be unevenly distributed in the space, and do not seem representative enough to generate informative contours. Because there are few hyperparameters here, would a simple grid search be applicable? This would yield an even distribution of points across the range that is plotted, and I think may be better suited for such contour plots.\n\n* Why are 3 timesteps used for DQN, while 4 are used for SAC? How were the timesteps chosen?\n\n* I think that the last paragraph of Section 5.2 (starting at line 269) is unnecessary; it only makes vague statements.\n\n* In Section 4.2, modality is not explained clearly enough. Does multimodality here refer to the distribution of performance values obtaiend when using the same hyperparameters with different random seeds? Also, the \"folding test of unimodality\" should be explained further here, as this is central to the quantitative results.\n\n* The submission did not provide source code.\n\n\nMinor\n\n* L121: \"Thus, An\" --> \"Thus, an\"\n\n* Eq. 3 can easily fit on one line.\n\n* In Eq. 3, I don't think that $\\mathcal{J}$ or $\\Pi$ are defined anywhere.\n\n* The first two sentences in Section 3.3 are redundant. In addition, the rest of Section 3.3 is wordy and seems overly drawn out.\n\n* In Section 4.1, it is not stated what the $ls$ in $t_{ls}$ stands for.\n\n* In Section 3.3, I think $\\pi_Z$ needs to be defined (e.g., as the policy that results from applying algorithm $Z$).\n\n* L49: \"The goal of AutoRL is to solve to facilitate\" --> \"The goal of AutoRL is to facilitate\"\n\n* L21: \"As the research\" --> \"As research\"\n\n* L171: \"we consider s set\" --> \"we consider a set\"\n\n* L188: \"then initial the configuration\" --> should this be \"then initialize the configuration\"?\n\n* L205: \"of the Sobol' sampler\" --> \"of the Sobol sampler\"\n\n* L243: \"train them on\" --> \"train them with\"\n\n* L246: Typo, \"10000\" should be \"100,000\"\n\n* L287: \"further analyses is\" --> \"further analyses are\"\n\n* L303: \"factore\" --> \"factor\"\n\n* L305: \"being in contrast\" --> \"in contrast\"\nreview_summary: Overall, this paper presents a fairly straightforward, visualization-centric, empirical investigation into hyperparameter loss landscapes for RL algorithms. The authors do not provide many takeaways from the results, other than that they support the idea of using hyperparameter schedules. There are no takeaways regarding what types of schedules would be useful, and what properties of the tasks would affect the optimal schedules. The experiments are limited, as they focus on two small-scale RL tasks, which may not be representative of larger, real-world problems. I do not think that the contribution of the paper is substantial enough to meet the bar for acceptance.\nactions_required_to_increase_overall_recommendation: I think that the writing and results need to be presented more clearly, and the diagrams (e.g., Figures 1 and 2) should be revised to improve clarity. I would encourage the authors to provide more analysis on larger-scale RL tasks and investigate which types of hyperparameter schedules are best for a given problem setting, to provide more practical takeaways.\nRating: 2 reject, significant issues present\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Initial Response\nThank you for your detailed review. We have addressed your points regarding the experiments and sampling strategy in our general comments. Please find our responses to your individual points below.\n\n## Selected Environments\n\nTo provide more evidence, we added further results in the results section on PPO for three phases on the Bipedal-Walker-v2 environment. See global answer for more details, too.\n\nGiven the problem of compute expense in running multiple sweeps across different environments, our choice was based on the following factors:\n\n- Diversity in environmental dynamics\n- How well does an algorithm perform in the given environment \n\nGiven that the selected algorithms (DQN, PPO, and SAC) cover three different kinds of objectives in the RL literature, we selected environments from both the classic control suite and MuJoCo to cover both simple and more nuanced dynamics. Given that we see similar trends, we expect that our results should generally hold. However, our method is agnostic to environments and agents, and thus, it can be transferred to any environment-agent combination.\n\n ## Clarity \n\nThank you for raising the point of clarity regarding the figures. We updated the figures for better accessibility. In more detail:\n\n- Figure 2: We redesigned Figure 2 to more intuitively show the data generation.\n- Figures 3 + 4: We denote the middle as mean surfaces for better clarity. The normal triangles represent local maxima, while the inverted ones, the local minima. The star represents the configuration selected for the next phase. We have added text in the caption to clarify this.\n\nWe will fix the result figures’ font sizes until the end of the rebuttal.\n\n### Figure 1\n\nWe agree that the notion of the different phases to tell it apart from the standard HPO procedure, as you rightfully mentioned, was missing in the figure and in the description, which we now updated. However we believe that this figure summarizes our work and serves as a primer for the upcoming sections.\n\n### Figure 2\n\nWe have significantly simplified Figure 2 to focus on the core aspects of our data collection strategy. We believe it should be more intuitive now since it concretely demonstrates an exemplary case of 3 configurations for 3 phases. \n\n## Basics of RL\n\nOur rationale for explaining the basics of RL in such detail was to address the AutoML community in general while being more self-contained in our writing. While these points are generally well understood in the subcommunities related to RL, we believe these points help an uninitiated reader to better understand our methodology.\n\n## “We try to imitate an optimal AutoRL optimizer”\n\nWe apologize for the misleading wording here and agree that it does indeed imply the chicken and the egg problem you mention. What we meant to say in that paragraph was that our approach to building the landscape involved focusing on the best configuration instead of all possible configurations, and we determine the quality of the configuration based on the final performance instead of the performance at that timestep. This would in fact correspond to a very well informed and expensive approach for an AutoRL optimizer. We have corrected this line to make this clearer.\n\n## Source Code\n\nWe would like to point you to line 247 which is where we provided the link to our source code. We have additionally added this link to the abstract for easier accessibility\n\nWe have corrected all the minor errors pointed out by you and hope to have a clearer version.\n\n## Modality\n\nWe have added an intuitive explanation of the folding test of modality that hopefully elucidates the meaning of $\\phi$ \n\n Global/general Response:\n```\nWe thank all the reviewers for their helpful feedback. We are glad that the reviewers consider our approach to be relevant to the AutoRL community by filling the gap through empirical evidence to commonly held assumptions (Reviewer 9o8e), and potentially instrumental in analyzing hyperparameters and finding better configurations (Reviewer 7fjf). We also appreciate the reviewers acknowledging the solidity of our experimental procedure (Reviewer W8u8), noticing the benefit of our methodical study of HP landscapes, and considering it a crucial step in creating a practical guide for HPO for DeepRL applications (Reviewer FvT5).\n\nWe would like to reiterate that the rationale behind our paper was to show the importance of landscape analyses and their importance to the field of AutoRL. This study is a first step in that direction, and our hope is that with this, we can bring the attention of larger parts of the AutoML community, a community known for adding data-driven principles to the design of ML pipelines, to work on further properties of AutoRL. We believe our methodology can be built upon by future work to explore a more nuanced interplay between hyperparameters, exploration, and credit assignment in RL. \n\nWhile we will address individual points as separate replies, we would like to address some general points in this comment. \n## Additional Experiments\nWe additionally evaluated PPO for three phases on the Bipedal-Walker-v2 environment, which can be found in the results section. For this experiment, we varied the learning rate, discount factor, and the generalized advantage estimate factor (gae_lambda). The plots show that PPO is able to achieve higher returns in very specific hyperparameter regions as opposed to SAC which achieves a generally stable performance across a wide range. This adds to previous arguments about PPO being highly sensitive to code-level optimizations [1,2].\n\nWe believe that, along the same lines as DQN and SAC, these results demonstrate that our methodology adds principled analyses of RL algorithms that can be potentially applied to understand the nuances between dynamic hyperparameter changes and the learning dynamics of the algorithms that they impact. \n\nOur general comments regarding multi-fidelity optimization still hold. However, we would additionally like to point out that these results, the generated meta-data and the approach itself are meant as a gateway to more interesting analyses that can be possible using our methodology. Along this line, we believe work in understanding learning dynamics and exploration behavior of the algorithms is a promising and crucial future direction. \n\n## Sampling Strategy\n\nWhile grid search is potentially more interpretable, we don't know whether dimensions are correlated [3]. This creates issues since a simple grid would be inefficient in effectively covering the search space and potentially hampers the scalability of our approach. \n\nOur rationale for choosing Sobol sequences, as a low-discrepancy sequence, was to exploit its pseudo-random nature and its ability to cover more values along one dimension of the search space, addressing the potential low effective dimensionality. Thus, we see Sobol sequences as a way to get best of both worlds of grid search and random search. We further highlight that Sobol sequences are a well-known strategy for initial designs in optimizers such as Bayesian Optimization.\n\n\n[1] Engstrom, Logan, et al. \"Implementation matters in deep policy gradients: A case study on ppo and trpo.\" arXiv preprint arXiv:2005.12729 (2020).\n[2] Huang, Shengyi, et al. \"The 37 implementation details of proximal policy optimization.\" The ICLR Blog Track 2023 (2022).\n[3] Bergstra, James, and Yoshua Bengio. \"Random search for hyper-parameter optimization.\" Journal of machine learning research 13.2 (2012).\n\n## Changelog\nAs a result of the feedback received we made several updates to the paper:\n\n- Inclusion of more experimental results: PPO on Bipedal Walker\n- Updated Figure 1 and its description\n- Redesigned Figure 2 (data collection) and added a more comprehensive caption\n\nWe marked the updated text with green color.\n\n```\n", "role": "user" }, { "content": "I value the authors' response. It answers most of my concerns, and I am considering revising the rating.", "role": "assistant" } ]
Ec09TcV_HKq	AutoML 2023 Conference	Reviewer_FvT5	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_FvT5, and the following are the paper you need to review:\n```<<Ec09TcV_HKq>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper focuses on the study of dynamic landscapes of hyperparameters during the training of RL algorithms. It focuses on DQN and SAC as representative DeepRL algorithms, and Cartpole and Hopper-v3 as the target environment for the study. \nPaper introduces a pipeline for methodical collection of experiences with the hyperparameters during the \"phases\" of DeepRL training. The results are then visualized using IGPR and modality plots primarily in the paper and ILM, ICE plots shared in the appendix. \nThe paper highlights the change in optimality peaks across the phases and provides practical recommendation for HPO for DeepRL pipelines. \noverall_review: The paper provides useful insights into the landscape for HPO for DeepRL algorithms. Clearly defined experiments and shared results are valuable for the DeepRL community and will provide more concrete steps toward the realization of AutoRL. \nThe minor limitation of the study is in terms of the number and characteristics of the environments. It will be interesting to see if the results shared in the paper hold for more complicated and diverse environments (eg. discrete environment - Atari etc).\nreview_summary: Great paper with fundamental (better understanding of the evolution of HPO landscape during training) and practical (a helpful guide to plan HPO for DeepRL algorithms) impact.\nactions_required_to_increase_overall_recommendation: Further diversification of the environments under study.\nRating: 9 strong accept, excellent work\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Initial Response\nThank you very much for your positive review of our work. We are elated that you see the benefit of the methodology we present and the potential it has to help the field of AutoRL. For your comment regarding environmental diversity, we have added the experiment of PPO on BipedalWalker to demonstrate the applicability of our approach to a different problem. \nWe aim to scale this up to even more diverse environments in the future.\n Global/general Response:\n```\nWe thank all the reviewers for their helpful feedback. We are glad that the reviewers consider our approach to be relevant to the AutoRL community by filling the gap through empirical evidence to commonly held assumptions (Reviewer 9o8e), and potentially instrumental in analyzing hyperparameters and finding better configurations (Reviewer 7fjf). We also appreciate the reviewers acknowledging the solidity of our experimental procedure (Reviewer W8u8), noticing the benefit of our methodical study of HP landscapes, and considering it a crucial step in creating a practical guide for HPO for DeepRL applications (Reviewer FvT5).\n\nWe would like to reiterate that the rationale behind our paper was to show the importance of landscape analyses and their importance to the field of AutoRL. This study is a first step in that direction, and our hope is that with this, we can bring the attention of larger parts of the AutoML community, a community known for adding data-driven principles to the design of ML pipelines, to work on further properties of AutoRL. We believe our methodology can be built upon by future work to explore a more nuanced interplay between hyperparameters, exploration, and credit assignment in RL. \n\nWhile we will address individual points as separate replies, we would like to address some general points in this comment. \n## Additional Experiments\nWe additionally evaluated PPO for three phases on the Bipedal-Walker-v2 environment, which can be found in the results section. For this experiment, we varied the learning rate, discount factor, and the generalized advantage estimate factor (gae_lambda). The plots show that PPO is able to achieve higher returns in very specific hyperparameter regions as opposed to SAC which achieves a generally stable performance across a wide range. This adds to previous arguments about PPO being highly sensitive to code-level optimizations [1,2].\n\nWe believe that, along the same lines as DQN and SAC, these results demonstrate that our methodology adds principled analyses of RL algorithms that can be potentially applied to understand the nuances between dynamic hyperparameter changes and the learning dynamics of the algorithms that they impact. \n\nOur general comments regarding multi-fidelity optimization still hold. However, we would additionally like to point out that these results, the generated meta-data and the approach itself are meant as a gateway to more interesting analyses that can be possible using our methodology. Along this line, we believe work in understanding learning dynamics and exploration behavior of the algorithms is a promising and crucial future direction. \n\n## Sampling Strategy\n\nWhile grid search is potentially more interpretable, we don't know whether dimensions are correlated [3]. This creates issues since a simple grid would be inefficient in effectively covering the search space and potentially hampers the scalability of our approach. \n\nOur rationale for choosing Sobol sequences, as a low-discrepancy sequence, was to exploit its pseudo-random nature and its ability to cover more values along one dimension of the search space, addressing the potential low effective dimensionality. Thus, we see Sobol sequences as a way to get best of both worlds of grid search and random search. We further highlight that Sobol sequences are a well-known strategy for initial designs in optimizers such as Bayesian Optimization.\n\n\n[1] Engstrom, Logan, et al. \"Implementation matters in deep policy gradients: A case study on ppo and trpo.\" arXiv preprint arXiv:2005.12729 (2020).\n[2] Huang, Shengyi, et al. \"The 37 implementation details of proximal policy optimization.\" The ICLR Blog Track 2023 (2022).\n[3] Bergstra, James, and Yoshua Bengio. \"Random search for hyper-parameter optimization.\" Journal of machine learning research 13.2 (2012).\n\n## Changelog\nAs a result of the feedback received we made several updates to the paper:\n\n- Inclusion of more experimental results: PPO on Bipedal Walker\n- Updated Figure 1 and its description\n- Redesigned Figure 2 (data collection) and added a more comprehensive caption\n\nWe marked the updated text with green color.\n\n```\n", "role": "user" }, { "content": "Thanks for the authors' response. The response addresses most of my concerns. I remain supportive of my positive recommendation for this paper.", "role": "assistant" } ]
Ec09TcV_HKq	AutoML 2023 Conference	Reviewer_9o8e	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_9o8e, and the following are the paper you need to review:\n```<<Ec09TcV_HKq>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper provides an empirical analysis of the hyperparameter landscapes for AutoRL. The authors collect a dataset using DQN and SAC on two RL environments that comprises trajectories collected from various hyperparameter and seed combinations, and divide the RL training into phases so as to study the temporal variation of the landscapes. From these empirical results, the authors conclude that the hyperparameter landscapes vary strongly over time, which is consistent with the assumptions that various AutoRL algorithms already depend on, but such phenomenon also questions the applicability of the multi-fidelity approaches in the context of RL. The finding that many hyperparameter landscapes are highly multi-modal also further hints the complexity of the AutoRL problems.\noverall_review: ## Strengths\n- The paper addresses an important problem in AutoRL, and fills the gap of providing empirical support for an assumption widely used in the literature, and as such it is definitely relevant to the AutoRL/ML audience for this conference.\n- The methods presented are technically sound, and the findings are in general well-supported and could be useful for future AutoRL works.\n\n## Weaknesses\n- On the flip side of Strengths point 1, the paper largely stops at only confirming the assumptions held by the community but did not go further. As I mentioned in `Potential Impact`, I do expect an analysis paper like this to provide further insights especially on their implications on the existing AutoRL practices and suggestions for future AutoRL directions. I feel this is particularly true given that the analysis techniques used in this paper are not novel in their own right, and thus the resulting analyses and insights should form the bulk of this paper's contributions.\nreview_summary: My initial recommendation is a borderline reject: while the paper shows promise and is largely technically sound, the suggested revisions may be material, and thus the paper would benefit from another round of reviewing.\n\nI will reconsider the score based on the author feedback and the other reviews.\nactions_required_to_increase_overall_recommendation: Please answer my questions and address my comments especially in the `Impacts` section.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Initial Response\nThank you very much for your review. In addition to the points addressed in the general response (motivation, additional experiments and sampling) we would like to address your points in detail.\n\n## Future Work\n\nWe absolutely agree that linking the dynamic hyperparameter landscapes to different AutoRL methods and thus analyzing the adaptability capabilities is the next step for future work.\nEspecially with respect to the prevalence of multi-modality, which is also highly relevant for multi-fidelity methods. In multi-fidelity, we need to promote good hyperparameter configurations with only a very limited number of evaluations which probably exhibit major multi-modality. This is a general challenge in AutoML and far from being solved.\n\n## Determination of Phases\n\nAs a first step we chose regular intervals as you mentioned. We agree that very probably the learning dynamics vary throughout the training and probing at those points in time might yield more insightful landscapes. Following this thought, this would mean determining the learning dynamics in the static case. However, we cannot guarantee the same dynamics for our dynamic case where we change hyperparameters during training. \n\n## Surrogate Model Quality\n\nWe thank you for raising this point which we deem very useful to select the model for representing the hyperparameter landscape. We will update the paper with information about the modeling quality until the end of the rebuttal.\n\n Global/general Response:\n```\nWe thank all the reviewers for their helpful feedback. We are glad that the reviewers consider our approach to be relevant to the AutoRL community by filling the gap through empirical evidence to commonly held assumptions (Reviewer 9o8e), and potentially instrumental in analyzing hyperparameters and finding better configurations (Reviewer 7fjf). We also appreciate the reviewers acknowledging the solidity of our experimental procedure (Reviewer W8u8), noticing the benefit of our methodical study of HP landscapes, and considering it a crucial step in creating a practical guide for HPO for DeepRL applications (Reviewer FvT5).\n\nWe would like to reiterate that the rationale behind our paper was to show the importance of landscape analyses and their importance to the field of AutoRL. This study is a first step in that direction, and our hope is that with this, we can bring the attention of larger parts of the AutoML community, a community known for adding data-driven principles to the design of ML pipelines, to work on further properties of AutoRL. We believe our methodology can be built upon by future work to explore a more nuanced interplay between hyperparameters, exploration, and credit assignment in RL. \n\nWhile we will address individual points as separate replies, we would like to address some general points in this comment. \n## Additional Experiments\nWe additionally evaluated PPO for three phases on the Bipedal-Walker-v2 environment, which can be found in the results section. For this experiment, we varied the learning rate, discount factor, and the generalized advantage estimate factor (gae_lambda). The plots show that PPO is able to achieve higher returns in very specific hyperparameter regions as opposed to SAC which achieves a generally stable performance across a wide range. This adds to previous arguments about PPO being highly sensitive to code-level optimizations [1,2].\n\nWe believe that, along the same lines as DQN and SAC, these results demonstrate that our methodology adds principled analyses of RL algorithms that can be potentially applied to understand the nuances between dynamic hyperparameter changes and the learning dynamics of the algorithms that they impact. \n\nOur general comments regarding multi-fidelity optimization still hold. However, we would additionally like to point out that these results, the generated meta-data and the approach itself are meant as a gateway to more interesting analyses that can be possible using our methodology. Along this line, we believe work in understanding learning dynamics and exploration behavior of the algorithms is a promising and crucial future direction. \n\n## Sampling Strategy\n\nWhile grid search is potentially more interpretable, we don't know whether dimensions are correlated [3]. This creates issues since a simple grid would be inefficient in effectively covering the search space and potentially hampers the scalability of our approach. \n\nOur rationale for choosing Sobol sequences, as a low-discrepancy sequence, was to exploit its pseudo-random nature and its ability to cover more values along one dimension of the search space, addressing the potential low effective dimensionality. Thus, we see Sobol sequences as a way to get best of both worlds of grid search and random search. We further highlight that Sobol sequences are a well-known strategy for initial designs in optimizers such as Bayesian Optimization.\n\n\n[1] Engstrom, Logan, et al. \"Implementation matters in deep policy gradients: A case study on ppo and trpo.\" arXiv preprint arXiv:2005.12729 (2020).\n[2] Huang, Shengyi, et al. \"The 37 implementation details of proximal policy optimization.\" The ICLR Blog Track 2023 (2022).\n[3] Bergstra, James, and Yoshua Bengio. \"Random search for hyper-parameter optimization.\" Journal of machine learning research 13.2 (2012).\n\n## Changelog\nAs a result of the feedback received we made several updates to the paper:\n\n- Inclusion of more experimental results: PPO on Bipedal Walker\n- Updated Figure 1 and its description\n- Redesigned Figure 2 (data collection) and added a more comprehensive caption\n\nWe marked the updated text with green color.\n\n```\n", "role": "user" }, { "content": "Title: Thank you for the response\nI'd like to thank the authors for responding to my review, answering my questions and the additional experiments provided on the new environment. While my view of the paper improved somewhat, I still think this paper is borderline given my expectation that analysis paper like this provides deeper insights into the existing AutoRL algorithms, as I mentioned in the original review, a revision that I acknowledge is material and not necessarily feasible within the response period. I have updated the score, but I still think that the paper can be strengthened significantly from more time for a substantial revision and an additional round of review.", "role": "assistant" } ]
Ec09TcV_HKq	AutoML 2023 Conference	Reviewer_7fjf	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_7fjf, and the following are the paper you need to review:\n```<<Ec09TcV_HKq>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The article tries to improve the impact of hyperparameters on Reinforcement Learning (RL) via Automated RL (AutoRL).\nThe article proposes an approach to build and analyze hyperparameter landscapes at multiple points in time throughout training, showing strong variations over time for different RL algorithms in various environments. This supports the theory that hyperparameters should be dynamically adjusted during training and provides the potential for more insights on AutoRL problems through landscape analyses.\noverall_review: In general, I think this is an interesting paper, with clear methods and concrete results.\nI hope the authors could address my questions, especially how the RL community would concretely benefit from this landscape analysis.\nreview_summary: Based on my understanding and questions, I would rate weak acceptance.\nactions_required_to_increase_overall_recommendation: I am willing to adjust my score based on the authors' responses if they could properly address my questions above.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Initial Response\nThank you very much for your review. Please find the answers to your questions below.\n\n## Phases\n\nOur motivation for inspecting the hyperparameter landscape at different points during training stems from previous work showcasing the potential of dynamic algorithm configuration [1] as well as the empirical performance of dynamic AutoRL methods like PBT and variants [2]. \n\nRL is an especially non-stationary process as the learning data is generated by the also changing policy. Thus, in building the HP landscape, we wanted to not just cover the space of hyperparameters effectively, but also do it tractably and in a way that is intimately tied to the data generation process. Consequently, we divided the training into phases that allowed us to characterize the landscape at different training points, and we generated configurations from a policy performing best in the end. An ideal number of phases as well as the duration is indeed an interesting question and requires further investigation in our opinion, and we plan to look into this further in follow-up work. \n\n## Leveraging the landscapes for RL research\n\nOur method is a principled way to analyze the impact of HP configurations on the RL pipeline, opening up opportunities to better understand the interplay between exploration, credit assignment, and HP configurations. We believe it can be utilized by practitioners in the future to perform more in-depth studies on how hyperparameters impact SOTA methods in their optimization procedure, or how HPs impact monolithic and non-monolithic exploration strategies. Further, we see a big potential for using these trends to design better optimizers for RL. These are also the directions that we intend to follow up on.\n\n## Accuracy of Models\n\nWe will add a paragraph on the accuracy of the selected models until the end of the rebuttal.\n\n## Applicability\n\nThe landscape analysis, as proposed in the current state, is applicable to only RL pipelines. However, similar analyses have previously been performed for supervised learning settings [3], and have traditionally been a part of the convex and evolutionary optimization community [4]. Along this line, one intersection of immediate interest is the evolutionary [5] and Quality Diversity subcommunities [6] within RL since they tackle sequential decision-making through parallel optimization methods. \n\n[1] Adriaensen, Steven, et al. \"Automated dynamic algorithm configuration.\" Journal of Artificial Intelligence Research 75 (2022): 1633-1699.\n[2] Jaderberg, Max, et al. \"Population based training of neural networks.\" arXiv preprint arXiv:1711.09846 (2017).\n[3] Pushak, Yasha, and Holger Hoos. \"AutoML Loss Landscapes.\" ACM Transactions on Evolutionary Learning 2.3 (2022): 1-30.\n[4] Malan, Katherine Mary. \"A survey of advances in landscape analysis for optimisation.\" Algorithms 14.2 (2021): 40.\n[5] Salimans, Tim, et al. \"Evolution strategies as a scalable alternative to reinforcement learning.\" arXiv preprint arXiv:1703.03864 (2017).\n[6] Parker-Holder, Jack, et al. \"Effective diversity in population based reinforcement learning.\" Advances in Neural Information Processing Systems 33 (2020): 18050-18062.\n\n Global/general Response:\n```\nWe thank all the reviewers for their helpful feedback. We are glad that the reviewers consider our approach to be relevant to the AutoRL community by filling the gap through empirical evidence to commonly held assumptions (Reviewer 9o8e), and potentially instrumental in analyzing hyperparameters and finding better configurations (Reviewer 7fjf). We also appreciate the reviewers acknowledging the solidity of our experimental procedure (Reviewer W8u8), noticing the benefit of our methodical study of HP landscapes, and considering it a crucial step in creating a practical guide for HPO for DeepRL applications (Reviewer FvT5).\n\nWe would like to reiterate that the rationale behind our paper was to show the importance of landscape analyses and their importance to the field of AutoRL. This study is a first step in that direction, and our hope is that with this, we can bring the attention of larger parts of the AutoML community, a community known for adding data-driven principles to the design of ML pipelines, to work on further properties of AutoRL. We believe our methodology can be built upon by future work to explore a more nuanced interplay between hyperparameters, exploration, and credit assignment in RL. \n\nWhile we will address individual points as separate replies, we would like to address some general points in this comment. \n## Additional Experiments\nWe additionally evaluated PPO for three phases on the Bipedal-Walker-v2 environment, which can be found in the results section. For this experiment, we varied the learning rate, discount factor, and the generalized advantage estimate factor (gae_lambda). The plots show that PPO is able to achieve higher returns in very specific hyperparameter regions as opposed to SAC which achieves a generally stable performance across a wide range. This adds to previous arguments about PPO being highly sensitive to code-level optimizations [1,2].\n\nWe believe that, along the same lines as DQN and SAC, these results demonstrate that our methodology adds principled analyses of RL algorithms that can be potentially applied to understand the nuances between dynamic hyperparameter changes and the learning dynamics of the algorithms that they impact. \n\nOur general comments regarding multi-fidelity optimization still hold. However, we would additionally like to point out that these results, the generated meta-data and the approach itself are meant as a gateway to more interesting analyses that can be possible using our methodology. Along this line, we believe work in understanding learning dynamics and exploration behavior of the algorithms is a promising and crucial future direction. \n\n## Sampling Strategy\n\nWhile grid search is potentially more interpretable, we don't know whether dimensions are correlated [3]. This creates issues since a simple grid would be inefficient in effectively covering the search space and potentially hampers the scalability of our approach. \n\nOur rationale for choosing Sobol sequences, as a low-discrepancy sequence, was to exploit its pseudo-random nature and its ability to cover more values along one dimension of the search space, addressing the potential low effective dimensionality. Thus, we see Sobol sequences as a way to get best of both worlds of grid search and random search. We further highlight that Sobol sequences are a well-known strategy for initial designs in optimizers such as Bayesian Optimization.\n\n\n[1] Engstrom, Logan, et al. \"Implementation matters in deep policy gradients: A case study on ppo and trpo.\" arXiv preprint arXiv:2005.12729 (2020).\n[2] Huang, Shengyi, et al. \"The 37 implementation details of proximal policy optimization.\" The ICLR Blog Track 2023 (2022).\n[3] Bergstra, James, and Yoshua Bengio. \"Random search for hyper-parameter optimization.\" Journal of machine learning research 13.2 (2012).\n\n## Changelog\nAs a result of the feedback received we made several updates to the paper:\n\n- Inclusion of more experimental results: PPO on Bipedal Walker\n- Updated Figure 1 and its description\n- Redesigned Figure 2 (data collection) and added a more comprehensive caption\n\nWe marked the updated text with green color.\n\n```\n", "role": "user" }, { "content": "Thanks to the authors for their clarification. The response addresses most of my concerns. I still support my favorable recommendation for this manuscript.", "role": "assistant" } ]
Eyzx7rDo-JNh	AutoML 2023 Conference	Reviewer_Sh2H	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_Sh2H, and the following are the paper you need to review:\n```<<Eyzx7rDo-JNh>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper proposes an efficient multi-objective NAS method for concealed weapon detection in X-rays. The contributions are twofold. The paper first introduces computationally cheap heuristics to make genetic multi-objective NAS algorithms for this application efficient and effective. The approach is evaluated on publicly available and private datasets. Secondly the paper proposes an approach to effectively ensemble multiple architectures discovered. The results show that their approach outperforms state-of-the-art methods for concealed weapon detection. \noverall_review: Positives:\nThe application and the necessity of automl for it is well motivated \nThe paper studies the effectiveness of automl in a real-world industry application.\nExploiting zero cost proxies for multi-objective NAS is fairly novel to the best of my knowledge\n\nNegatives:\nBackground and example definitions of “threats and modifications” is nor provided.\nIt is not clear why optimizing for different proxy scores is formulated as a multi-objective problem. In traditional multi-objective NAS the objectives are often user centric (eg: hardware efficiency, performance quality). Why is optimization of different zero cost proxy scores simultaneously desirable to search for optimal architectures? How were the proxy metrics selected?\nEvaluation on a private dataset makes some results presented not reproducible. Could the authors evaluate on more publicly available benchmarks? \nThe paper does not provide details on the search space for the task. Is it a set of all the architectures/operations from table 1?\nFor the results table 1,2,3 it would be nice to evaluate across multiple search seeds. \nCould the author’s provide an intuition on why the performance gains do not translate to semantic segmentation in table 4?\nPresentation of the approach could be improved.\nreview_summary: I find the application quite interesting and well motivated. However I think the paper misses some important technical details and description of terminologies used. This makes reading the paper and deriving connections between different parts difficult. \nactions_required_to_increase_overall_recommendation: Convincing answers to my questions above could lead to an increase in score.\nRating: 4 possibly reject, but has redeeming facets\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Thank you very much for the insightful comments and useful input. Please find our answers to some of the main points.\n1. Further evaluation: To boost the generalisability we include a 10-run case study on CIFAR-10 in the updated version. We have chosen CIFAR given its popularity in NAS, which ensures a fair comparison. Please refer to the updated Appendix for the CIFAR-10 results.\n\n2. Threats and modifications: We have added a brief explanation of the different threats and modifications to the paper as requested. The threat types are similar to the ones of SIXray and the modification types refer to how the particular item has been modified. For example, since the dataset is produced by taking multiple images, let's say the two images are of a scanned baggage and there is a gun in the first and then there is no gun in the second. This will be a threat modification of \"gun\" and the modification type would be that the gun is \"removed\". The reason we can't disclose the actual modification types and threats is that this could lead to establishing a link between the industrial project and this research which we are legally not allowed for security reasons, our apologies.\n\n3. Multiple proxy scores: Thank you very much for your comment and interest in using multiple proxy scores. During the search, we use three objectives: the NASWOT, the SYNFLOW and a complexity score. The last one aims to steer the searching process towards smaller non-over-parameterised networks, rendering the discovery of faster and more compatible architectures, as shown in the results (especially in semantic segmentation). This is now clarified in the revised version. \n \n4. Proxy scores selection: We briefly outlined (page 5 top) the need for the synergy between the NASWOT score (a data-dependent score) and the SYNFLOW score (a data-agnostic measure). This explanation was admittedly too brief. The motivation behind the choice is that NASWOT was originally designed to evaluate the difference between the linear regions produced by the model on the data at hand. In contrast, SYNFLOW provides a more general overview of the network's capacity by using an input of a vector of ones, regardless of the data and hence they provide \"orthogonal\", complementary views of the system's performance. We picked the two scores since they are, in our view, the most robust scores in the literature. They consistently score highest when compared to alternatives in studies such as Krishnakumar et al.'s 2022 \"NAS-Bench-Suite-Zero\" or Abdelfattah's 2015 \"Zero-Cost Proxies For Lightweight NAS\". In addition, our suggested approach is compatible with any proxy score or an aggregated proxy of multiple scores at once. As shown in Abdelfattah's 2015 \"Zero-Cost Proxies For Lightweight NAS\", such a combination of proxy scores can lead to better results since some of them (such as SYNFLOW and NASWOT) carry information about different aspects of the network. \n\n5. Search space for the task: Thank you for your comment, we are adding these details in the Appendix of the revised version.\n\n6. Semantic segmentation performance: Thank you for pointing out the need for clarifying further the segmentation results. The reason for these results is the multi-objective nature of the process. The fact that it did not achieve the highest test IOU or F1 score is due to the optimal balance established by the model between these scores and the computational cost.\nPerformance gains translated to this problem, but the process suggested that it was optimal to slightly sacrifice accuracy to reduce the model size by several orders of magnitude. Even though these 2 metrics are indeed worse than 2 of the other models, the MEOW model is more than x50 and more than x100 times smaller than the 1st and 2nd ones, respectively, to make it much more computationally efficient and compatible with a plethora of devices and settings.\n\nMEOW is also the only approach of the selection capable of real-time inference, which is crucial for the application it was designed for. Perhaps more important, the analysis of the gap between the validation and test results reveals that the MEOW model is the most robust out of the full selection: note that the difference between the scores is less than 1%, and this is not observed with any of the other models. This property makes it significantly more trustworthy and reliable, hence, superior to the rest of the models. Moreover, if compared to similarly sized models, it does outperform them.\n\n7. Multiple seeds: All the experiments were run with multiple seeds (5). This was indicated in point (j) in the checklist, this is now clearer in the revised version. In addition, we have also run CIFAR with 10 different seeds. The scope of this paper is to present a proof-of-concept, but, interestingly, it is straightforward to run even more seeds using the provided code in the future. \n\nThank you again for your comments and we are looking forward to your views and reply. Please find also the revised manuscript.", "role": "user" }, { "content": "Thanks for the authors' clarification. It answers most of my concerns, and I will consider raising the rating score.", "role": "assistant" } ]
Eyzx7rDo-JNh	AutoML 2023 Conference	Reviewer_no5F	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_no5F, and the following are the paper you need to review:\n```<<Eyzx7rDo-JNh>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper introduces MEOW, a multi-objective NAS approach for weapons detection in X-ray images that are commonly used in security applications. The proposed approach combines neuroevolution NAS with proxy scores to find better-performing architectures in terms of training time, size/parameters for concealed weapons detection in X-ray images. NSGA-II is used to evolve architectures, and two complementary proxy scores (NASWOT and SYNFLOW) are modified to evaluate the performance of each architecture. These proxy scores are used as objectives within NSGA-II to generate an optimal and ensemble architecture from the pareto front. The MEOW approach results in a considerable speedup over other AutoML approaches, and results in a new ensemble and optimal architecture that are improve multi-class, multi-label mean prediction accuracy by 10% compared to popular CV architectures. The benefits of using MEOW are demonstrated on the publicly available SIXray dataset and the proprietary Residuals dataset. \noverall_review: The paper is very well written and easy to follow. Concepts are explained clearly and using NAS for weapons detection in X-ray images is technically interesting. THe proposed approach combined multi-obective search with proxy scores to evaluate discovered architecture performance. Experiments on two datasets resulted in accuracy improevemts and NAS search time speedups. The authors also mention that the datasets ued for experimentations are larger than those used in related works, making the speedup remarkable. These finding have implications for the broader AutoML community in developing neuroevolution NAS approaches for specific applications. \n\n\nThe negative aspects of this paper pertain to the limited number of runs (the authors used 5 runs of NSGA-II). More runs could lead to deeper understanding of performance, and differences in architectures found. \n\nreview_summary: This paper presented an approach that used multi-objective search with proxy scores as objectives for NAS. The paper also demonstrated the effectiveness of MEOW on a relatively new problem - that of concealed weapons detection in X-ray images. The results improved accuracy by 10% and led to an atleast 8x speedup compared to existing CV architectures. Furthermore, the datasets used for evaluations were larger than those used in related work. \n\nactions_required_to_increase_overall_recommendation: Please see comments above.\nRating: 8 accept, good paper\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Thank you for the detailed input and for your interest in our study. Please find below our answers to your comments.\n1. More number of runs: We fully agree with the need for computing as many runs as possible to draw reliable, statistically significant conclusions. To this end, we have now benchmarked our approach with an extra public dataset for 10 runs instead of 5, as requested. This was also done to further foster reproducibility on another public dataset, even beyond weapon detection. Given the limited time to address the points in this rebuttal, this was the only feasible public dataset to complete the experiments in time across more than five runs and in line with other reviewers' requests. Please find them in the updated Appendix A (Section 7 currently). \n\nThe reason why we did not present further runs in the previous submission was two-fold: \nFirst, the realisation that all runs provide very similar results for all datasets (less than 1\\% variance in the scores). Thus error vars (either defined by variance or standard errors of the mean) were hardly visible. Indeed with more time ahead, we can provide well over 10 runs for all datasets. Thus this is precisely a distinctive advantage of our approach w.r.t previous models: the feasibility of comprehensively employing AutoML for industrial datasets.\n\nAlso, the NAS process takes 1 GPU hour, but the retraining of the architectures also takes around the same or even more, depending on the dataset. Hence, in real-world settings (the scenario this approach is designed for), it takes 2-3 GPU hours to find the architecture and benchmark it from scratch (for selecting the one with the highest contributing hypervolume).\n\nHowever, please note that we need to retrain all produced architectures from all runs to faithfully compare them, which will require 200x (given a population size of 20 and 10 runs) more time than in a real setting. That is why, and simply due to the limited time for drafting the rebuttal and updating the study, we have focused on rerunning and including 10 runs on CIFAR-10 for this rebuttal (instead of some larger dataset). \n\nWe would also like to note that the scope of this feasibility study was to present a proof-of-concept to challenge the (sometimes implicit) hypothesis that AutoML neuroevolution NAS methods cannot outperform expert-designed ones by showing otherwise in some key use cases.\nWe thank you very much for highlighting this key aspect; we will stress this distinctive advantage more, as indicated in this and in all follow-up studies.\n\n3. Clarification of the 200x speedup. Thank you for spotting the missing explanation behind the number, and we apologise for it. We have added a reference to the works we talk about, and these two works are NSGANET (Lu et al., 2019) which uses 8 GPU days (192 GPU hours) and SqueezeNASNet (Shaw et al., 2019), which uses between 11 and 15 GPU days for discovering an optimal architecture. \n\n4. Couple of sentences describing NSGA-II. Sorry for the omission (for space limitations). We have now added some additional sentences to the beginning of the methodology to explain NSGA-II briefly.\n\n6. Legibility of the optimal architecture. We agree that the architecture is not legible. Thank you, we have now included a much larger version of it in Section 9, which will become Appendix C (Section 9 currently). Visualisations and exports of all used architectures (including the ones for the ensemble) will be uploaded to the repository. \n\nThank you again for your interest in our study and for your input; we are looking forward to hearing your views and any further requests you may have.", "role": "user" }, { "content": "I acknowledge the authors' response. Most of my concerns have been addressed. I reaffirm my positive endorsement for this paper.", "role": "assistant" } ]
Eyzx7rDo-JNh	AutoML 2023 Conference	Reviewer_SdUM	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_SdUM, and the following are the paper you need to review:\n```<<Eyzx7rDo-JNh>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper proposes a new method to generate neural network architectures for the task of concealed weapon detection from X-ray images.\nIt combines multiple efforts from computer vision, first by adopting a neural architecture search procedure that doesn't involve training the model being evaluated and thus saving a considerable amount of time.\nSecond, the paper uses different methods of assessing the quality of the evaluated model, making up a multi-objective optimization problem.\nThis optimization problem is solved with an evolutionary search strategy where each input encodes the architecture of the evaluated model.\n\nFurther, in order to make use of different solutions found during the optimization phase, an ensemble is built from the top performing models, which avoids over-committing to a specific model.\nThe experiment section shows the proposed method performs well on two weapon detection tasks, especially better than other benchmarks on rare instances depicting uncommon weapons.\noverall_review: The paper presents a promising approach to NAS for concealed weapon detection, an important problem.\nThe experiments show results in favor of the proposed approach and contain interesting insights.\n\nClarity is a concern for the paper, as a lot of details are currently missing, making it difficult to fully appreciate the proposed algorithm.\nMore setup is also needed to fully crystalize the different metrics being used to evaluated the models in the experiment section.\nAs of now, \"performance\" is quite a vague term that's being used liberally in the paper.\n\nIn terms of technical insights, the paper can benefit from further ablation studies that pinpoint what exactly makes the proposed method work.\nIs it the multi-objective approach or simply just one of the scores being optimized?\nWhat are the different ways of building the final ensemble and how does that affect performance?\nreview_summary: I think the paper contributes a reasonable approach to an important problem and presents promising results.\nHowever, more details are needed to make the submission a strong one, and I recommend a weak rejection.\nactions_required_to_increase_overall_recommendation: More clarity about the evaluation metrics being used in the experiment section will help the paper.\nI will appreciate more intuition building in Section 3, maybe via a small illustration, where the different objectives are explained.\n\nAblation studies for the multi-objective aspect will make me get a better insight.\nDoes optimizing individual objectives work just as well?\nCan we do better than picking out the top solutions during optimization to build an ensemble?\nRating: 4 possibly reject, but has redeeming facets\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Thank you very much for raising these points. Please see more details on the manuscript updates below, as requested.\n1. Performance metrics in Table 1: We provided a too succinct explanation of the metrics used in Table 1. The caption has been expanded in the revised version. In short, we used Average Precision (labelled \"AP\" in the manuscript) and mean average precision (mAP) in order for our results to be comparable to the original study of SIXray. Typically, precision is the metric used for SIXray given that most of the images are negative, and hence reducing the number of false positives is fundamental in this setting.\n\n2. Further clarity on the advantages of the multi-objective approach: We apologise for not communicating this key point more clearly. A multi-objective approach is needed to balance the network's performance and size since the efficiency and compatibility of the produced architecture can be as important as its performance in realistic settings. \n\nWe agree that ablation studies are interesting to provide an even more comprehensive view of the robustness of the approach and bring further insights. Along these lines, please note that the proposed method outperforms single-objective ones, showcasing the distinct advantage of this feasibility study. \n\nIn the revised manuscript, we have highlighted these advantages of the multi-objective results, as requested (see reworked Section 5, discussion about the segmentation, stating that results show the compelling need for considering multiple objectives). In these results, it can be observed that the process suggested that it was optimal to slightly sacrifice accuracy while reducing the model size by several orders of magnitude (the optimal MEOW model is more than x50 and more than x100 times smaller than the 1st and 2nd ones, respectively), to make it much more computationally efficient and compatible with a plethora devices and of realistic industrial settings.\n\n3. Further clarity on the ensemble approach: We fully agree that, in principle, alternative ways of combining models can also be useful. In this work, we aimed to conduct a feasibility study and establish that this might potentially be a promising research direction. As indicated, more diversity can be added for other scenarios not tested here, and we now acknowledge this in the revised version, as requested. We did not mention this aspect in the previous study since this was not needed for our benchmarks (note please that we have added a further public dataset); thank you for the input.\n\n4. Clarity in NSGA-II and other clarifications in the Introduction: We have added extra argumentation about the need for multi-objective optimisation in the Introduction and also some very brief discussion of the benefits of NSGA-II in the Related work. Please consult the revised manuscript. Thank you for raising this point and for your advice.\n\nThanks a lot also for bringing these two typographical errors to our attention; we apologise for any confusion they may have caused. Regarding the reference to \"Dimanov et al. (2022) encoding and algorithm,\" we would like to please clarify that our work was inspired by the neuroevolution algorithm presented in their work, as well as the novel encoding technique used in conjunction with this algorithm. We would like to thank you very much for your useful input, and we are looking forward to your reply and discussing any further queries. ", "role": "user" }, { "content": "Title: remaining questions\nHi authors, thanks for their responses.\n- Regarding the precision metric, wouldn’t you care about false negatives, which are more catastrophic than false positives?\n- Thanks for the clarification of the multiple objectives; my understanding of the paper is now better. Where exactly is the comparison between your approach and the single-objective methods? Perhaps an enumeration of what the specific objectives are and how they’re optimized would help clarify things\n", "role": "assistant" }, { "content": "Title: Thank you very much for your interest, reply, and follow-up questions.Of course, please find below the clarification to your queries. \n1. False negatives: We did not report the false negatives rate or the recall (this was not reported in the original SIXray paper either) because all recall values were above 90%, and thus, false negative rates were not very informative of the performance, due to the 10:1 ratio of negatives to positives. Moreover, in this domain, industry partners were much more interested in the false positive rate than the false negatives since current checks in many places are done at random intervals, and any system which would not disturb the throughput of items/people would be considered of great value. Our view fully aligns with yours though, all values should have been reported, we cannot edit the manuscript now (it is not allowed) but we can mention this in the camera-ready version.\n\n2. Different objectives: We apologise for not stressing the different objectives enough. They are the following:\n Objective 1: NASWOT score\n Objective 2: SYNFLOW score \n Objective 3: Level of complexity (number of parameters)\n\nThese are shown in Figure 2. In addition, in the revised manuscript, we also added an additional paragraph at the end of Section 4. Experimental Setup. Thank you for noticing this inconsistency, and we hope this modification in the updated manuscript has resolved this issue.\n\n3. Single objective benchmark: The reviewer makes a great point about showing the benchmark to single-objective methods explicitly. We adopted this for the revised version, shown in Figure 4 in the Appendix with our CIFAR-10 experiments, as some of the chosen reference methods are indeed single-objective. In this study, we didn't focus on single-objective methods since the main challenge in the area was on finding the optimal balance between cost and performance. However, it is perfectly feasible to adjust the underlying RAMOSS algorithm MEOW is based on such that single-objective optimisation is supported. We will complete this further benchmark and add it to the repository as requested; thank you very much for the idea. We cannot edit the manuscript further (it is locked) but rest assured that for the camera-ready version the following sentence will be added:\n\"In our experiments, we optimise for three different objectives which are (1) NASWOT score, (2) SYNFLOW score and (3) Level of complexity (as described in Dimanov et al. 2021) which measures the model's complexity in terms of the number of parameters.\"\n", "role": "user" }, { "content": "I am grateful for the opportunity to discuss these points with the author, which has clearly contributed to strengthening this work.", "role": "assistant" } ]
JQwAc91sg_x	AutoML 2023 Conference	Reviewer_7csY	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_7csY, and the following are the paper you need to review:\n```<<JQwAc91sg_x>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: While existing hyperparameter optimization (HPO) methods efficiently identify well-performing hyperparameter configurations, they often lack transparency and insights. In this paper, the authors propose a method that uses symbolic regression on meta-data obtained through Bayesian optimization (BO) during HPO. Symbolic regression allows them to get explicit formulas that quantify the relationship between hyperparameter values and model performance. This approach makes the HPO process more explainable and human-centered. \n\nSince symbolic regression directly to meta-data collected during HPO can be influenced by the sampling bias introduced by BO. The authors propose to fit the symbolic regression on the surrogate model trained during BO, which can approximate the true underlying loss landscape. \n\noverall_review: Pro\n\nThe state of the art is meticulously presented, offering a comprehensive overview of relevant literature in the field.\n\nDespite being extensive and technical, the background information is clearly articulated, making it accessible to a wide range of readers.\n\nThe proposed method is intriguing and offers several opportunities for future refinement and improvement.\n\nThe limitations of the approach are transparently highlighted, such as its focus only on numerical parameters.\n\nThe experiments use an up-to-date tool (SMAC), which enhances the reliability and validity of the study.\n\nA critical issue, the trade-off between faithfulness and interpretability, is thoroughly explored, providing a thought-provoking discussion on this topic.\n\n\nCons \n\nA more detailed description of the GP part is missing to my opinion\n\nConsiderations on landscapes analysis are still limited\n\nreview_summary: The main purpose of the paper was to make the HPO process more human-centered by proposing a practical workflow for practitioners to gain insights into the underlying HPO problem. The proposed approach uses symbolic regression to derive a closed-form expression that captures the relationship between hyperparameter values and their performance in an interpretable manner. The authors point out that simply learning a symbolic model on meta-data collected during the HPO process can be biased. To address this issue, they have proposed learning on randomly sampled configurations whose performance is approximated by the HPO process' surrogate model. The experimental evaluation demonstrates that a balance between the interpretability and faithfulness of the symbolic model can be achieved by systematically setting a parsimony hyperparameter.\n\nThis is clearly a relevant and good-quality contribution to AutoML\n\nactions_required_to_increase_overall_recommendation: No specific suggestion\nRating: 8 accept, good paper\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer 7csY\nThank you very much for your comments and feedback. \n\nWe have carefully considered your comments, and would like to clarify one point you raised. Concerning your comment “A more detailed description of the GP part is missing”: Are you referring to the part on Bayesian optimization in the background section (Section 3), or on how we use the GP surrogate to generate predictions for the randomly sampled points? Concerning the background section, we are referring to the book on GPs by Rasmussen, which readers, who are not too familiar with the topic can, can skim. Concerning the latter, we added a comment in Section 5 to clarify that to obtain the predictions for the random samples, we leverage the surrogate model’s posterior mean.\n\nThank you again for your valuable input.\n Global/general Response:\n```\nDear reviewers, \n\nthank you for your valuable feedback and questions, which we answer below in comments to each reviewer. We are currently running additional experiments with another baseline to the symbolic regression model. Furthermore, we are running experiments on how to deal with problems featuring more than two hyperparameters to optimize, along the lines of the additional workflow element in Section 4.1.3. We will alert you again, once these additional experimental results are included in the paper. However, we have already integrated all requested changes not requiring additional experiments in the current form of the manuscript. For convenience, these changes are highlighted in red. \nDear reviewers, \n\nwe would like to follow up on our recent response to your comments on our paper. If you have any further concerns or questions, we would greatly appreciate it if you could let us know at your earliest convenience. We want to ensure that we have addressed all of the questions that you raised. Thank you for your time and consideration.\nDear reviewers, \n\nas announced in our previous response, we ran additional experiments with another baseline to the symbolic regression model. We now added the results to Table 1 in Section 5.1.\nFurthermore, we provided an example on how to deal with problems featuring more than two hyperparameters to optimize in Section 5.2.\n```\n", "role": "user" }, { "content": "I'm grateful for the authors' response. It satisfactorily addresses the major issues I had. I continue to recommend this work positively.", "role": "assistant" } ]
JQwAc91sg_x	AutoML 2023 Conference	Reviewer_8gQo	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_8gQo, and the following are the paper you need to review:\n```<<JQwAc91sg_x>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper suggests an approach to make the black-box HPO process more transparent and explainable. In order to do so, the meta-data generated by a Bayesian Optimization based HPO tool (SMAC) is used as training data for a symbolic regressor. The goal of using symbolic regression is to learn the relationship between the hyperparameters and the loss value. By looking at the explicit formula discovered by symbolic regression, insights about the loss landscape of the HPO problem can be gained. However, the meta-data generated by BO-HPO tools is biased towards certain regions of the loss landscape. To avoid this, the meta-data of randomly sampled configurations is used by leveraging the Gaussian Process model learned during the BO process to approximate the loss values of randomly sampled configurations. Experimental evaluations of the HP-Bench benchmark show this approach to be favorable. \noverall_review: Pros:\n\n1. The idea to use symbolic regression for better explainability of the HPO process is a computationally cheap method to generate inisghts and could thus be widely adopted. \n2. The proposed workflow is technically sound and straight-forward to apply. \n3. Evaluations on the HPO-Bench show the approach to be able to accurately approximate the true cost values, thus the methods seems to working well.\n\nCons (Major):\n\n1. The GP baseline in Table 1 is the only baseline that the approach is compared to. What about other “simple” regression techniques? Is it obvious that linear regression or logistic regression (that are both interpretable in some sense as well) would produce worse results?\n2. The main results of the experiments, namely the formulas in Section 5.2 are not explained at all. How can these formulas help practitioners of HPO to understand the loss landscape better? What do the factors (e.g. 0.078) mean? \n\nCons (Minor): \n\n1. In the experimental evaluations just two hyperparameters for each type of machine learning model are analyzed. This is a very small subset of available (numerical) hyperparameters\nreview_summary: Overall, the idea and the proposed workflow with symbolic regression for explainability seem to be novel and well-suited for the problem. But adding comparisons to additional baselines is necessary to assess its real impact and a more detailed explanation of the final results is absolutely crucial for practitioners to use this approach. \n\nactions_required_to_increase_overall_recommendation: See review summary.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer 8gQo\nThank you very much for your comments and feedback. Below, we provide responses to each of the specific comments you raised in your review. We hope that our responses adequately address your concerns, and are happy to answer any follow-up questions.\n\n> While Figures 1 and 2 foster the understanding of the problem and the workflow, the added value from Figure 3b is limited. The sampling bias of BO is already mentioned in the text and should not require a separate figure. \n\nThe insight is indeed limited for those who are aware how prominent the sampling bias is. However, we believe that those readers who are not too aware of the problem might find this a useful visualization of the concept.\n\n> The GP baseline in Table 1 is the only baseline that the approach is compared to. What about other “simple” regression techniques? Is it obvious that linear regression or logistic regression (that are both interpretable in some sense as well) would produce worse results?\n\nThank you for the idea to compare against linear regression as an alternative to symbolic regression. We will add this as additional baseline in the paper within the next days. However, in view of the landscape analysis by Pushak and Hoos [1], we do not believe that linear regression will be a good fit in most cases. Furthermore, we believe that logistic regression is not a good fit for our case, as it estimates the odds of an event while we are interested in modeling the HPO loss landscape as a regression problem. \n\n> The main results of the experiments, namely the formulas in Section 5.2 are not explained at all. How can these formulas help practitioners of HPO to understand the loss landscape better? What do the factors (e.g. 0.078) mean?\n\nWe agree that some more explanation on the resulting formulas would be valuable for the reader and added an explanation in Section 5.2.\n\n> In the experimental evaluations just two hyperparameters for each type of machine learning model are analyzed. This is a very small subset of available (numerical) hyperparameters. \n\nIn general, the proposed approach should be agnostic to the number of hyperparameters. Nevertheless, we decided to consider a small number of hyperparameters in our explanations mainly for two reasons:\n1) Previous work [2, 3] has shown that most performance variation is attributable to just a few hyperparameters.\n2) Interpretability decreases with the number of hyperparameters involved in the resulting formula. Furthermore, with more than two hyperparameters, plotting the formula becomes more complex or impossible. \n\nIn Section 4.1.3, we added a more detailed recommendation on how to obtain useful explanations when optimizing multiple hyperparameters, and will add an example for this within the next days.\n\nWe will inform you once the open points are added to the paper. Thank you again for your valuable input.\n\n[1] Pushak, Yasha, and Holger Hoos. \"AutoML Loss Landscapes.\" ACM Transactions on Evolutionary Learning 2.3 (2022): 1-30. \n[2] Rijn, Jan N. van and Frank Hutter. “Hyperparameter Importance Across Datasets.” Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2017). \n[3] Hutter, Frank et al. “An Efficient Approach for Assessing Hyperparameter Importance.” International Conference on Machine Learning (2014).\n Global/general Response:\n```\nDear reviewers, \n\nthank you for your valuable feedback and questions, which we answer below in comments to each reviewer. We are currently running additional experiments with another baseline to the symbolic regression model. Furthermore, we are running experiments on how to deal with problems featuring more than two hyperparameters to optimize, along the lines of the additional workflow element in Section 4.1.3. We will alert you again, once these additional experimental results are included in the paper. However, we have already integrated all requested changes not requiring additional experiments in the current form of the manuscript. For convenience, these changes are highlighted in red. \nDear reviewers, \n\nwe would like to follow up on our recent response to your comments on our paper. If you have any further concerns or questions, we would greatly appreciate it if you could let us know at your earliest convenience. We want to ensure that we have addressed all of the questions that you raised. Thank you for your time and consideration.\nDear reviewers, \n\nas announced in our previous response, we ran additional experiments with another baseline to the symbolic regression model. We now added the results to Table 1 in Section 5.1.\nFurthermore, we provided an example on how to deal with problems featuring more than two hyperparameters to optimize in Section 5.2.\n```\n", "role": "user" }, { "content": "Title: Response to Authors\nI thank the authors for addressing my concerns and running additional experiments. Regarding one of my main concerns: As the authors point out correctly in the current manuscript version, linear regression indeed does not match the SR performance consistently (though it is sometimes close). \n\nI also appreciate the idea of adding an ANOVA to the workflow to include a larger number of hyperparameters in the analysis, but do have one more question regarding this new addition: Are there any new limitations that arise from the usage of an ANOVA in connection with SR? ", "role": "assistant" }, { "content": "Title: Response to Reviewer 8gQo\nThe limitations of fANOVA itself apply in principle. In particular:\n\n1) fANOVA can yield misleading results when dealing with correlated features due to extrapolation to unlikely combinations of feature values (i.e., hyperparameter values, in our case) [1]. However, in contrast to the standard setting of interpreting machine learning models, there is no inherent data distribution on the hyperparameter configuration space in the context of HPO. Thus, this limitation is no concern for the HPO setting.\n\n2) When interaction effects are present, the main effects obtained by fANOVA can be misleading, as they capture the effect of varying the hyperparameter averaging across all instantiations of all other hyperparameters [2]. However, recent findings by Pushak and Hoos [3] suggest that there are few strong interactions in HPO problems, which makes this limitations less of a concern in practice.\n\n[1] G. Hooker. Generalized functional anova diagnostics for high-dimensional functions of dependent variables. Journal of Computational and Graphical Statistics, 2007. \n[2] Hutter, Frank et al. “An Efficient Approach for Assessing Hyperparameter Importance.” International Conference on Machine Learning (2014). \n[3] Pushak, Yasha, and Holger Hoos. \"AutoML Loss Landscapes.\" ACM Transactions on Evolutionary Learning 2.3 (2022): 1-30.", "role": "user" }, { "content": "Title: Response to Authors\nI thank the authors for the quick clarification - it would be helpful to include this information in the paper (as part of the appendix) as well.\n\nAs all my concerns have been addressed I am raising my score and now recommend the acceptance of the paper. ", "role": "assistant" }, { "content": "Title: Response to Reviewer 8gQo\nThank you for your feedback. As you suggested, we added the information on the limitations of fANOVA in the appendix.", "role": "user" }, { "content": "I am grateful for the opportunity to discuss these points with the author, which has significantly improved the quality of this work.", "role": "assistant" } ]
JQwAc91sg_x	AutoML 2023 Conference	Reviewer_bYyg	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_bYyg, and the following are the paper you need to review:\n```<<JQwAc91sg_x>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper focuses on introducing explanability and interpretability to the very critical but opaque hyperparameter optimization or HPO process. To this end, the paper utilizes the symbolic regression framework to model the hyperparameter-loss relationship utilizing the meta-data obtained from any HPO execution. The paper highlights the common tradeoff in symbolic regression between expression fidelity (the ability of learned expression to fit the data) and the expression complexity. Furthermore, the paper highlights a challenge in applying the symbolic regression framework to meta-data from a Bayesian Optimization based HPO scheme, and presents the advantages and disadvantages of two ways to mitigating this challenge.\n\noverall_review: As I have detailed in the Technical Quality and Correctness section, the main strengths of this paper are the well-motivated problem and a natural, intuitive solution.\n\nHowever, as detailed in the same section, one of the main concerns with the proposed idea is that it is not clear how much interpretability a simple expression for the (surrogate or actual) loss surface brings to the HPO process given the interpretability-fidelity tradeoff and the instability of the interpreted dependencies.\nThe choice of small HPO problems (problems with just 2 hyperparameters) does not appropriately highlight the advantage of utilizing symbolic regression over existing explanability schemes based on partial dependency plots.\n\nreview_summary: Based on the above review, the proposed idea seems an interesting direction, but in my opinion, the current conclusions and explanations seem limited or not well-supported. Hence, I recommend a rejection.\n\nactions_required_to_increase_overall_recommendation: I am happy to increase my score if the authors (or other reviewers) can point out how my concerns regarding the fundamental issues with the proposed idea is not justified or if it is a result of my misunderstanding of the idea or results.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer bYyg\nThank you very much for your comments and feedback. Below, we provide responses to each of the specific comments you raised in your review. We hope that our responses adequately address your concerns, and are happy to answer any follow-up questions.\n\n > One main concern is that it seems that we might be conflating the use of an interpretable modeling of the hyperparameter-loss relationship with the interpretability or explainability of the HPO process. Even if we assume that we are able to obtain a perfectly faithful yet very low-complexity expression with Symbolic Regression (SR), this expression highlights what loss function the HPO process is estimating and optimizing for. It is not clear how this leads to an answer of the critical questions of \"why were these hyperparameters selected\".\n\nFirst of all, we would like to clarify that we do not claim that our approach answers the question of why certain hyperparameters or hyperparameter configurations were selected. However, as we detail in Section 4.2, we envision several ways how the symbolic expression can improve the understanding of the HPO process. In particular, our approach allows (i) to characterize the HPO loss landscape mathematically through a closed-form expression and (ii) allows one to plot it. We believe that these are relevant insights into the problem gained through our approach.\n\n> This issue gets more complicated when we consider the parsimony hyperparameter. It is not clear how the faithfulness or fidelity of the SR model plays into the explanation. As the complexity of the SR-induced expression increases, the interpretability decreases while the fidelity improves. In this case, as we span the tradeoff spectrum, it would seem that the explanations (as obtained via the SR-induced expression) themselves would be changing, which is counter-intuitive. How can we address the fact that the SR model does not really match the true loss landscape in our explanations?\n\nAs Figure 5 shows, the SR model trained based on the randomly sampled points evaluated by the GP model performs quite well even for a rather low complexity. Consider, for example, the XGBoost example on the blood-transfusion-service-center dataset, where we already achieve a very faithful expression with an operation count of 1-3. This coincides with previous findings on the HPO loss landscape, which indicate that HPO loss landscapes are rather benign [1] and thus, it should be feasible to model them with a low complexity of an SR model. There are certainly learner and dataset combinations where we need a higher complexity to achieve a very faithful model, as, for example, in the SVM case on the Australian dataset. However, even in that case, we can achieve a reasonable faithfulness with still a small operation count of 4-5. Correspondingly, there is no apparent need to address that the SR model will not perfectly match the true loss landscape, from our point of view, as one apparently achieves a reasonable faithfulness with a high default parsimony value in practice. Figure 5 only highlights a simple way to fine-tune the parsimony for those users who are not satisfied by the default setting.\n\n> This issue is also related to the stability issue in symbolic regression -- if different runs produce different expressions (of potentially same level of fidelity), then it is not clear which one is the right one. This issue is highlighted in the results of this paper itself -- for the same problem (hyperparameters and dataset), there are two possible relationships that SR finds; and one of them does not depend on one of the hyperparameters batch_size at all. This is a significant level of instability in the explanations. Furthermore, $\\exp(\\text{alpha})$ and $\\sqrt{\|\\sin\\text{alpha}\|}$ are very different functions -- $\\exp(\\text{alpha})$ can grow forever with alpha, while $\\sqrt{\|\\sin\\text{alpha}\|}$ is bounded with a periodicity.\n\nConcerning the stability issue in symbolic regression: There often are various explanations for a specific behavior, e.g. a multitude of different equations in a class of functions giving about the same minimum error rate in a bounded space. This is a general issue of many explanation techniques, called the Rashomon effect [2, 6], and thus also applies to the technique proposed in our paper.\nNevertheless, we argue that leveraging symbolic regression, the user can obtain a reasonable number of explanations and leverage these as a usable hypothesis space of explanations for further studies.\n\nFurthermore, one could make use of conformal prediction for quantifying the uncertainty of the different expressions [3]. We added this as possible future work in Section 6.\n\nConcerning the concrete functions shown in the paper: After re-checking the functions, we actually spotted a typo in the first function, which we corrected now. While indeed it is correct that $\\exp(\\text{alpha})$ and $\\sqrt{\|\\sin\\text{alpha}\|}$ are different functions with different behaviors for growing alpha, it needs to be considered that the resulting functions only describe the functional relationship within the bounds that were defined for the hyperparameters. Within these bounds, the relationship described by the two functions is actually quite similar. Notably, the second function does not depend on the batch size at all, indicating that alpha is more important than the batch size. In Section 5.2, we added a plot showing the importance values for both hyperparameters calculated using functional ANOVA [5], indicating that alpha is indeed by far more important than the batch size. This is quite in line with recent findings [8] and shows a useful degree of faithfulness of our approach. \n\n> Another (relatively smaller) concern is the use of HPO problems with just 2 hyperparameters. The PDP based schemes are mentioned to be only limited to studying the effect of 2 or less hyperparameters at a time. The proposed SR based scheme should be relatively agnostic to the number of hyperparameters -- this aspect of the proposed scheme should have been highlighted in the empirical evaluations with HPO problems with a large number of hyperparameters.\n\nWe agree that the proposed approach should be agnostic to the number of hyperparameters. We decided to consider a small number of hyperparameters in our explanations mainly for two reasons:\n1) Previous work [4, 5] has shown that most performance variation is attributable to just a few hyperparameters.\n2) Interpretability decreases with the number of hyperparameters involved in the resulting formula. Furthermore, with more than two hyperparameters, plotting the formula becomes more complex or impossible. \n\nIn Section 4.1.3, we have added a more detailed recommendation on how to obtain useful explanations when optimizing multiple hyperparameters, and will add an example for this within the next days.\n\n> One thing that is not clear is regarding the SR(BO-GP) variant -- If we are modeling the surrogate function, how are the predictions for the random samples selected for regions with high uncertainty? Are we just selecting the mean of the posterior or are we sampling from the posterior distribution? How would the results be affected, if at all, by these different choices.\n\nThank you for pointing out this lack of clarity in our formulation. We clarified this in the paper: To obtain the predictions for the random samples, we leverage the surrogate model’s posterior mean.\nGiven the favorable results obtained using this method, we did not attempt to sample from the posterior distribution instead. However, we assume that the additional noise introduced by this might not be helpful to learn a low-complexity SR model. Leveraging the mean of the predictive distribution can be seen as a form of label smoothing [7] in order to stabilize the learning process. We hypothesize that sampling from the posterior would lead to even more complex and varying explanations.\n\n> Moreover, one can argue that the SR expression for the surrogate model will probably be evolving through the HPO -- how does that play into the explanation?\n\nAt the moment, it does not play into the explanation at all, as we do a posthoc analysis with all available data at the end. We could also compute intermediate explanations, which would most likely be more noisy in the beginning as the meta-dataset is smaller. However, although this is an interesting idea, it is out of scope of this paper. \n\nWe will inform you once the open points are added to the paper. Thank you again for your valuable input.\n\n[1] Pushak, Yasha, and Holger Hoos. \"AutoML Loss Landscapes.\" ACM Transactions on Evolutionary Learning 2.3 (2022): 1-30. \n[2] Breiman, Leo. “Statistical modeling: The two cultures.” Quality Engineering 48 (2001): 81-82. \n[3] Angelopoulos, Anastasios Nikolas and Stephen Bates. “A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification.” ArXiv abs/2107.07511 (2021). \n[4] Rijn, Jan N. van and Frank Hutter. “Hyperparameter Importance Across Datasets.” Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2017). \n[5] Hutter, Frank et al. “An Efficient Approach for Assessing Hyperparameter Importance.” International Conference on Machine Learning (2014). \n[6] Leventi-Peetz, Anastasia-M., and Kai Weber. \"Rashomon Effect and Consistency in Explainable Artificial Intelligence (XAI).\" Proceedings of the Future Technologies Conference (FTC) 2022, Volume 1. Cham: Springer International Publishing, 2022. \n[7] Müller, Rafael, Simon Kornblith, and Geoffrey E. Hinton. \"When does label smoothing help?.\" Advances in neural information processing systems 32 (2019). \n[8] Godbole, Varun et al. (2023). Deep Learning Tuning Playbook. Version 1.\n\n Global/general Response:\n```\nDear reviewers, \n\nthank you for your valuable feedback and questions, which we answer below in comments to each reviewer. We are currently running additional experiments with another baseline to the symbolic regression model. Furthermore, we are running experiments on how to deal with problems featuring more than two hyperparameters to optimize, along the lines of the additional workflow element in Section 4.1.3. We will alert you again, once these additional experimental results are included in the paper. However, we have already integrated all requested changes not requiring additional experiments in the current form of the manuscript. For convenience, these changes are highlighted in red. \nDear reviewers, \n\nwe would like to follow up on our recent response to your comments on our paper. If you have any further concerns or questions, we would greatly appreciate it if you could let us know at your earliest convenience. We want to ensure that we have addressed all of the questions that you raised. Thank you for your time and consideration.\nDear reviewers, \n\nas announced in our previous response, we ran additional experiments with another baseline to the symbolic regression model. We now added the results to Table 1 in Section 5.1.\nFurthermore, we provided an example on how to deal with problems featuring more than two hyperparameters to optimize in Section 5.2.\n```\n", "role": "user" }, { "content": "Title: Thank you for detailed thorough response\nI thank the authors for responding to my comments in detail. The authors raise a lot of good counterpoints. For example, I appreciate the explanation regarding why the two quite different symbolic expressions might be similar is a particular range of value which of interest in the hyperparameter optimization. I also appreciate the additional clarifications and the discussion and new experiments on hyperparameter importances.\n\nI am happy to increase the score. However, I still maintain the following reservation.\nWe are still just modeling the surrogate function with a symbolic expression and then using this expression as an \"explanation for hyperparameter optimization\". Furthermore, based on the hyperparameter importance discussion, we are also just focusing on modeling the surrogate model with a symbolic expression of the hyperparameters we consider the most important. I cannot speak for how explainable partial dependency plots of the surrogate function are either, but in my opinion, I am not sure if such a symbolic expression as an explanation is providing any insights about the main problem since we are modeling the surrogate function, not the real one. And it is known that the accuracy of the surrogate model is fairly variable across the whole hyperparameter space. The symbolic regression is not taking that into account so it is not clear how faithful the symbolic regression is to the true function, and not just having low RMSE on the data generated from the surrogate model. If we are modeling the wrong thing, a succinct clean expression is not going to be the right explanation.", "role": "assistant" } ]
JQwAc91sg_x	AutoML 2023 Conference	Reviewer_PYEc	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_PYEc, and the following are the paper you need to review:\n```<<JQwAc91sg_x>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper introduces a new approach to make the HPO process more explainable via applying symbolic regression to the metadata collected with BO. To this end, the authors develop the inference of a closed-form expression that captures the hyperparameter values and its performance with controllable tradeoffs. Empirical experiments show that the proposed method and the systematic approach is able to faithfully model the HPO loss landscape with the tradeoff between faithfulness and interpretability controllable by the practitioners.\noverall_review: Overall the idea of this paper is quite interesting. The authors first observe the feature gap that hinders wider adoption of the existing HPO tools. Then, the authors propose simple yet effective approaches to provide interpretability without infeasible overhead. Further, the authors do an extensive evaluation to illustrate the effectiveness of their approach, namely how well the symbolic models provide explanation and meanwhile accurately to capture the dynamics. The downside of this work is that it is unclear how much contribution is original to the authors as opposed to just applying and trying a combination of existing algorithms and tune the approach to work for certain settings.\nreview_summary: Simple yet effective approach. The paper is well written. Clarification is needed to further explain the technical contribution and to distinguish from directly applying related work.\nactions_required_to_increase_overall_recommendation: Further explain the technical contribution and to distinguish from directly applying related work.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer PYEc\nThank you very much for your comments and feedback. Below, we provide responses to each of the specific comments you raised in your review. We hope that our responses adequately address your concerns, and are happy to answer any follow-up questions.\n\n> For example, in Sec 4.1, the authors mention that gradient-based methods work less reliably compared to traditional genetic programming works. It would be better to explain why because it does not seem intuitive enough to skip. \n\nThe gradient-based methods were brittle in some preliminary experiments we did, meaning that in quite a few cases, we could not recover simple functions, such as polynomial functions, with the approach. Thank you for pointing this out - we adapted the corresponding part of the paper to be more precise.\n\n> The downside of this work is that it is unclear how much contribution is original to the authors as opposed to just applying and trying a combination of existing algorithms and tune the approach to work for certain settings.\n\nWhile our proposed workflow leverages well-known, existing algorithms, we argue that our contributions are far beyond the simple combination of existing algorithms. Concretely, we show that learning a symbolic model naively on metadata collected during the HPO process does not yield a good explanation, due to bias caused by the HPO sampling strategy, and propose a procedure to alleviate this problem. Furthermore, we propose a systematic approach to control the trade-off between the symbolic model's faithfulness and interpretability based on a parsimony hyperparameter. In addition to the previous contributions, we now added a section (Section 4.1.3.) in which we propose steps to extend the workflow such that useful explanations can be obtained when optimizing multiple hyperparameters, for which we will add an example within the next days.\n\nWe will inform you once the open points are added to the paper. Thank you again for your valuable input.\n\n Global/general Response:\n```\nDear reviewers, \n\nthank you for your valuable feedback and questions, which we answer below in comments to each reviewer. We are currently running additional experiments with another baseline to the symbolic regression model. Furthermore, we are running experiments on how to deal with problems featuring more than two hyperparameters to optimize, along the lines of the additional workflow element in Section 4.1.3. We will alert you again, once these additional experimental results are included in the paper. However, we have already integrated all requested changes not requiring additional experiments in the current form of the manuscript. For convenience, these changes are highlighted in red. \nDear reviewers, \n\nwe would like to follow up on our recent response to your comments on our paper. If you have any further concerns or questions, we would greatly appreciate it if you could let us know at your earliest convenience. We want to ensure that we have addressed all of the questions that you raised. Thank you for your time and consideration.\nDear reviewers, \n\nas announced in our previous response, we ran additional experiments with another baseline to the symbolic regression model. We now added the results to Table 1 in Section 5.1.\nFurthermore, we provided an example on how to deal with problems featuring more than two hyperparameters to optimize in Section 5.2.\n```\n", "role": "user" }, { "content": "I appreciate the authors' reply. The response covers nearly all of the issues I raised. I continue to recommend this work for acceptance.", "role": "assistant" } ]
LV-5kHj-uV5	AutoML 2023 Conference	Reviewer_GfZK	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_GfZK, and the following are the paper you need to review:\n```<<LV-5kHj-uV5>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper takes a closer look at Differentiable Architecture Search and analyses its performance in the context of inverse problems: more specifically using cosine waves of varying magnitude, amplitude and offset, with the goal of searching for models that can recover the samples from distorted measurements. The key findings as described in the paper are 1) DAS has large variance for inverse problems, 2) selection of hyperparameters has a large influence on the performance of DAS, 3) the performance of the weight-sharing architecture used during training does not reflect the final performance of the found architecture well.\n\noverall_review: Positives:\n* The general idea of the paper is interesting: take a deeper look at the performance of differentiable NAS, using lightweight inverse problems setting.\n* The paper identifies specific points that highlight potential limitations of DAS: instability across trials, instability across HPO tuning of the NAS hyperparameters, differences between one-shot performance ranking and full re-training.\n* The paper comes with a new problem setting that could be potentially of interest to other NAS researchers thanks to being lightweight: inverse problems.\n* The paper is well-written and easy to read.\n\nNegatives:\n* The significance of the main findings - instability across trials and instability across sets of hyperparameters - may be overrated given the results presented in the paper. Disparity between one-shot performance and re-training performance appears to be investigated already in multiple studies based on Section 2: Related Work, Stability of DAS.\n* Analysis of the results may not be the most appropriate to draw the given conclusions: the paper reports max, mean and median across 75 or 100 repetitions, but reporting stds or confidence intervals may be more appropriate to judge if the main messages of the paper hold and to what extent. Min instead of max is likely to be more interesting as min corresponds to worst-case performance here.\n* Given that the general goal seems to be to find general findings about differentiable NAS methods rather than just within the context of inverse problems, some evaluation also on other problems (e.g. CIFAR-10 classification) would be helpful. Are there any parts of the analysis that would be viable also on other problems than inverse problems so that we can find how general the findings can be? Generally it would be good to provide some intuitions for how well the findings from the inverse problem transfer to other settings of interest.\n* All analysis seems to be only done on validation data - test data (especially in combination with re-training) would appear more appropriate.\n* When not using harmful operations in the search space, DAS is worse than pure random sampling within the sequential search space.\n* When using all operations in Table 2, there is only DAS and random sampling. Is it really the case there is no other simple baseline that could be evaluated here too?\n\nOther questions:\n* To what extent is including harmful operations in the search space realistic? It certainly makes sense as part of evaluation (the method should avoid those), but if there is not much hope the operations would help, then these normally would not be included in well-designed search space? If that is the case then a good ops search scenario is more realistic?\n* What are the times to run the experiments? Information is provided in Table 1, but it is missing in Table 2. Also times to run HPO for DAS would be useful to know.\n* Why are there 75 trials in Table 1 and 100 in Table 2?\n\nreview_summary: The paper raises interesting points about differentiable NAS methods, but there are questions about how these findings follow from the results provided. I’m happy to increase the rating and recommend acceptance if the discussion resolves my concerns.\n\nactions_required_to_increase_overall_recommendation: Key areas that could improve the paper:\n* Update how the results are reported in the tables - especially include stds or confidence intervals\n* Do some experiments that would give us insights how the findings transfer from DAS for simple inverse problems to more complex and realistic problem settings\nRating: 5 marginally below the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewers I and II\n\nWe appreciate the reviewer’s acknowledgment of the potential of our work and the suggestions made.\n\n\n1. For repetition of the experiments multiple times: [...] To be able to compare the stability across repetitions, reporting standard deviation or confidence interval may be more appropriate. Analysis of the results may not be the most appropriate to draw the given conclusions)\n\nWe added minimal PSNR values as well as the standard deviation to all tables in our paper. Indeed, we can now further analyze the stability across all trials with regard to worst-case scenarios. For the sequential search space, we see that the random baseline has the highest minimal value, whereas this picture changes for the non-sequential search space, in which DAS has the highest minimal value. \n\n\n 2. When applying the right hyperparameters the results appear to be quite stable, so I am not sure about the significance of saying that the selection of hyperparameters has a large influence on the performance of DAS.\n \nAlso here, we can analyze the stability of DAS with regard to worst-case scenarios in a better way, after including the minimal PSNR and the standard deviations. As we can see in Tab. 3, the standard deviation for the blur data formation is rather small for all hyperparameter sets and operation sets, whereas the minimal value for the supposedly best hyperparameter set (BOHB-Blur) is the lowest but has the highest standard deviation. Therefore, different hyperparameter sets do have a large influence on the stability of the DAS search approach.\n\n 3.Provide more details about how the validation set is constructed. All analysis seems to be only done on validation data - test data (especially in combination with re-training) would appear more appropriate.\n\n All models are tested on random 1D signals contaminated by noise that are generated on-the-fly for every validation. More details can be found in Section 4.1. Because of this sampling procedure there is no re-use of validation data and all analysis could be equally understood as being run on test data.\n\n 4. Visualization of the search space is skewed (Figure 1 and 2)\n \nWe updated Figures 1 and 2. \n\n 5. Proof-read the supplementary for typos\n \nWe have proof-read the supplementary again. \n\n 6. Table 1 typo - there is sec for runtime, but the format is min:sec - would also be good to add to Table 2\n \n We changed the runtime format, to make it more clear and also added the runtimes to all tables.\n\n7. How were the manually designed HPs selected?\nManually designed hyperparameters were chosen based on best practices in this domain. Here, they correspond broadly to the hyperparameters for DnCNN, for training-sepcific hyperparameters. DAS hyperparameters are chosen based on the original implementation.\n\n8. Figure 1 could ideally be closer to where it is mentioned\n\nWe updated that.\n\n\n9. Instability across trials and instability across sets of hyperparameters - may be overrated, already other related work investigated this.\n\nWe agree, that previous work already pointed out the rank disorder and poor test generalization using DAS for image classification task. In this paper, we focus on analyzing, if these problems also hold in other domains, i.e., inverse problems, and if changes like search space, hyperparameters, and also improvements on the random initialization (Appendix B in the revised version) improve the named problems. We can indeed see, that (i) a proper search space for the task of interest, in our case the sequential search space, influences the performance outcome substantially, and (ii) overcoming the random initialization and just searching for the operations at each layer by using a single-level optimization further improves the test generalization.\n\n10. Given that the general goal seems to be to find general findings about differentiable NAS methods rather than just within the context of inverse problems, some evaluation also on other problems would be helpful. Are there any parts of the analysis that would be viable also on other problems than inverse problems so that we can find how general the findings can be?\n\nIn general, this study shows, that the problems which are already known with DAS on image classification also transfer to this domain of inverse problems. In this paper, we analyze DAS in the domain of inverse problems on a more fundamental level and investigate if the same problem also occur here. We show, that also in this simple 1D example, we face similar problems as the domain of image classification, which is the most commonly used domain for DAS: the ranking disorder and the poor test generalization.\n\n\n\n\n\n\n\n\n11. Is it really the case there is no other simple baseline that could be evaluated here too?\n\nIn this paper, the focus is on DAS and its ability to perform one search run to find a well-performing architecture in the domain of inverse problems. In order to evaluate its ability for being a one-shot model, we focused on the most simple existing baseline. Since DAS fails in some cases to improve already over this simple baseline, we did not include other baselines. \n\n12. To what extent is including harmful operations in the search space realistic?\n\nWe consider this point relevant for several reasons, e.g.: (1) One fundamental hope is that NAS would facilitate the design of well performing neural architectures even for non ML experts, e.g.~companies in need for a solution to a particular practical problem. In a new scenario, could a non-expert confidently exclude harmful operations? (2) In terms of robustness, we would argue that a NAS approach has to handle this case of harmful operations in a predictable way. When applied to a new search space, small programming mistakes can easily turn a good operation into a poor one that should not be chosen. (3) Would it not be disappointing, from a theoretical perspective, if a methods towards optimizing a neural architecture fails to identify poor operations? To our understanding, this should be an easy task for a practically relevant tool.\n\n\n13. What are the times to run the experiments? Information is provided in Table 1, but it is missing in Table 2.\nWhy are there 75 trials in Table 1 and 100 in Table 2?\n\nWe updated Tab. 2 with the run times. Given the overall worse performance on the non-sequential search space, we conducted more runs in this search space to better analyze the general convergence behavior. \n\n14. Do some experiments that would give us insights how the findings transfer from DAS for simple inverse problems to more complex and realistic problem settings.\n\nWe additionally evaluated DAS on the Berkeley Segmentation Dataset and Benchmark (BSDS300) for image denoising in Sec. 4.4 in our updated paper. We compare DAS to the DnCNN architecture (Zhang et al., 2017) and to a random sampling baseline, where the latter random sampling approach improves over both, DAS and DnCNN. However, we want to point out here, that this study is possible since the search and evaluation times are rather small and thus allow for these exhaustive evaluations, including hyperparameter search. This study is not easily feasible for larger datasets.\n\n\nWe thank the reviewer again for the input and hope we addressed all remaining concerns. \n Global/general Response:\n```\nWe thank the reviewers for their feedback and their acknowledgment, that this paper provides an interesting idea (R GfZK) with a good study (R wSGU) on a fundamental level (R 4Z5E) for DAS in a new domain of inverse problems, being a very important topic (R 4Z5E). \n\nWe agree, that additional information about the results, such as minimal value and standard deviation, is helpful for further investigation of the stability of DAS and the random sampling-based baselines. Therefore, we included these values in each table. \n\nIn addition, we provided experimental results on image data (BSDS for image reconstruction) in which we compare DAS to random sampling and the common network for inverse problems, DnCNN, in Sec. 4.4. \n\nWe thank the reviewers for their input and hope we addressed all remaining concerns.\n```\n", "role": "user" }, { "content": "Title: Response\nThanks for the additional details, explanations and experiments, it certainly improves the quality of the paper. However, I still have some key worries about the paper, especially after the stdevs have been added.\n* Given the magnitude of the stdevs, it is quite hard to say if we can make some of the claims in the paper - e.g. “DAS improves over a random search baseline by a significant margin”.\n* Regarding the size of variance and instability: I’d agree the variance is large when including all operations, but not when including only the good ones. After adding stdevs, various sets of hyperparameters seem to give similar results, contrary to saying “different hyperparameter sets do have a large influence on the stability”.\n* Simple baselines: if random search is too expensive, you could run something simple like selecting the best model after running one (or more generally N) training iteration(s) with each candidate configuration. Maybe this would be competitive in the scenario where random search is not evaluated due to too large time costs.\n* There is also the broader issue that the paper seems to mainly identify NAS challenges already known in literature.\n\nFor now I maintain my rating, primarily because I do not find the claims in the paper sufficiently supported with the results.\n", "role": "assistant" }, { "content": "Title: Response to Reviewer\nDear Reviewer, thank you for the question. \nThe observation of large standard deviations in DAS is actually one of the strongest motivations of our study setting. As specified for example in lines 51-53 in our paper, the variance in the performance of DAS over different runs calls for an evaluation in a statistically significant way. To do so, we propose the computationally manageable signal reconstruction setting and report results over a sufficiently large amount of results. This is not reflected in the standard deviation but in the standard error (std err), which also takes the number of runs into account. \n\nFor example, consider the std err of the results from Table 1 (e.g. blur, only good ops) where we evaluate 75 runs, and let us compare DAS against Random Sampling as you suggest. For DAS, the std dev is 1.03 so the std err is ~0.12. Thus, statistically, the result of DAS will be within the interval [21.33, 21.79] with a confidence of 95%. Random Sampling with std dev of 0.77 has a std error of ~0.09. The result of Random Sampling is thus expected to be within the interval [21.96, 22.14] with confidence of 95%.\nIn fact, we think our paper will be useful to the community in highlighting exactly this need for a large number of runs, in order to prove the statistical significance of results of DAS.\n\nIf it is helpful, we can of course include the standard error instead of the standard deviation in all our tables and and the explicit 95% confidence intervals of our evaluations in the appendix.\n\nDoes this address your concern? Are there any further questions we could address?", "role": "user" }, { "content": "Title: Response\nThanks, statistical evaluation using std err would be more appropriate here and would support the claims. For simplicity it could be enough to only report mean and std error, but of course the other metrics can be included too (especially the minimum). However, even after using more appropriate statistical evaluation the other concerns remain (baselines, findings mentioned here already known in literature, ...).", "role": "assistant" } ]
LV-5kHj-uV5	AutoML 2023 Conference	Reviewer_PEcP	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_PEcP, and the following are the paper you need to review:\n```<<LV-5kHj-uV5>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper applied differentiable architecture search to inverse problems. The experiments showed problems of one-shot neural architecture search, including a large variance, the performance gap between supernet and final performance, and the importance of hyperparameters. The interesting point of this paper is the design of search spaces. Harmful operations are included to check if the search algorithm can avoid harmful operations. \noverall_review: [+] Add harmful operations to evaluate search algorithms. \n\n[-] I did not find new experimental results. Experiments are not complete and solid.\nreview_summary: This paper re-discovered problems of one-shot search. It is not necessary to use inverse problems to repeat these problems again. I hope the authors can solve these problems when they apply differentiable search to inverse problems. \nactions_required_to_increase_overall_recommendation: I do not have room to adjust my score. However, I am still happy to hear back from the authors and will reconsider it.\nRating: 2 reject, significant issues present\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer\nThank you for your feedback. To the best of our knowledge, existing work towards analyzing \"one-shot search\" using DAS has so far focused on well studied image classification problems, leading to the following limitations: (1) each search run is expensive so that the strive for significant (negative) results is heavily challenged, and (2) the search hyperparameters are already established so that an unbiased analysis can not be provided. \n\nOur paper closes this gap. \n\nAre there particular studies that you feel are not well represented in our discussion? \nWe would have been happy to clarify and include these. \nFurther, we would happily respond to additional actionable suggestions. \n\nThank you for your time and consideration.\n Global/general Response:\n```\nWe thank the reviewers for their feedback and their acknowledgment, that this paper provides an interesting idea (R GfZK) with a good study (R wSGU) on a fundamental level (R 4Z5E) for DAS in a new domain of inverse problems, being a very important topic (R 4Z5E). \n\nWe agree, that additional information about the results, such as minimal value and standard deviation, is helpful for further investigation of the stability of DAS and the random sampling-based baselines. Therefore, we included these values in each table. \n\nIn addition, we provided experimental results on image data (BSDS for image reconstruction) in which we compare DAS to random sampling and the common network for inverse problems, DnCNN, in Sec. 4.4. \n\nWe thank the reviewers for their input and hope we addressed all remaining concerns.\n```\n", "role": "user" }, { "content": "Thank you for the authors' reply. It addresses most of my concerns, and I am inclined to revise the rating upwards.", "role": "assistant" } ]
LV-5kHj-uV5	AutoML 2023 Conference	Reviewer_wSGU	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_wSGU, and the following are the paper you need to review:\n```<<LV-5kHj-uV5>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This work identifies the problem that differentiable architecture search (DAS) has mainly been studied in a very specific architectural search space for image classification. To understand how it realistically performs in a new domain, it is applied to simulated inverse problems with either a sequential or an interconnected structural space, either a good or a mixed-quality operator space, and various hyperparameter optimization strategies.\noverall_review: Positive aspects:\n\n* You did well to search across hyperparameter optimization settings, from hand-selected to various approaches of transferring hyperparameters or searching on just the search phase or also the retraining phase.\n\nNegative aspects:\n\n* What do the searched architectures actually look like? Only two such architectures are shown (and only in the supplementary material), but looking at trends of architectures across trials could help explain why DAS can't outperform baselines when only good operations are used. For example, what do the architectural parameter patterns look like over the course of search? \n\n* Why do you think random search and random sampling got the same max scores for all combinations of data generation methods and operation spaces? This seems like a red flag and should be explained, in addition to clarifying the difference between these two random baselines. From my understanding, random sampling selects and trains a single model per trial, while random search uses the same compute time as DAS to repeatedly train random architectures?\n\n* In Figure 3, the outliers seem to greatly affect the regression line and correlation score: maybe also add the line and score for the non-degenerate data points (or at least those without a degenerate search model score, since these models could in theory be pruned out)?\n\n* Make sure to include some measure of variance in your result tables. These can be qualitatively estimated from the plots, but standard practices in DAS are to report maximum, mean, and standard error of performance.\n\n* What is the scale and intuitive meaning of PSNR? This measure is not often used in AutoML papers, so describing it would help the audience interpret the results.\n\n* The only datasets are single dimensional and generated. The results would be strengthened by also using a more realistic dataset. \n\n* Computational cost is only mentioned in Table 1; total costs are not reported (although we can guess they are pretty minimal based on Table 1). 75 to 100 repetitions seems rather excessive: generally for AutoML 5-10 runs (that only differ by random seed) suffice, so you could reduce the number of repetitions in order to perform more variations of experiments, such as another dataset as suggested previously.\n\n\n\nMinor comments:\n\n* Line 83: \"smoothen\" -> \"smooth\"\n\n* Line 169: add a paragraph break between \"...beneficial by design.\" and \"The first benign...\"\n\n* Line 186: add a paragraph break between \"...convolution layer.\" and \"These two layers...\"\n\n* FIgures 1 and 2 seem to have been stretched horizontally. It would be better to regenerate these plots with the desired aspect ratio so that it is easier to read.\n\n* Table 3 caption \"hyperparameter.\" -> \"hyperparameter sets.\"\nreview_summary: This paper provides a good preliminary study on how DAS may realistically perform in a new domain not as extensively studied as image classification. However, the provided studies are only conducted on generated toy datasets, and there are few novel claims regarding problems with DAS beyond those already identified for the image classification domain in other DAS works.\n\nPost-rebuttal update: I have raised my score from 5 to 6.\nactions_required_to_increase_overall_recommendation: * Extend the analysis of already-run experiments, including computation cost and analysis of DAS architectures\n* Include a realistic dataset\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer\nWe appreciate the reviewer’s recognition of our work and the suggestions made.\n\n1. What do the searched architectures actually look like?\n\nWe added visualizations of the best and worst found architecture by DAS and random sampling for blur and only beneficial operations in Fig. 4. As we can see, random sampling's found best architecture iterates in the first 6 layers between Net and Learnable Grad, followed by only Net layers, while the best-found architecture using DAS has less Learnable Grad operations but also stacks the Net operation at the last layers. \n\n2. Why do you think random search and random sampling got the same max scores for all combinations of data generation methods and operation spaces? \n\nRandom search is an extended random sampling approach, which searches for several architectures with the same computing time as one DAS run. It does not train a random architecture repeatedly. We reran these experiments and updated the values.\n\n3. In Figure 3, the outliers seem to greatly affect the regression line and correlation score: maybe also add the line and score for the non-degenerate data points.\n\nWe included new plots in Appendix A without the degenerated results (search results, and both search and eval results). These plots show that the correlation decreases if only degenerated search outcomes are removed but increases as soon as all degenerated results are removed. \n\n4. Make sure to include some measure of variance in your result tables.\n\nThank you for pointing that out. We included the minimal value and the standard deviation for each experiment in the updated paper. \n\n5. What is the scale and intuitive meaning of PSNR? This measure is not often used in AutoML papers, so describing it would help the audience interpret the results.\n\nThe Peak signal-to-noise ratio (PSNR) is based on MSE and is a logarithmic measurement for reconstruction quality. The higher the PSNR the better is the reconstruction.\n\n6. The only datasets are single dimensional and generated. The results would be strengthened by also using a more realistic dataset.\n\nWe additionally evaluated DAS on the Berkeley Segmentation Dataset and Benchmark (BSDS300) for image denoising in Sec. 4.4 in our updated paper. We compare DAS to the DnCNN architecture (Zhang et al., 2017) and to a random sampling baseline, where the latter random sampling approach improves over both, DAS and DnCNN. However, we want to point out here that our extensive study including unbiased hyperparameter search is possible, since the search and evaluation times on the 1D data are rather small. Such a study is not easily feasible for larger datasets.\n\n\n7. Computational cost is only mentioned in Table 1; total costs are not reported (although we can guess they are pretty minimal based on Table 1). 75 to 100 repetitions seems rather excessive: generally for AutoML 5-10 runs (that only differ by random seed) suffice, so you could reduce the number of repetitions in order to perform more variations of experiments, such as another dataset as suggested previously.\n\nWe included computational costs for the other experiments in Tab. 2, Tab. 3, and Tab. 4.\nIndeed in the general AutoML setting only 5-10 runs are conducted, given rather large computation costs. In this paper, we want to analyze the convergence behavior of DAS over several trials without being biased to different better-performing seeds. Especially, since the computational costs are small, we are able to conduct this large amount of trials to investigate the convergence behavior. \n\nThank you for pointing out the typos in the minor comments, we changed them. \n\nWe thank the reviewer again for the input and hope we addressed all remaining concerns. \n Global/general Response:\n```\nWe thank the reviewers for their feedback and their acknowledgment, that this paper provides an interesting idea (R GfZK) with a good study (R wSGU) on a fundamental level (R 4Z5E) for DAS in a new domain of inverse problems, being a very important topic (R 4Z5E). \n\nWe agree, that additional information about the results, such as minimal value and standard deviation, is helpful for further investigation of the stability of DAS and the random sampling-based baselines. Therefore, we included these values in each table. \n\nIn addition, we provided experimental results on image data (BSDS for image reconstruction) in which we compare DAS to random sampling and the common network for inverse problems, DnCNN, in Sec. 4.4. \n\nWe thank the reviewers for their input and hope we addressed all remaining concerns.\n```\n", "role": "user" }, { "content": "Title: Response to Authors\nThank you for your updates and responses. While most of the \"negative aspects\" were addressed, there is still the overarching problem that this paper doesn't have many novel claims or insights beyond existing DAS papers. I will raise my score slightly.", "role": "assistant" }, { "content": "Title: question for authors\nDear authors,\nReviewer wSGU's main question here, also mentioned by reviewers GfZK and PEcP, is that of what readers should learn from your paper that they didn't already know from previous papers. Specifically, it seems to me that these 3 reviewers (and perhaps many other readers too) believe the problems you highlight are already known. While you demonstrate the problems in the scenario of toy inverse tasks more thoroughly than before, is there any sense in which toy inverse tasks are representative of the problems readers may care about? For example, do the performance of search algorithms on toy inverse tasks correlate with their performance on important tasks?\nAC\n", "role": "user" }, { "content": "Title: Contributions\nDear AC,\n\nthank you for specifying the remaining concerns on our paper.\n\nFrom the \"inverse problems\" perspective, our paper contributes the first extensive analysis of the performance of differentiable architecture search (DAS). We can show that (unlike SotA methods suggest) hybrid architectures containing learnable gradient and net operations have the potential to outperform architectures that purely consist of either learnable gradient or net operations. We also show that, in this particular (yet commonly addressed) application, there is little benefit in considering non-sequential search spaces. This observation is of high practical relevance for the design of architectures for inverse problems. Our newly added experiments on BSDS 300 confirm these findings beyond toy experiments.\n\nFrom the \"differentiable architecture search\" perspective, several of our findings have been suspected to some extent to the community. E.g., we know that DAS can be outperformed and retraining of the found architecture is required to achieve good performance. Yet, existing literature focuses on well-established search spaces, that mostly contain well-performing architectures and for which overall well-performing hyperparameters are known. As a result, the fluctuation of DAS results can only be observed to a limited extent in this scenario. The application in a different setting allows us to observe DAS under less controlled conditions: neither the search space nor suitable hyperparameters are given a priori. Please consider that this is a far more realistic setting than the classical \"image classification in CIFAR or ImageNet\" setting for which the community has been optimizing hyperparameters for years. As a result, we can make, for the first time, statistically significant observations of the behavior of DAS under varying (yet optimized) hyperparameters. To our understanding, the results are striking and challenge the understanding of DAS as a one-shot method altogether. \n\nAs a consequence, we conclude that DAS approaches should not only report results for several runs including std errors but also that the correlation between search validation accuracy and final validation accuracy has to be monitored. If this correlation is weak in a specific setting, the entire DAS paradigm is prone to failure. Yet, such analysis is lacking is SotA DAS publications. Our paper aims to create awareness to this issue in the community.\n\nLast but not least, we argue that, from a research perspective, even such aspects that have been suspected by the community (e.g. DAS is not always stable), have to be validated in a statistically significant way. Our setting allows a statistically significant analysis of DAS in a simple and computationally manageable setting and thereby closes this important gap.\n\nWe hope that our explanation could address your remaining concerns are are happy for any further discussion!\n\nBest regards,\nauthors of submission #25", "role": "assistant" } ]
LV-5kHj-uV5	AutoML 2023 Conference	Reviewer_4Z5E	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_4Z5E, and the following are the paper you need to review:\n```<<LV-5kHj-uV5>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper presents a methodical evaluation of the Differentiable Architecture search (DAS). It uses a simple 1-D inverse problem as a tractable and unbiased problem for the evaluation of the effectiveness of DAS as an AutoML technique and also specifically as a one-shot solution. The paper presents empirical studies which provide practically useful insights into DAS as a generic AutoML technique.\nIt uses Sequential and non-sequential search spaces to further understand the impact of search space design.\nPaper reveals the effectiveness of DAS in general and its limitation as single-shot solution based on empirical evidence. \noverall_review: It is a well-written paper on a very important topic. There are many AutoML techniques currently available for NAS problems, with no clear choices for specific problems. This paper attempts at understanding the issues for DAS algorithms at fundamental level. More of such studies (in breadth and scale) are needed to make progress in the field.\nreview_summary: -do-\nactions_required_to_increase_overall_recommendation: Scale the problem to at least MNIST level image datasets.\nRating: 9 strong accept, excellent work\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer \nWe appreciate the reviewer's acknowledgment of the usefulness and importance of our study to investigate DAS on a more fundamental level. \n\nScale to image datasets.\n\nWe additionally evaluated DAS on the Berkeley Segmentation Dataset and Benchmark (BSDS300) for image denoising in Sec. 4.4 in our updated paper. We compare DAS to the DnCNN architecture (Zhang et al., 2017) and to a random sampling baseline, where the latter random sampling approach improves over both, DAS and DnCNN. Important here is, that the 1D setting allows for this exhaustive study, which is not easily feasible with larger image dataset. \n\n\nWe thank the reviewer again for the input.\n\n\n Global/general Response:\n```\nWe thank the reviewers for their feedback and their acknowledgment, that this paper provides an interesting idea (R GfZK) with a good study (R wSGU) on a fundamental level (R 4Z5E) for DAS in a new domain of inverse problems, being a very important topic (R 4Z5E). \n\nWe agree, that additional information about the results, such as minimal value and standard deviation, is helpful for further investigation of the stability of DAS and the random sampling-based baselines. Therefore, we included these values in each table. \n\nIn addition, we provided experimental results on image data (BSDS for image reconstruction) in which we compare DAS to random sampling and the common network for inverse problems, DnCNN, in Sec. 4.4. \n\nWe thank the reviewers for their input and hope we addressed all remaining concerns.\n```\n", "role": "user" }, { "content": "I'm grateful for the authors' response. The response covers nearly all of the issues I raised. I continue to recommend this work for acceptance.", "role": "assistant" } ]
MeCwOxob8jfl	AutoML 2023 Conference	Reviewer_5CEF	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_5CEF, and the following are the paper you need to review:\n```<<MeCwOxob8jfl>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper discusses the use of post hoc ensembling in AutoML systems and proposes the use of a gradient-free numerical optimization method called Covariance Matrix Adaptation Evolution Strategy (CMA-ES), comparing its performance with the commonly used Greedy Ensemble Selection (GES) method. The authors apply these methods to AutoGluon on 71 classification datasets from the AutoML Benchmark and show that the choice of metric (balanced accuracy or ROC-AUC) affects whether GES or CMA-ES performs better. Specifically, for the metric balanced accuracy, CMA-ES outperforms GES, while for ROC AUC, GES outperforms CMA-ES due to overfitting. Additionally, the authors propose a method to avoid overfitting for CMA-ES, which involves a novel normalization method inspired by GES's implicit constraints during optimization. This method allows CMA-ES to perform as well as GES while keeping the ensemble size small.\noverall_review: Strengths:\n\n-The paper presents a detailed analysis of the performance of two post hoc ensembling methods, GES and CMA-ES, and compares their performance on 71 classification datasets from the AutoML Benchmark. This comparison provides valuable empirical contribution to the field of AutoML.\n-The authors propose a novel way to avoid overfitting in post hoc ensembling using CMA-ES, which involves a normalization method inspired by GES's implicit constraints during optimization. This method can improve the performance of CMA-ES on the ROC AUC metric, which is known to be prone to overfitting.\n-The paper provides insights into the effect of different validation data qualities on post hoc ensembling performance, highlighting the importance of using high-quality validation data.\n\nWeakness:\n\n-The paper does not apply the methods to other AutoML systems and limits the analysis to AutoGluon. Extending the evaluation to other AutoML systems would provide more generalizable insights.\n\n\n\nreview_summary: The paper provides a detailed comparison of two post hoc ensembling methods, GES and CMA-ES, and proposes a novel approach to avoid overfitting in post hoc ensembling using CMA-ES. Additionally, the paper provides empirical evidence to highlight the importance of using high-quality validation data for better results in post hoc ensembling. While there are some limitations to the paper, the insights provided are valuable for practitioners in the field.\nactions_required_to_increase_overall_recommendation: NA\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Clarification Regarding a Statement in Technical Quality And Correctness\nDear Reviewer 5CEF,\n\nThank you for your feedback. We request a short clarification regarding your review:\n\nYou stated under Technical Quality And Correctness: “The authors conclude that overfitting does not depend on the metric when tested on systems that use high-quality validation data.” – The “not” contradicts your previous statements. Was this intended as feedback or an error?", "role": "user" }, { "content": "Title: Clarification\nDear Authors,\n\nI apologize for the error. The \"not\" was unintended. I appreciate your attention to detail and for pointing out the inconsistency.", "role": "assistant" }, { "content": "Title: Updated Manuscript Available \n Global/general Response:\n```\nDear Reviewers, \n\nThe updated manuscript is available. All changes are colored green.\n\nIn detail, we changed the following major points, among other minor things:\n\n* We explained our choice of CMA-ES in the introduction.\n* We explicitly state our motivation for analyzing GES in Section 4. \n* We fully formalized the definitions of our normalization methods (Section 4.3). This includes an algorithmic description of our proposed approach. \n* We motivated including stacking in our final comparison. \n* We provide an additional discussion in the result section on the performance difference between balanced accuracy and ROC AUC. \n* We included an overview of the rank change from validation to test data for all compared methods in the Appendix. \n* We extended the Appendix on the comparison of normalization method w.r.t. an ablation study and a table for the average ensemble sizes. \n* We provide a detailed description of the used statistical test and normalized improvement metric. \n* We added tables showing the mean score and standard deviation per dataset in the Appendix. \n\nWe believe these changes address the concerns raised by you. \nTherefore, we politely ask you to verify if your concerns have been resolved and reconsider your rating.\n\nThank you for your time and valuable feedback. \nWe are happy to participate in any discussions and to resolve additional concerns. \n```\n", "role": "user" }, { "content": "I truly enjoyed the above exchange with the author, which has certainly strengthened the paper.", "role": "assistant" } ]
MeCwOxob8jfl	AutoML 2023 Conference	Reviewer_Ro2y	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_Ro2y, and the following are the paper you need to review:\n```<<MeCwOxob8jfl>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper proposes CMA-ES for post-hoc ensembling in AutoML. Therefore, first, the covariance matrix adaption evolution strategy (CMA-ES) is compared to Greedy Ensemble Selection (GES) as well as the single best model. Afterwards, CMA-ES is modified for three normalization techniques: softmax, softmax and implicit GES normalization, as well as softmax and explicit GES normalization. Those modifications are compared to each other as well as the base version of CMA-ES, stacking, and single best model. Furthermore, pseudo-discrete and sparseness of weight vectors are formalized, which is hypothesized to be properties that help to avoid overfitting.\n\nThe evaluation is done based on 10-fold cross-validations of 71 classification datasets from the AutoML benchmark of AutoGlueon, comparing the performance of balanced accuracy as well as ROC AUC both for binary and multi-class classification. The authors report that CMA-ES can prevent overfitting, and that it performs similary or better than GES.\noverall_review: ## Positive\n\n- General idea of improving ensembling to prevent overfitting due to normalization is interesting.\n\n## Negative (see comments in other fields)\n\n- The correctness and quality of the presentation of experimental results is low and needs improvements. Furthermore, the conclusions need to be drawn with care.\n- The clarity of the paper as well as motivation and precise descriptions needs to be improved.\n- The contribution and novelty of the paper is very limited.\n\nreview_summary: The paper covers an interesting idea, but unfortunately not carefully written and analyzed. Especially the presentation of the experimental results is missing important information. Overall, I am wondering about the novelty of this paper due to the limited contribution. Additionally, I was wondering why AutoSklearn2 has not been used for comparison.\nactions_required_to_increase_overall_recommendation: All in all the negative aspects described under correctness and clarity need to be addressed to convince me to raise my score. Additionally, I am interested in an argument why AutoSklearn2 has not been used instead of AutoSklearn(1)?\nRating: 4 possibly reject, but has redeeming facets\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Review of Ro2y\nDear Reviewer Ro2y,\n\nThank you for your feedback. The following addresses the negative aspects raised by you. As we want to start a potential discussion early, given the short timeframe of the rebuttal, we are sharing our comments now and are stating below where we plan to adjust the manuscript. The updated manuscript will then follow in the next few days.\n\n### Technical Quality And Correctness\n\n> “The captions of all tables state that there is a comparison of three methods, but only a single rank is given.”\n\nThe captions of the tables state that one method is compared to two other methods. Hence, only the results for one method are shown. We will provide tables with all rank changes and the details you requested in the updated manuscript. \n\n> “Within the text, some numbers from evaluation results are mentioned without having any table nor plot showing any of these (Section 3.4).” \n\nThe section refers to Appendix E for a figure. An additional table for the mean number of base models was out of place, considering the page limit and small amount of numbers. \n\n> “The description of the results and the significance thereof does not match my interpretation of the result presentation.”\n\nWe ask you to clarify this statement because the critical difference plots (Figures 1, 2, and 4) clearly show statistical significance using a Friedman test with the Nemeny post hoc test. How does your interpretation differ? \n\n\n### Clarity \n\n> “Furthermore, the normalized improvement in Section 4 is described, but not mathematically introduced, which I would have expected to be precise. Additionally, the penalization with -10 is not further explained nor motivated.”:\n\nAs we followed the AutoML benchmark, we did not deem it necessary to introduce it mathematically in our paper. We will add more details on this to explain the mathematical aspect. The penalization was motivated by PAR10 from algorithm selection. We only mentioned this in the code and failed to add it to the manuscript. We will update this part accordingly.\n\n> “The reason and motivation to formalize pseudo-discrete and sparse is not completely clear to me, i.e., why the authors hypothesize that both properties might help to avoid overfitting.”\n\nWe partially answer this in Section 3.2. We will add more motivation by specifically stating that the good performance of GES for ROC AUC and its wide adaption (as indicated in the introduction) motivated us to look more into GES’s properties. \n\n> “The appendix is not only used for additional information, but also for main statements of this paper, i.e., Figure 4 is referred within Section 3.4 but can be found only in the Appendix.”\n\nFigure 4 provides additional detail but no main statement, as clarified in the first two sentences of Section 3.4. We did not select the method based on performance. Moreover, we could not include this figure in the main paper due to the page limit. \n\n> “Within overall experiments in Section 4, stacking is included. There is no reason given why stacking has been used for comparison, neither why other standard ensembling techniques like bagging or boosting are not considered.”\n\nWe used Stacking since it relates to Auto-sklearn’s statement (Line 31-33). Moreover, stacking is used by H2o AutoML. Finally, stacking is a post hoc ensembling method, unlike bagging or boosting. Both require re-training of the base models and are therefore not considered post hoc ensembling but rather part of model selection by us. Moreover, neither has so far, to the best of our knowledge, been used for post hoc ensembling in any AutoML system. \n\n> “Additionally, I am interested in an argument why AutoSklearn2 has not been used instead of AutoSklearn(1)?”\n\nWe are using AutoGluon. We ask you to clarify what you are referring to with this comment. In the introduction, we referred to Auto-Sklearn 1 because that is the version from the paper we are referring to (Auto-Sklearn 2’s paper did not comment on this again). \nMoreover, Auto-Sklearn 1 is the default when using Auto-Sklearn while Auto-Sklearn 2 is in an experimental phase. \n\n\n \n\nWe are happy to follow up on any concerns you may have, especially the point regarding the significance analysis. We will also update the manuscript based on your remaining feedback. ", "role": "user" }, { "content": "Title: Highlight Changes\nDear Authors, \n\nThanks for the fast reply. I don't have any concrete response to your message so far. Nevertheless, I would like to point out, that any changes you make to the paper should be highlighted, e.g. with a different color, so that reviewing the next version will be much easier. \n\nThanks in advance!", "role": "assistant" }, { "content": "Title: Updated Manuscript Available \n Global/general Response:\n```\nDear Reviewers, \n\nThe updated manuscript is available. All changes are colored green.\n\nIn detail, we changed the following major points, among other minor things:\n\n* We explained our choice of CMA-ES in the introduction.\n* We explicitly state our motivation for analyzing GES in Section 4. \n* We fully formalized the definitions of our normalization methods (Section 4.3). This includes an algorithmic description of our proposed approach. \n* We motivated including stacking in our final comparison. \n* We provide an additional discussion in the result section on the performance difference between balanced accuracy and ROC AUC. \n* We included an overview of the rank change from validation to test data for all compared methods in the Appendix. \n* We extended the Appendix on the comparison of normalization method w.r.t. an ablation study and a table for the average ensemble sizes. \n* We provide a detailed description of the used statistical test and normalized improvement metric. \n* We added tables showing the mean score and standard deviation per dataset in the Appendix. \n\nWe believe these changes address the concerns raised by you. \nTherefore, we politely ask you to verify if your concerns have been resolved and reconsider your rating.\n\nThank you for your time and valuable feedback. \nWe are happy to participate in any discussions and to resolve additional concerns. \n```\n", "role": "user" }, { "content": "Title: Response\nDear Authors, \n\nthanks for your detailed answers and the colored changes in the PDF. A lot of valuable information has been added to the paper. Nevertheless, besides the novelty, my main concern was about the structure and presentation of the content of your paper, which has not been reworked. Therefore, I am still not convinced to raise my score. ", "role": "assistant" }, { "content": "Title: Structure of the paper\nHi reviewer Ro2y,\n\nAre you sure that the structure of the paper is so bad that it justifies a score of 3? I see that the authors did not follow your advice to chance the structure, but I also see that your main argument was that the structure of the paper is different from what you view as the standard structure of a scientific paper. That alone is not enough. Maybe the structure is even better than the standard structure? That said, to be honest, I do not think that there is a unique best structure of a paper. In fact, my own papers have very different structures depending on the circumstances. If there is little previous work, you don't need a \"previous work\" section, for example. So I wonder if you could reconsider this case. Either, you make a convincing argument for the 3 (which is fine), or you take a fresh look at the paper, review the pros and cons, and give it a new grade? At the moment, it looks a little like you're annoyed by the fact that the authors did not change the structure and that's why you insist on your grade. I really value your opinion on the paper, but you should also consider the possibility that the authors are more experienced, have put more time into the paper, or have any other good reason to proceed as they do. \n\nAll the best\nYour area chair", "role": "user" }, { "content": "Title: Regarding structure of the paper\nDear area chair,\n\nI took a closer look at the paper again. For me, the structure of this paper is not just unusual, but also irritating. I agree that there is no standard paper structure every good paper has to follow and everyone has their own style of writing. While reading the paper I was wondering multiple times where information is coming from or if I missed something, but mostly it was due to the fact that those information was presented later in the paper (or even in the appendix).\n\nThe main contributions of this paper are the application of CMA-ES to post-hoc ensembling, but more importantly, the 3 presented normalization methods to prevent overfitting. For me, the structure of the paper is not beneficial for the contributions, which is why I earlier (mistakenly) also criticized the novelty. Most importantly, the reader is extensively presented with CMA-ES post-hoc ensembling results, although that is not the main contribution. In contrast, the results of the main contribution, i.e. the normalization methods for CMA-ES for post-hoc ensembling are only very briefly discussed, but presented in the appendix. Therefore, in my opinion CMA-ES is over presented and the normalization methods are under presented and the paper is not self-contained as some of the main results are only presented in the appendix. \n\nOverall, for me, the paper was not easy to read and I had the impression that it was not carefully checked before submission, but due to mistakenly criticizing the novelty I will raise my score to a weak reject. ", "role": "assistant" } ]
MeCwOxob8jfl	AutoML 2023 Conference	Reviewer_Bj9B	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_Bj9B, and the following are the paper you need to review:\n```<<MeCwOxob8jfl>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper studies the problem of post-hoc ensembling in AutoML systems, and proposes schemes to improve upon the widely used Greedy Ensemble Selection or GES scheme. To this end, this paper considers the use of CMA-ES for ensemble selection in place of GES by posing the ensemble selection as a gradient-free numerical unconstrained optimization of the ensemble weights using validation data.\n\nThe paper empirically evaluates this use of CMA-ES for ensemble selection on AutoML classification problems, with the results highlighting that CMA-ES is able to improve upon GES when considering the balanced accuracy metric, but falls behind GES when considering the ROC-AUC metric. Based on these results, the paper then motivates the need for constraints on the ensemble weights during the numerical optimization with CMA-ES, and presents three ways of enforcing such constraints. The empirical evaluation of these constraint enforcement schemes demonstrates that CMA-ES with such constraints improves upon vanilla CMA-ES, and is able to match GES for the ROC-AUC metric. However, these constraints lead to performance degradation with the balanced accuracy metric.\n\noverall_review: ### Positive Aspects\n\nPost-hoc ensembling is a widely used technique in many AutoML toolkits. Thus, any significant improvement to this ensembling scheme can have significant practical impact. This paper demonstrates that CMA-ES is able to significantly improve over GES on the AMLB when considering the balanced-accuracy benchmark.\n\n\nThe proposed method is clearly motivated and thoroughly evaluated with the AutoML benchmark (AMLB) with 71 classification datasets with two classification metrics. The results are presented with statistical significance tests, making it suitable for drawing meaningful conclusions from the empirical evaluations. Such an elaborate evaluation is critical when addressing fundamental aspects of AutoML such as post-hoc ensembling.\n\nThe paper does a great job at highlighting the desirable properties of GES and how to incorporate such properties as constraints with CMA-ES. This was very clearly articulated.\n\n\n\n### Negative Aspects\n\nI have detailed my main technical concerns with this paper in the Technical Quality and Correctness section of the review. Among those, the main concern is the fact that the modifications made to CMA-ES to improve performance with ROC-AUC metric leads to degradation in the performance with the balanced-accuracy metric. This gives the impression that we are not clearly understanding why we are seeing improvements or degradation over GES, and that there is some implicit aspect we are missing. Some of the other technical issues such as (i) small number of base models, or (ii) the use of AutoGluon base models, or (iii) the thresholding effect with balanced accuracy, or (iv) the fact that SingleBest performs competitively to GES with balanced accuracy might all be related to the main concern that CMA-ES enhancements for ROC-AUC hurt the balanced accuracy performance.\nIt is not clear if there is any fundamental improvement that CMA-ES provides over GES in a metric-agnostic way. One fundamental aspect could have been that CMA-ES allows for negative weights, while GES restricts to non-negative weights. But the CMA-ES normalizations remove this distinguishing feature, implicitly implying that the negative weights hurt.\nAnother related issue is that, if the CMA-ES needs to be configured differently for different metrics to get improvements over GES (or even match it), then its usability is probably going to be somewhat limited relative to GES. Would this mean that if we consider other metrics (such as F1-score, or area under the precision-recall curve), we would again have to reconfigure CMA-ES ensembling?\n\nreview_summary: The paper is studying a very important problem and provides an elaborate evaluation worthy of current AutoML research. However, I feel that there are certain fundamental limitations (as described in detail) which leads me to recommend a rejection.\n\nactions_required_to_increase_overall_recommendation: I am happy to increase my score if the authors (or other reviewers) can point out how my concerns regarding the fundamental issues with the proposed idea is not justified or if it is a result of my misunderstanding of the idea.\n\nAlternately, if we can understand some fundamental reasons why certain ensembling strategies or weighing schemes help with certain metrics but can hurt other metrics, it would be a valueable contribution as it will improve the understanding of the community. But this might be out of scope within a review cycle.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Review of Bj9B\nDear Reviewer Bj9B,\n\nThank you for your feedback. The following replies to your concerns. Thereby, directly answering some of your concerns raised under negative aspects. As we want to start a potential discussion early, given the short timeframe of the rebuttal, we are sharing our comments now and are stating below where we plan to adjust the manuscript. The updated manuscript will then follow in the next few days.\n\n\n## Technical Quality And Correctness\n### Small number of base models\nFirst, to clarify, the number of base models is entirely controlled by AutoGluon.\nAutoGluon only produces a small number of base models in most cases, because it expensively trains and evaluates a configuration with n-repeated k-fold cross-validation and only has a small number of possible configurations in total. In contrast, an AutoML system, like Auto-Sklearn, uses a less expensive validation procedure (e.g., hold-out) to train and evaluates much more base models during Bayesian optimization (>>50). Moreover, such systems decided to prune to a smaller number of base models (e.g. 50); details on the reasons for this are missing in papers of AutoML systems.\n\nWe agree that this prompts an interesting research question. It would require detailed analysis, likely with additional AutoML systems, to determine the effect of fewer base models on post hoc ensembling. This is, however, out of the scope of this rebuttal. We will note in the limitations that we only evaluated our approach for one AutoML system with its specific properties and that we make no guarantees for AutoML systems with entirely different approaches to CASH. \n\nNevertheless, we think the smaller number of base models has no significant effect on our conclusion because the results of GES for ROC AUC indicate that the statement made for Auto-Sklearn, which has a large number of base models, holds – unaffected by the smaller number of base models and the entirely different approach of AutoGluon to CASH in AutoML. Moreover, AutoGluon uses GES, following Auto-Sklearn – despite the fundamental differences. \n\n### Thresholding effect\nWe would like to note that any effect that functions like calibration/changes to the threshold only holds for binary classification, as no compared ensemble method performs a schema for multi-class that is suitable for calibration (e.g., One-Vs-Rest classification). Nevertheless, we see the same trends for CMA-ES across binary and multi-class classification. \n\nMoreover, we hypothesize that post hoc ensembling methods benefit from performing something like calibration in binary classification. They reweight the impact of the probabilities such that the base models’ threshold of 0.5 might move given the ensemble’s static threshold of 0.5. But this is not exclusive to CMA-ES, GES achieves the same reweighting of base models. \n\nHence, we do not think that the thresholding effect impacts our work or conclusions. \n\n### SingleBest vs GES (for balanced accuracy) \n\nYou stated:\n> “This seems a bit concerning given that existing papers have shown that GES can significantly improve over the SingleBest.”\n\nWhich papers are you referring to? We are only aware of papers stating this, but no paper has ever shown this. Moreover, AutoML systems are usually also not evaluated on balanced accuracy (see the AutoML benchmark), which is likely why previous papers did not consider this in their statements. \n\nFigure 1 shows the mean rank, which is drastically dominated by CMA-ES such that GES and the SingleBest have the same mean rank in the end. The relative performance comparison (Figure 3 a/b), shows the (marginal) improvement of GES over the SingleBest more clearly (especially for multi-class). \n\nWe see no connection to the number of base models for this point. If this would be a reason, CMA-ES would be equally affected by it. Nevertheless, we would be thankful if you could elaborate on the last question. \n\n## CMA-ES with normalization hurts balanced accuracy performance\n\n> “For example, if \"pseudo-discrete\" and \"sparse\" properties are important, we should try to ablate the effect of each of these properties. Do the 3 variations exhibit such an ablation?”\n\nYes, the variations exhibit such ablation, but not explicitly. CMA-ES and CMA-ES-Softmax are versions without either constraint; CMA-ES-ImplicitGES satisfies “sparse”; and CMA-ES-ExplicitGES satisfies both. Moreover, as Appendix E shows, only the version that correctly satisfies both properties (CMA-ES-ExplicitGES), is significantly different from CMA-ES for multi-class and always has the highest mean rank. We will provide more details on this in Appendix E. \n\n\n> “Furthermore, the performance of CMA-ES with normalization seems to be significantly worse than that of CMA-ES with balanced accuracy. Does this mean that we would have to utilize different versions of CMA-ES ensembling for different metrics to get the best ensembling? This seems counter-intuitive.“\n\nYes, that would be a takeaway for AutoML developers. It is very common for AutoML systems to change components based on the task and metric. For instance, AutoGluon switches parameters based on the provided metric, e.g. how to calibrate the final predictions or entier methods, e.g., not using stacking for binary classification. This would be the first time someone would select a different post hoc ensembling method. But considering that this field was not explored so far and that the no-free lunch theorem likely also holds, it does not seem counter-intuitive to us. \nFinally, the increase in performance provided by CMA-ES for balanced accuracy would certainly justify such a parameterization based on the metric. Additionally, we see balanced accuracy as representative for threshold-dependent metrics (F1, ..) and ROC AUC for threshold-independent metrics (log loss,...). \n\n> “On a related note, there is no discussion why the GES properties help CMA-ES with ROC-AUC but hurt CMA-ES with balanced accuracy. What might be happening in the ensembling that leads to such different behaviours with different metrics?”\n\nThe reason is overfitting, i.e., the reason why we apply normalization. As mentioned in Section 2 and shown in Table 1, CMA-ES does overfit for ROC AUC but not for balanced accuracy. CMA-ES with normalization does not overfit while performing worse on validation and balanced accuracy – cases where there is no overfitting for CMA-ES anyways. \nAs we failed to describe this explicitly, we will provide more details and discussion on this by extending the paragraph “Normalization to Combat Overfitting” of the results section in our updated manuscript. \n\n### Use of AutoGluon base models\n\nWe selected AutoGluon because it is likely the SOTA AutoML system (see the AutoML benchmark), uses GES, and one of its key concepts is a strong validation procedure (e.g., n-repeated k-fold cross-validation). \n\nYes, the base model choices are pre-specified, and whether stacking is used or not. And what you are referring to with k-fold ensembling is what were are referring to as n-repeated k-fold cross-validation – as they build ensembles consisting of the fold models produced during cross-validation. HPO was not enabled. \n\nWe agree that this differs from, for example, Auto-sklearn’s post hoc ensembling setup. But we think the differences are not significant enough, considering that both (Auto-Sklearn and AutoGluon) use GES and are affected by overfitting for gradient-free numerical optimization (proven for AutoGluon in our paper; stated for Auto-sklearn in their paper).\n\n\n## Clarity\n\nWe will update the manuscript with your feedback to improve clarity. \n\nWe only have one minor immediate responses to clarify an open issue: An absolute rank of 1.5 occurs as CMA-ES-ExplicitGES is tied with GES. We only mentioned this in the results section. We will also explain this manifestation of the tie as part of the table. In detail, following common practices, we set the rank to the average of the ranks in case of a tie. As the methods here are tied for the first rank, resulting in taking rank 1 and 2; the rank is set to 1.5 for both. \n\n \n\nWe hope our explanations and clarifications resolved your concerns, especially as you considered them to be fundamental limitations. \nAgain, thank you for your detailed feedback and for sharing your ideas and thoughts for potential future work.\n\n\n Global/general Response:\n```\nDear Reviewers, \n\nThe updated manuscript is available. All changes are colored green.\n\nIn detail, we changed the following major points, among other minor things:\n\n* We explained our choice of CMA-ES in the introduction.\n* We explicitly state our motivation for analyzing GES in Section 4. \n* We fully formalized the definitions of our normalization methods (Section 4.3). This includes an algorithmic description of our proposed approach. \n* We motivated including stacking in our final comparison. \n* We provide an additional discussion in the result section on the performance difference between balanced accuracy and ROC AUC. \n* We included an overview of the rank change from validation to test data for all compared methods in the Appendix. \n* We extended the Appendix on the comparison of normalization method w.r.t. an ablation study and a table for the average ensemble sizes. \n* We provide a detailed description of the used statistical test and normalized improvement metric. \n* We added tables showing the mean score and standard deviation per dataset in the Appendix. \n\nWe believe these changes address the concerns raised by you. \nTherefore, we politely ask you to verify if your concerns have been resolved and reconsider your rating.\n\nThank you for your time and valuable feedback. \nWe are happy to participate in any discussions and to resolve additional concerns. \n```\n", "role": "user" }, { "content": "Title: Paper on balanced accuracy\nI am still digesting the response, but in the meantime here is the paper regarding the following comment:\n\n> “This seems a bit concerning given that existing papers have shown that GES can significantly improve over the SingleBest.”\n> \n> Which papers are you referring to? We are only aware of papers stating this, but no paper has ever shown this. Moreover, AutoML systems are usually also not evaluated on balanced accuracy (see the AutoML benchmark), which is likely why previous papers did not consider this in their statements.\n\nThe original auto-sklearn paper [R1] uses balanced error rate and shows that ensembling improves upon SingleBest -- for example see Figure 3 in the paper. Note that they compare 4 algorithms in this figure so the difference between vanilla auto-sklearn and ensembled auto-sklearn looks small but is significant on a 1-1 comparison.\n\nTo the best of my understanding, balanced error rate averages the per class error rate, and hence, is just 1 - balanced accuracy. So this is equivalently a result on balanced accuracy. Of course this evaluation pre-dates the AMLB.\n\n[R1] Feurer et al. 2015. [Efficient and Robust Automated Machine Learning](https://papers.nips.cc/paper_files/paper/2015/hash/11d0e6287202fced83f79975ec59a3a6-Abstract.html). NeurIPS. ", "role": "assistant" }, { "content": "Title: Response: Paper on balanced accuracy \nWe agree that the original auto-sklearn paper shows that GES improves over the SingleBest. Moreover, we agree that the balanced error rate is just 1 - balanced accuracy. But we disagree that the original auto-sklearn paper shows a statistically significant improvement; which we thought was what you are referring to. \n\nWe believe it is hard to tell from an average relative rank to conclude significance on a 1-1 comparison. Any of the other 2 methods could have been better than auto-sklearn + ensemble and thus worsen its average rank, but similarly, the other 2 methods could have also only been better than vanilla auto-sklearn and worsened its average rank. \n\nTherefore, we took the time to compare vanilla auto-sklearn and auto-sklearn + ensemble (the original data for Figure 3 can be found [here](https://figshare.com/articles/dataset/Efficient_and_Robust_Automated_Machine_Learning_-_Section_6/3824103)).\nWe computed the average rank over all 140 datasets when only comparing vanilla auto-sklearn and auto-sklearn + ensemble: vanilla auto-sklearn has a mean rank of $1.59$ and auto-sklearn + ensemble of $1.41$. Comparing the performances with a Wilcoxon Test (scipy's default implementation) returns a p-value of $0.095$; assuming a standard $\\alpha = 0.05$, not significant. \n\nTo additionally put our results into context, we computed the average rank of a 1-1 comparison of GES and SingleBest for our data of AutoGluon: GES ranks marginally better than SingleBest for balanced accuracy binary ($1.49$ vs. $1.51$; 41 datasets) and multi-class ($1.47$ vs. $1.53$; 30 datasets). \n\nConsidering that many of the datasets used in the original auto-sklearn paper are not considered good benchmark tasks anymore (e.g., following the requirements of the latest AutoML benchmark paper [R2]), we believe the results to be similar enough (in terms of rank and significance) to contradict your original statement that our results are concerning. \n\nFinally, we note that for subsequent benchmarks and comparisons of AutoML systems after the original auto-sklearn paper, balanced error rate/accuracy has not been evaluated again [R2, R3, R4, R5] -- which we believe to be a mistake given the differences in performance between ROC AUC and balanced accuracy found in our paper (for AutoGluon). In contrast, the latest auto-sklearn paper [R6] reports (normalized) balanced error rates for ablation and portfolio/meta-learning studies but not in comparison to other AutoML systems.\n\n \n\nReferences\n* [R1] Feurer et al. 2015. Efficient and Robust Automated Machine Learning. NeurIPS.\n* [R2] AutoML Benchmark Papers https://openml.github.io/automlbenchmark/papers.html\n* [R3] AutoMLBench: A Comprehensive Experimental Evaluation of Automated Machine Learning Frameworks, https://arxiv.org/abs/2204.08358\n* [R4] Can AutoML outperform humans? An evaluation on popular OpenML datasets using AutoML Benchmark, https://dl.acm.org/doi/abs/10.1145/3448326.3448353\n* [R5] AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data, https://arxiv.org/abs/2003.06505\n* [R6] Auto-Sklearn 2.0: Hands-free AutoML via Meta-Learning, https://arxiv.org/abs/2007.04074\n\n", "role": "user" }, { "content": "Title: Thank you for the response\nThank you for the thorough point by point response, the new empirical studies. I am happy to raise my score. The paper is covering an important topic and it makes sense to explore how we can improve over the GES, which is what this paper does. The updates in the paper make the paper much more clear, and the additional results in the supplement makes it clear how the comparison is done (beyond the summaries in the main paper).\n\nI also really appreciate the investigation into the autosklearn's comparison between SingleBest vs Ensemble. I wonder if the difference between warm-started autosklearn and warm-started+ensembled autosklearn is also similarly statistically insignificant.\n\n\nHowever I continue to have some reservations regarding the exploration considered in this paper:\n- First, the small number of base models considered for the evaluation seems a bit incomplete. I understand that this number is completely controlled by autogluon but our goal here is not the evaluate autogluon but rather a subset selection algorithm. A more complete evaluation should be considering performance with varying numbers of base models. However, such an evaluation might not be feasible in the review period (especially since I was not able to respond earlier).\n- The other reservation I have is the metric-dependent \"overfitting\" of CMAES. The results highlight that, with CMAES, the validation objective is much better than the test objective with auroc and thus we say CMAES overfits. But I don't think we have an understanding of why this phenomenon metric dependent. Shouldn't the proposed regularization help with balanced accuracy too? Why is there no overfitting with unconstrained CMAES on balanced accuracy? I don't feel like we have a good answer here. The authors claim that this phenomenon might be dependent on whether the metric is threshold based, and the response says that CMAES might do well other threshold based metrics. But that hypothesis needs to be supported. And similarly, it would good to show that unconstrained CMAES also \"overfits\" with area under the precision-recall curve metric (much like AuROC). Without these supporting evidence (or some mathematical analyses), we don't really understand what is happening here. I don't think just claiming no free lunch addresses this issue.\n\n\nAs a final comment, it would be great if the authors can point me to the autogluon paper that talks about it using GES. To the best of my understanding (of some of the papers and the various presentations), autogluon just ensembles everything they train during the given time period, with a careful (preset) selection of what gets trained when. Of course, there is a paper about how such a large model is distilled into a smaller model. But I was not aware of there being any base models selection in the ensemble.\n", "role": "assistant" }, { "content": "Title: Responses: Additional Concerns Reviewer Bj9B\nThank you for your response and reconsideration. The following provides our answers to your additional concerns.\n\n> \"I wonder if the difference between warm-started autosklearn and warm-started+ensembled autosklearn is also similarly statistically insignificant.\"\n\nHere are the results, as this only requires changing paths in our code: warm-started auto-sklearn $1.55$ vs. warm-started+ensembled auto-sklearn $1.45$ (Wilcoxon $0.087$). Note, the (comparably) poor performances of warm-started+ensembled auto-sklearn and auto-sklearn + ensemble likely follow from overfitting due to hold-out validation. On the validation scores, they are significantly different, and the SingleBest is heavily outranked. \n\n\n> \"First, the small number of base models considered for the evaluation seems a bit incomplete. I understand that this number is completely controlled by autogluon but our goal here is not the evaluate autogluon but rather a subset selection algorithm. A more complete evaluation should be considering performance with varying numbers of base models.\" \n\nWe agree that such an evaluation would be very insightful and clarify the significance of the impact of a small to large number of base models on overfitting. However, we believe that modifying the base models produced by AutoGluon (or any other AutoML system) would skew our result. Our goal was to evaluate a post hoc ensembling method for a state-of-the-art AutoML system with a good validation procedure. Therefore, following best practices (cf. AutoML benchmark, Auto-Sklearn papers, AutoGluon paper), we had to evaluate AutoML system with a time limit. This time limit is directly responsible for the number of base models, as AutoGluon (or any other AutoML system with such an expensive validation procedure) simply often won't have the time to validate more models or enough models such that we can freely choose the number of base models for our experiments. Therefore, we believe that post hoc ensembling method (i.e., GES with its weighting and selection or default CMA-ES with only weighting) must cope with the results of a time-limited AutoML system to draw realistic conclusions about their performance. Nevertheless, in future work, we would like to explore this in a \"lab-like\" setting where we can control the number of base models and ignore time constraints. \n\n\n\n> \"But I don't think we have an understanding of why this phenomenon metric dependent. [..]. I don't think just claiming no free lunch addresses this issue.\"\n\nWe agree that our work does not provide a complete answer to this question. Following the goal of our paper, our work started by trying to verify the statements of auto-sklearn for another AutoML system with a better validation procedure and ended up opening many new future research opportunities. Consequently, our work provides, to the best of our knowledge, the first evidence that overfitting is metric-dependent and a (baseline) solution for future work. Thereby, our solution provides hints at what could be the cause of this phenomenon (e.g. sparseness or continuous values) and what this could mean for real-world AutoML systems (e.g., selecting the correct ensemble method per metric due to no free lunch). \n\n> \"Shouldn't the proposed regularization help with balanced accuracy too?\"\n\nIf there is no overfitting, it should not help (see our initial comments on this). But maybe we can find a scenario in future work where there is overfitting for balanced accuracy and check if the normalization helps in this case (e.g. auto-sklearn 1 + ensemble).\n\n> \"As a final comment, it would be great if the authors can point me to the autogluon paper that talks about it using GES. [...]\" \n\nWe understand your perspective since AutoGluon usually refers to GES only as \"weighting\" or \"final stacking layer\". Erickson et al. state this in their paper [R1]. Specifically, see Figure 2 and the last paragraph of Section 2.5, which states: \"Our final stacking layer applies ensemble selection (Caruana et al., 2004) [...].\". We refer to the work by Caruana et al. as greedy ensemble selection (GES) in our paper. \n\n \n\nThank you for the thought-provoking and insightful discussions during this rebuttal.\n\n \n\nReferences\n\n* [R1] AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data, https://arxiv.org/abs/2003.06505", "role": "user" }, { "content": "I appreciate the fruitful discussion with the author, which undoubtedly enhances the strength of this paper.", "role": "assistant" } ]
O0XkkITcLJKu	AutoML 2023 Conference	Reviewer_7MEw	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_7MEw, and the following are the paper you need to review:\n```<<O0XkkITcLJKu>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper presents a framework called AutoPatch for segmenting visual anomalies. Essentially, it is a neural architecture search method that enables efficient segmentation of visual anomalies without any training. By leveraging a pre-trained supernet, a black-box optimization algorithm can directly minimize FLOPS and maximize wAP on a small validation set of anomalous examples. Finally, compelling results on the widely studied MVTec dataset are presented, demonstrating that AutoPatch outperforms the current state-of-the-art method. \noverall_review: Overall, the motivation of this paper is clear: try to balance performance and FLOPS for segmentation tasks using a weight-sharing NAS method. \n\nstrength:\n- motivation is pretty clear and the problem this paper tries to solve is important. \n- the presentation of the proposed method is clear\n\nweakness:\n- the method is a combination of weight-sharing NAS and NN score. Besides, the weight-sharing NAS part only involves inference, that is to say, this method takes a pre-trained weight-sharing supernet and directly uses it to do the multi-objective search. The novelty is limited. \n- the evaluation is also limited. For example, a comparison with model compression technique is encouraging, although the result will be a single point in the Pareto front. \nreview_summary: Given the novelty and empirical evaluation is limited, the reviewer does not support acceptance of this paper. \nactions_required_to_increase_overall_recommendation: - discuss more about the proposed metric\n- the method is very sensitive to the random seed ( please correct the reviewer if the understanding is wrong), which makes the proposed method less reliable\nRating: 2 reject, significant issues present\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer 7MEw\nDear Reviewer 7MEw,\n\nWe greatly appreciate the time and effort you put into reviewing our paper, and we value your insightful comments. We have carefully taken into account each point raised and have made revisions accordingly.\n\nRegarding the Metric: Your observation about the need for a more in-depth discussion of the newly introduced region-weighted Average Precision is well taken. We initially included a comparison with existing metrics in the Appendix, recognizing the importance of this, we have relocated this comparison to Section 5.4 in our revised paper, and have provided a more comprehensive analysis.\n\nAddressing Novelty: We acknowledge that our method primarily involves inference. The novelty, however, resides in the application of NAS to the visual anomaly segmentation problem. Specifically, we have innovated by using a pre-trained supernet, substantially reducing the computational load usually associated with NAS. This paper presents the first application of NAS to this demanding task. Additional innovative aspects include the distinctive approach of searching for each stage, extracting block and patch size, and setting the patch size per extraction block, rather than a singular patch size. Moreover, our dynamic depth feature enables further reduction of computational complexity. We also emphasize the value of our method's simplicity, as it can be effortlessly adapted and applied to real-world datasets.\n\nConcerning Limited Evaluation: We concur that the evaluation in the initial version of the paper was limited. To address this, we have broadened our experiments in the revised version. These include:\n\n- A comparison with PatchCore using a MobileNetV3-Large backbone (Section 5.3)\n- An evaluation utilizing different metrics (this content has been relocated from the Appendix to Section 5.4)\n- In-depth analysis of the architectures discovered by our method for each category (Section 5.5)\n- An ablation study on the search space (this content has been relocated from the Appendix to Section 5.6)\n\nIn terms of Seed Sensitivity: We agree that our method displays sensitivity to the random seed. We have explicitly highlighted this as a limitation of our method in the revised paper and suggested for future work to minimize this sensitivity through augmentation and synthesis of additional anomalous examples. Additionally, we noted that the search method MOTPE exhibits a variance of one percentage point in the oracle evaluation setting, indicating that the search algorithm itself could potentially be improved. While this improvement is beyond the scope of this paper, we aim to address it in future work.\n\nWe trust that these revisions adequately address your concerns and enhance the quality and impact of our work.\n\nThank you once again for your valuable feedback.", "role": "user" }, { "content": "I value the authors' response. It addresses most of my concerns, and I may consider adjusting the rating upwards.", "role": "assistant" } ]
O0XkkITcLJKu	AutoML 2023 Conference	Reviewer_jVhD	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_jVhD, and the following are the paper you need to review:\n```<<O0XkkITcLJKu>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper proposes a novel application of NAS to visual anomaly segmentation called AutoPatch. AutoPatch uses nearest neighbor scoring, weight-sharing NAS, and weighted average precision to provide a pareto-optimal model for image segmentation. The main novelty introduced with AutoPatch is the use of a NAS algorithm specifically tailored for visual anomaly segmentation. AutoPatch appears to improve performance in a multi-objective space, and is able to accurately identify a near-optimal neural architecture for that problem with only a few examples of the anomaly.\noverall_review: Strengths:\n\nThe authors introduce a new method for visual anomaly segmentation, through the use of a NAS algorithm searching for subnetworks from MobileNetv3, in order to generate near pareto-optimal network architectures after being given only a few examples of a visual anomaly in a dataset. A novel weighted average precision metric is being used in this space in order to create a more robust performance indicator.\n\nWeaknesses:\n\nOne minor weakness within the paper is the limited scope of experimentation. While there are good and potentially generalizable results shown on the MVTec dataset, this would be more substantial with further experimentation on other datasets.\nThis is partially addressed within the future works section along with another weakness I agree with. Synthesizing artificial anomalies could provide this angle for further experimentation. Also, further ablation studies on the generalization of the model on larger anomaly set sizes would improve the empirical strength of this paper.\n\nreview_summary: This paper introduced a novel application for NAS, and a novel metric for image anomaly segmentation. It was presented in a clear manner with robust experimentation. There is a downside in its limitation to the field of visual anomaly segmentation, but this can be overcome with the usage of weighted average precision to address class imbalance.\nactions_required_to_increase_overall_recommendation: I would recommend re-organizing the paper. Moving Figure 2 to the appendix, and moving the appendix after the citations would help improve the readability of the paper.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer jVhD\nDear Reviewer jVhD,\n\nWe sincerely appreciate your comprehensive review of our paper and your constructive feedback.\n\nRegarding the Experimental Setup: Your suggestion about refining the presentation in Section 5.1 is well received. We have replaced the bullet points with a table to better organize and communicate the information.\n\nOn the Use of Other Datasets: We agree that diversifying the datasets used for testing would be beneficial. It's worth noting that we have treated MVTec as a meta-dataset, running separate experiments for each of its 15 distinct categories. This approach effectively means we've tested our method across 15 different datasets. We understand, however, the value of further diversifying the datasets used and will reserve this for future work.\n\nAddressing Larger Anomaly Sets: You raise a valid point about the potential value of larger anomaly sets. Our focus in this paper was to handle situations where very few anomalous examples are available, not enough for training a model via conventional supervision. Yet, we acknowledge the potential improvements larger anomaly sets could bring to the robustness of our method.\n\nIn Terms of Reorganizing the Paper: We appreciate your suggestions on enhancing the readability of the paper. We have moved Figure 2 to the Appendix and relocated the Appendix itself to appear after the references as recommended.\n\nWe believe that these revisions should address your concerns and further enhance the quality of our work.\n\nThank you once again for your constructive feedback.", "role": "user" }, { "content": "Thank you for the authors' feedback. It resolves the majority of my concerns. I continue to recommend this work positively.", "role": "assistant" } ]
O0XkkITcLJKu	AutoML 2023 Conference	Reviewer_nyA3	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_nyA3, and the following are the paper you need to review:\n```<<O0XkkITcLJKu>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The authors propose AutoPatch, a NAS method that efficiently segments visual anomalies. It proposes a new metric, weighted average precision (wAP), to measure anomaly segmentation quality, and shows that AutoPatch outperforms the current state-of-the-art method with significantly fewer computational resources using only one example per anomaly type. The research question is how well Pareto optimal neural architectures can be approximated with limited anomalies. The authors conclude that while promising, reliable NAS for anomaly segmentation remains a complex challenge that depends on the specific anomalies available and can exhibit considerable variance.\noverall_review: Positives: The application of NAS on Anomaly Segmentation is interesting. The presented approach is technically sound. \nThe empirical results presented in this paper are attractive and prove that NAS methods on Anomaly Segmentation are effective.\n\nNegatives: The authors compared their work only with one previous method. Is it the only method in the literature? It is also unclear how to construct a small size of the training set with one example per anomaly type. \nreview_summary: The reviewer is confident of the review. The reviewer has reviewed the search space, algorithm, and evaluation criteria to check the correctness. The reviewer has concerns with the novelty of the approach taken to search the neural architecture. \nactions_required_to_increase_overall_recommendation: The authors could improve the paper by giving more insights into the search method used in the paper by comparing it to other search methods. \n\nThe authors should also give more details on the efficiency of the search algorithm. What is the search time? \n\nWhat is the reason for AutoPatch to obtain better performance than PatchCore? Is it because of the search space or the search method? \n\nThe authors talk about inference time multiple times in the paper. However, there is no mention of the inference time of the searched models. The inference time should also be compared with PatchCore and previously searched or manually designed models. \n\nIt is also highly recommended to show the architecture of searched models for different scenarios. The authors did not discuss how the searched architectures look or explain them.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer nyA3\nDear Reviewer nyA3,\n\nWe greatly appreciate your thoughtful review of our paper and the suggestions you've provided to improve the quality of our work.\n\nRegarding the Contributions: We concur that our search space builds upon existing works such as Once-for-All and MobileNetV3. However, our approach introduces novel aspects, such as searching for each stage, extraction block, and patch size. This allows for a more dynamic and adaptable methodology than previous methods, like PatchCore, which uses a single patch size and a complicated pooling strategy. We also innovatively utilize dynamic stage depth to further reduce computational complexity. We have attempted to clarify these points in our manuscript to highlight the originality of our approach.\n\nIn relation to Once-for-All: We agree with your point, and the statement in the paper has been amended to avoid any possible misunderstanding.\n\nAbout the Abstract: Your feedback on the abstract is valuable, and we have made the necessary revisions to ensure it stands alone without needing specific knowledge of other works.\n\nOn the Method: We agree that our method can work with other performance metrics and black-box optimization algorithms, and we have updated the method section to reflect this.\n\nIn terms of Previous Methods: To the best of our knowledge, our work is the first to directly apply neural architecture search to the task of visual anomaly segmentation/detection. Therefore, there isn't any other method that we could compare ours to in this context. We believe that our approach has opened up a new area of research in this field and look forward to more works in this direction that we can compare with in the future.\n\nAbout the Validation Set: We apologize for any confusion, and have clarified in Section 5.1 of our revised manuscript that the training set consists solely of normal images. The validation set is constructed by taking one anomalous example of each type of anomaly from the MVTec dataset. For example, for the bottle category, the k=1 validation set consists of 3 anomalous images (one broken_large, one broken_small, one contamination example) along with 19 good images. We believe that the updated table in Section 5.1 should offer a clearer understanding of how we split the data.\n\nConsidering the Novelty: The novelty resides in the application of NAS to the visual anomaly segmentation problem. Specifically, we have innovated by using a pre-trained supernet, substantially reducing the computational load usually associated with NAS. This paper presents the first application of NAS to this demanding task. We also emphasize the value of our method's simplicity, as it can be effortlessly adapted and applied to real-world datasets.\n\nRegarding the Search Time: We have included the exact average search time in Section 5.1 of our revised manuscript.\n\nOn Better Performance: We have included additional experiments to investigate why AutoPatch outperforms PatchCore:\n\n- Comparison with PatchCore with a MobileNetV3-Large backbone (Section 5.3)\n- An ablation study on the search space (moved from Appendix to Section 5.6)\n\nIn terms of Inference Time: We agree with your observation. Given that the focus of this paper is computational complexity and not inference time, we have revised the manuscript to remove mentions of inference time.\n\nAbout Found Architectures: We have added a new section (Section 5.5) in our revised manuscript to describe the exact configurations of the architectures found for each category, providing insights into the oracle architectures.\n\nWe hope that these revisions have addressed your concerns and further improved the quality of our paper. \n\nOnce again, thank you for your valuable insights.", "role": "user" }, { "content": "Title: Reply to Authors\nThe reviewer would like to thank the authors for their detailed response on the issues raised. The authors made a good job of explaining the merits of the search algorithm, such as searching for each stage, extraction block, and patch size and updating the paper accordingly. The authors made most of the changes pointed out by the reviewer. The visualization of the searched architecture gives the reader a good understanding of how the searched model looks and how it is different from the manually designed Patchcore. Therefore, the reviewer is satisfied with the detailed response given by the authors and would like to update the rating to 7 (weak accept) from 5 (borderline learning reject). \n\nThe reviewer is still skeptical regarding the original contribution of the search method. It is up to the Program Committee members to decide whether the application of NAS to a problem (visual anomaly segmentation in this case) is relevant to this conference. \n\nA few other minor issues: The references should not be referred to directly. For example, “in [12]” in line 298 should rather be written as Roth et al. [12]. A sentence should be able to read as if all the citations are removed. This style should be corrected throughout the paper. \n\n\n", "role": "assistant" } ]
O0XkkITcLJKu	AutoML 2023 Conference	Reviewer_LNNk	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_LNNk, and the following are the paper you need to review:\n```<<O0XkkITcLJKu>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: In this paper, authors applied neural architecture search to the visual anomaly segmentation problem to optimize the neural architecture for the best weighted Average Precision (wAP) performance with respect to FLOPS. The proposed AutoPatch utilizes MobileNetV3 as its backbone and optimizes various width, expansion ratio, kernel size, extraction block, and patch size values for each stage using FLOPS and wAP based Multi-Objective Tree-structured Parzen Estimator (MOTPE). The experimental results show that AutoPatch outperforms the existing state-of-the-art (SOTA) PatchCore on the MVTec dataset, demonstrating superior performance.\noverall_review: The strengths of this paper lie in finding a highly efficient neural network structure for visual anomaly segmentation using neural architecture search, as well as proposing a new metric for performance measurement and increasing the learning efficiency to enable learning with fewer anomaly samples. The proposed technique has demonstrated performance surpassing the existing state-of-the-art, PatchCore, in experimental results. However, the lack of evaluation on multiple datasets and the absence of ablation study act as limitations.\nreview_summary: Although this paper has some shortcomings in experimental design and performance validation, it is technically correct, and the proposed technique consistently demonstrates better performance compared to the comparing methods. Therefore, my recommendation is a weak accept.\nactions_required_to_increase_overall_recommendation: I think it would be recommended to include comparisons with techniques other than PatchCore, comparison experiments on different datasets, and a thorough ablation study.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer LNNk\nDear Reviewer LNNk,\n\nWe appreciate your time and effort in reviewing our paper and providing valuable feedback.\n\nRegarding Other Datasets: We agree that diversifying the datasets used for testing would be beneficial. It's worth noting that we have treated MVTec as a meta-dataset, running separate experiments for each of its 15 distinct categories. This approach effectively means we've tested our method across 15 different datasets. We understand, however, the value of further diversifying the datasets used and will reserve this for our future work.\n\nAbout Ablation Studies: We took your feedback into account and have included a more thorough ablation of the method in the revised version of the paper. We have moved the ablation study on the search space from the Appendix to Section 5.6 and have included a comparison with PatchCore using a MobileNetV3-Large backbone in Section 5.3. We believe these additions will provide a more comprehensive understanding of our method's effectiveness.\n\nConcerning Other Techniques: To the best of our knowledge, our work is the first to directly apply neural architecture search to the task of visual anomaly segmentation/detection. Therefore, there isn't any other method that we could compare ours to in this context. We believe that our approach has opened up a new area of research in this field and look forward to more works in this direction that we can compare with in the future.\n\nWe hope these responses address your concerns and further clarify the contributions of our paper. \n\nOnce again, we thank you for your insightful feedback.", "role": "user" }, { "content": "I appreciate the authors' efforts in addressing my concerns. The response addresses most of my concerns. I continue to recommend this work for acceptance.", "role": "assistant" } ]
O0XkkITcLJKu	AutoML 2023 Conference	Reviewer_63A5	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_63A5, and the following are the paper you need to review:\n```<<O0XkkITcLJKu>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: In this paper, the authors propose an architecture search method, AutoPatch, which is designed to search for effective segmentation algorithms for anomalies, especially under the scenario of only having relatively few anomalies. With this, they begin by proposing a new weighted average precision (wAP) metric to serve as an alternative to AUROC and AUPRO. The motivation for this is in some industrial applications you might not be able to set some fixed False Positive Rate (FPR) and wish to be flexible over multiple thresholds. The authors then propose an extension of Once-for-All to their scenario where they can extract features from training images, without any additional training, to help segment images during inference, and use MOTPE to maximize their proposed weighted mAP metric subject to minimize the number of flops used. The authors then analyze their method in comparison with an existing method for the problem of anomaly segmentation across different amounts of anomalies.\noverall_review: Pros:\n+ The wAP metric seems well-motivated and is an unintuitive way to tackle the problem of class imbalance under the scenario when you don't have a fixed false positive rate.\n+ The neural architecture search seems very thorough and the formulation of solving their black-box optimization scheme with MOTPE.\n+ The author's proposed method doesn't require any additional training, something that I believe is underplayed \n+ The authors do a good job of answering their proposed question of \"How close can they get to the Pareto frontier'.\n\nCons:\n- The combination of the new wAP metric alongside proposing the neural architecture search scheme slightly complicates the comparison to the baseline chosen.\n- The use of the name 'Weighted Average Precision' seems overloaded to something that is common elsewhere in the literature and refers to something different which is slightly confusing.\n- One of the motivations in section 3 is \"However, these metrics lack flexibility for diverse requirements and could potentially lead to inaccurate performance indications in the absence of an appropriate FPR limit\"; however, it's unclear from the setup that there exist such scenarios where an ~some FPR limit cannot be set that is semi-optimal.\n- The use of the memory bank here seems to be a major limitation because the amount of memory required to both store the examples and the amount of computing required to search them is linear in the amount of training data. The authors do recognize this in their limitations; however, I could not find details of the necessary size of the memory bank that was used to achieve their results and I think that might be a crucial factor in their performance.\n- Unless I misunderstood, it seems like this method is very prone to failing under domain shift, where images are at inference.\n\nreview_summary: Overall, my current recommendation for this paper is a Weak Reject. The authors in this work solve a specific problem with pretty intuitive architectural search solutions and introduce a nice metric to help address their setup of anomaly detection. The main weaknesses of the paper stem from the lack of comparison using traditional metrics to their chosen baseline and some issues with their modeling. While they don't need to include any additional training, they have a method that is bandlimited by the number of training frames you can store, which makes it also seemingly brittle to domain shift. I also strongly believe that wAP here is causing some confusion and that the authors could benefit by changing the naming. I could be misunderstanding something here about their intended use case, and do think that for their problem setup, the authors offer a strong piece of work.\nactions_required_to_increase_overall_recommendation: I am happy to work with the authors to consider increasing my score if the authors can help me better assess the quality of this paper:\n\n1. How does their method compare to their baseline using the metrics that were proposed by PatchCore?\n\n2. How many frames are really necessary to store in order to attain good performance? This seems to be a fundamental question that I didn't see answered in the main text.\n\n3. Does their method suffer from issues in generalization under even small amounts of domain shift?\n\n4. I'm curious as to the effect of the selection of examples chosen for k = 1, 2, 4. I understand that this randomization is done over seeds but am wondering if there was additional work to verify the average performance over different k for both the proposed model and the baseline. Additionally, were the same examples used between the proposed method and the baseline in the evaluation?\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer 63A5\nDear Reviewer 63A5,\n\nThank you for your review of the paper. We appreciate your feedback and have addressed your concerns in the revised version of the paper.\n\nMetric Naming: We acknowledge the confusion caused by the name \"weighted Average Precision (wAP)\" and have changed it to \"region-weighted Average Precision (rwAP)\" to prevent confusion and emphasize that weighting is done over regions. This name is generic, as the metric may indeed be useful outside of visual anomaly segmentation.\n\nFair Evaluation: In the initial version of the paper, all experiments were performed using the newly introduced region-weighted Average Precision. However, we understand your concern regarding the potential disadvantage of the compared baseline due to the new metric. Thus, we have moved the comparison table using existing metrics, including AUROC without an FPR limit as used by PatchCore, from the Appendix to Section 5.4 in the revised version, along with a more elaborate analysis. Additionally, this section highlights a scenario where a semi-optimal FPR limit makes evaluation troublesome.\n\nDomain Shift: The robustness to domain shifts of AutoPatch should be comparable to PatchCore, and it is not the main focus of this work. However, it is hypothesized that ImageNet features do show some robustness to domain shift. We refer to the following work that finds PatchCore segmentation performance is competitive under domain shifts: [https://arxiv.org/pdf/2304.02216.pdf](https://arxiv.org/pdf/2304.02216.pdf)\n\nEfficiency Analysis: We agree that the large figure with the Pareto fronts was crowded and difficult to interpret. Therefore, we have moved this figure to the Appendix and added Section 5.3 to evaluate the performance of architectures under a computational complexity constraint. This section includes an easier-to-parse table with an additional comparison to PatchCore with a MobileNetV3-Large backbone.\n\nMemory Bank: Your point about the memory bank is valid, and we acknowledge that the memory bank size is crucial to the amount of compute required. Subsampling as introduced by PatchCore allows to maintain a sufficient memory size to query from, instead of the whole training set, which is intractable for larger real-world datasets of normal examples. The results presented in this paper are without subsampling, both for PatchCore and AutoPatch. As a result, the size of the memory bank is equal to the number of patches in the training set for each category (see Section 5.1 Experimental Setup). We have included a better elaboration on the memory bank and subsampling in Section 7 Limitations and Future Work of the revised version. AutoPatch allows subsampling of the memory bank with the found architecture after completing the search. Performance is not significantly affected up to a certain subsampling ratio, as observed by PatchCore. To answer your question on how many frames are needed to attain good segmentation performance, it is between 1% to 10% of all patches from the training sets in MVTec (Figure S6 in the supplementary of PatchCore).\n\nValidation Sets: The images included in the validation set are not based on the seed, as they are not random. As described in Section 5.1 Experimental Setup, the validation sets are the first 1, 2, or 4 images of the corresponding anomaly type of the corresponding category in MVTec. PatchCore uses the full test sets of anomalies for all categories to determine a general choice of extraction layers and patch size. However, access to a large test set of anomalies is unrealistic. Therefore, the AutoPatch paper proposes to use small category-specific validation sets instead and shows that the optimal extraction layers and patch sizes are data-dependent and that one configuration, as proposed by PatchCore, does not rule all categories.\n\nWe hope these revisions address your concerns and improve the quality of our paper.\n\nThank you for your valuable feedback.", "role": "user" }, { "content": "Title: Thank you for your considerate response!\nI appreciate your careful consideration of my comments.\n\nI believe that with the changes that you have proposed that this work now merits acceptance to the conference, but the fact that the memory bank is a significant limitation of the method, that additional results with the subsampling, I believe that simply the comparison and improvement over the current state of the art with PathCore merits a minor accept. Another thing that influences this decision is that while the goal of patch core isn't necessarily domain shift, the intended industrial use case would suggest that it would be necessarily robust (similar to the citation of PatchCore). I think this would be a valuable future contribution.\n\nValidation Set Response\nThis didn't play a significant role in my final decision, but I am slightly confused as to your clarification about validation sets. In comparison to PatchCore, wouldn't it be a more fair comparison to evaluate both methods against the same sets of examples? Perhaps I am misunderstanding something about your evaluation.", "role": "assistant" }, { "content": "Title: Response to Reviewer 63A5\nDear Reviewer 63A5,\n\nThank you for your updated assessment of our paper. We are glad to see that our revisions have addressed some of your concerns.\n\nRegarding the memory bank, we acknowledge that this is a crucial aspect of our method, and we will consider conducting more experiments with subsampling in the future to further analyze its impact on performance.\n\nAs for the validation set question, our aim in using category-specific validation sets is to demonstrate that one configuration, as proposed by PatchCore, does not rule all categories. AutoPatch is designed to adapt to different categories and find optimal extraction layers and patch sizes specific to the validation data it encounters. PatchCore's method assumes a general predetermined choice for extraction layers and a single patch size. Both methods are trained on the same train set and evaluated on the same test set. However, it's non-trivial how to use the validation images for PatchCore.", "role": "user" }, { "content": "I am delighted to have engaged in the above conversation with the author, which has clearly contributed to strengthening this work.", "role": "assistant" } ]
QXKWSM0rFCK1	AutoML 2023 Conference	Reviewer_tX7y	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_tX7y, and the following are the paper you need to review:\n```<<QXKWSM0rFCK1>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: \nIn this paper, the challenging problem of using a high-dimensional domain in Bayesian optimization is tackled through gaussian process based variable selection. The method presents an alternative to the current state-of-the-art embedding-based approach. The paper shows that the model improves convergence iterations and time when compared with existing approaches. However, as the number of dimensions in input increases, the variable selection approach becomes less accurate.\n\noverall_review: Strength:\n\n1.\tThe main strength of the approach is that it tries to overcome the sensitivity of embedding dimensions.\n2.\tThe paper presents a novel variable selection approach that can be expanded into other areas.\n3.\tThe paper is well written, with a clear background distinguishing the contribution from the potential downsides.\n \nWeakness:\n\n1.\tThe paper fails to clarify to the audience the marginal benefits compared with other approaches on high dimensions.\n2.\tPaper relies on the assumption of separating the important from the non-important, but the robustness and stability \n\nreview_summary: \nThe paper is rated borderline leaning accept due to the following reasons:\n\n1.\tIssues in running the code to create reproducible results.\n2.\tThe robustness and stability of the approach cannot be quantified.\n3.\tUnable to distinguish the benefits of the approach since it was noted that it doesn’t perform as well as the embedding-based approach at high dimensions, which is the goal of the work.\n\nactions_required_to_increase_overall_recommendation: Following can be improved/addressed,\n\n1. A step-by-step instruction in the readme file would be helpful to run the project for reproducibility. Also, rerunning the code following the instructions will be helpful so that the reviewer will also be able to validate.\n\n2. Specify the primary benefits of the work over using embedding based BO approach given that the authors have suggested using embedding based approach as first preference.\n\n3. Robustness and stability of the variable selection at high dimensions.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to the reviewer's comments\nThank you for reviewing our work. Here are our responses to your comments:\n\n`A step-by-step instruction in the readme file would be helpful to run the project for reproducibility. Also, rerunning the code following the instructions will be helpful so that the reviewer will also be able to validate.`\n\nWe have significantly revised our codes in the repository. Please refer to the updated README.md for instructions on how to run it.\n\n`Robustness and stability of the variable selection at high dimensions.`\n\nWe have added new results to the synthetic experiments by changing the importance weights of the test function from [1, 0.1, 0.01] to [1, 0.5, 0.1] and [1,0.9,0.8] and running our method on the new functions. The results in Figure 7&8 demonstrate that our variable selection step can still detect the actual important variables, which further indicate the robustness of the variable selection module. Please refer to Figure 7&8 for a detailed description of the results.\n\nThe real-world experiments have higher dimensions than synthetic experiments, and we actually use a sampling experiment to demonstrate that our variable selection module identifies the actual important variables. Please refer to Section 5.2 and Figure 2(c, d) for a detailed description of the experiment and the results.\n\nAll of these results demonstrate the robustness and stability of the variable selection module, particularly at high dimensions.\n\n`Specify the primary benefits of the work over using embedding based BO approach given that the authors have suggested using embedding based approach as first preference.`\n\nWe did not suggest using the embedding-based approach as the first preference. In fact, all of our experimental results demonstrate that our method outperforms many embedding-based methods, such as REMBO, ALEBO, and HeSBO, particularly in terms of computational efficiency.\n\nIn the conclusion section, we do mention that in some extreme cases, such as when the function has thousands of dimensions, the embedding-based approach may still be the preferred choice because the variable selection step of our method may not be as accurate. However, these cases are extremely rare and difficult to encounter.\n Global/general Response:\n```\nWe would like to express our sincere gratitude to all the reviewers for their time and effort in reviewing our manuscript and providing valuable feedback. We have carefully considered the feedback provided by the reviewers and revised our manuscript accordingly, particularly in the experimental section. We have also responded to each reviewer individually for addressing the comments and questions.\n```\n", "role": "user" }, { "content": "Thank you for the authors' feedback. The authors have answered almost all of my concerns. I continue to recommend this work positively.", "role": "assistant" } ]
QXKWSM0rFCK1	AutoML 2023 Conference	Reviewer_3DxR	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_3DxR, and the following are the paper you need to review:\n```<<QXKWSM0rFCK1>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: 1. This paper propose a high-dimensional BO algorithms based on variable selection. \n2. Experimental results show the superior performance of VS-BO.\noverall_review: Positive\n- The proposed method is easy to understand and the paper is easy to follow.\n\nNegative\n- The experimental results seem weak. The statistical significance test may be further strengthen the results.\n- The types of real-world problems are limited.\n\nreview_summary: This paper is well-written, but the experimental part is weak.\nactions_required_to_increase_overall_recommendation: More discussions about the application scenarios of the algorithm and more types of real-world experiments under these problems.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to the reviewer's comments\nThank you for reviewing our work, we have revised our manuscript based on your comments. Here are some further responses: \n\n`More discussions about the application scenarios of the algorithm`\n\nWe have included additional related work on the applications of Bayesian Optimization (BO) in the introduction section. Furthermore, in the Broader Impact Statement, we emphasized that BO can be applied to various scientific fields, including but not limited to machine learning, computational chemistry, and bioinformatics.\n\n`The statistical significance test may further strengthen the results.`\n\nWe have updated our manuscript by using mean and standard error instead of mean and standard deviation. This metric can provide a better comparison between different methods.\n\nAdditionally, we have included more results in the experimental section of our revised manuscript.\n\n\n Global/general Response:\n```\nWe would like to express our sincere gratitude to all the reviewers for their time and effort in reviewing our manuscript and providing valuable feedback. We have carefully considered the feedback provided by the reviewers and revised our manuscript accordingly, particularly in the experimental section. We have also responded to each reviewer individually for addressing the comments and questions.\n```\n", "role": "user" }, { "content": "I acknowledge the authors' response. Most of my concerns have been addressed. I continue to recommend this work positively.", "role": "assistant" } ]
QXKWSM0rFCK1	AutoML 2023 Conference	Reviewer_MVQB	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_MVQB, and the following are the paper you need to review:\n```<<QXKWSM0rFCK1>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper addresses an important problem in the field of Bayesian optimization (BO): high-dimensional BO. They proposed a new computational efficient BO algorithm that leverages variable selection by performing the following steps: (1) select variables that are considered “important”; (2) recommend experiments using GP models fitted using these variables; (3) get observations and adjust variables deemed “important” every few iterations. The method is compared against other SOTA algorithms for high-dimensional BO problems based on both # iterations and wall/CPU times. \noverall_review: Positive aspects \n- This paper is well organized, with clearly stated framework/algorithm and motivation.\n- The proposed algorithm is compared against multiple SOTA algorithms. It seems to achieve comparable or better performance on synthetic and real problems and is computationally efficient.\n\nNegative aspects:\n- Some parts of the paper can benefit some clarification and critical analysis.\n- There might be some room for improvement for the experiments and the definition of the IS score.\nreview_summary: This is a well-written and thoughtful paper that can potentially have good impact in the BO and AutoML community. There are some issues in the papers but can be fixed through some additional experiments / clarification texts.\nactions_required_to_increase_overall_recommendation: I would like to see the author address the questions/concerned raised in the \"Technical Quality And Correctness\" and \"Clarity\" sections.\nRating: 8 accept, good paper\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to the reviewer's comments \nThank you for reviewing our work. Here are our responses to your comments:\n\n`When computing the importance score (IS) on page 4, have you explored other strategies than sampling uniformly from the design space? It is possible that the sensitivity of the function w.r.t. a feature is high when the function value is also low.`\n\nIn our manuscript, we demonstrate in equations (4)-(5) that computing the Importance Sampling (IS) is not based on the average of function values, but rather on the average of the gradients of the function (scaled by the standard deviation). We acknowledge that it is possible for the sensitivity of a feature to be high when the function value is low, which is why we opt to use the function gradient. Besides, using Monte Carlo sampling to do estimation in (5) seems to be a very natural choice.\n\n`I find it difficult to have an good interpretation for the distribution p(x \| D) which is used to sample unimportant features. When you say D, did you mean just the X’s or the Y’s as well? What happens some features are just purely noise? Could you clarify on that in the paper?`\n\n`Despite having a relatively good understanding of the CMA-ES algorithm, I find it difficult to fully understand what was done. Some pseudo-code/clarification might help.`\n\nIn our manuscript, we discuss in section 3 and Algorithm 1 that D represents a set of x,y pairs: $D=((x^i,y^i))_{i=1}^{N}$. $p(x\|D)$ is actually a learned distribution so that when we sample $x$ from $p(x\|D)$, the corresponding $y$ value should be more likely to be higher. In cases where certain features are simply noise, $p(x\|D)$ should ideally resemble random sampling for those features.\n\nCMA-ES is an algorithm to learn the distribution $p(x\|D)$. One of the advantages of CMA-ES is that the learned distribution $p(x\|D)$ is a multivariate Gaussian distribution, which implies that its conditional distribution is also multivariate Gaussian and can be sampled easily.\n\n`I wonder what the performance of the algorithm woud look like when (1) in the test problem we have only pure noise in addition to the important features (2) the test problem (e.g., the hartmann problem) has a similar specification as its current form but has weights like [1, 0.9, 0.8] instead of [1, 0.1, 0.01]. I am interested because (1) helps understand the performance of sampling unimportant features better and (2) helps understand the algorithm performance when you would need to sequantially fit more GPs in variable selection. To me these experiments would strengthen the arguments of paper even if it doen’t perform well in some of the situations. To the very least, one should hypothesize what would happen if such problems were tested.`\n\nWe have added some additional experimental results with weights [1,0.5,0.1] and [1,0.9,0.8], and Figure 7&8 in the manuscript shows the total frequency of being chosen as important for each variable in these cases. As shown, our method is able to detect correct important variables in this case. Please refer to the updated manuscript for further details.\n\n`A plot of \|x_ipt\| (i.e. number of variables deemd “important”) versus iteration/time would help.`\n\nWe have added these experimental results. Please see Figure 6 in the updated manuscript. As shown, only a small number of variables are deemed important in each case, and this number is close to the number of important variables in reality. \n\n`It seems like SAASBO performs the best when progress is plotted against iterations. Taking a quick look at the repo, the author seems to be using the default setup for the SAASBO algorithm. I am interested in seeing its performance when it’s less computationally expensive (e.g., smaller # samplers, # warm-ups, etc.)`\n\nBy default, num_warmup=256 and num_samples=256 in SAASBO. We evaluated a modified SAASBO that used num_warmup=128 and num_samples=128. This version of SAASBO ran faster, but was still significantly slower than other methods. SAASBO (num_warmup=256 and num_samples=256) requires approximately 9 hours to complete 200 iterations on the Branin case, SAASBO (num_warmup=128 and num_samples=128) takes around 5.5 hours on the same task, whereas VS-BO can accomplish 500 iterations on this case in around 20 minutes.\n\nWe did not incorporate these findings into our updated manuscript because we were unable to complete all of the test cases using the less computation-intensive version of SAASBO. Further decreasing the value of num_warmup or num_samples may speed up SAASBO, but we are not sure if it would affect its performance. We may include these results later if our work is accepted.\n\n\n Global/general Response:\n```\nWe would like to express our sincere gratitude to all the reviewers for their time and effort in reviewing our manuscript and providing valuable feedback. We have carefully considered the feedback provided by the reviewers and revised our manuscript accordingly, particularly in the experimental section. We have also responded to each reviewer individually for addressing the comments and questions.\n```\n", "role": "user" }, { "content": "Title: Response to the authors\nThanks for the additional experiments and clarification. Most of my concerns have been addressed and I'll raise my score accordingly. ", "role": "assistant" } ]
QXKWSM0rFCK1	AutoML 2023 Conference	Reviewer_YTEk	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_YTEk, and the following are the paper you need to review:\n```<<QXKWSM0rFCK1>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper addresses high-dimensional Bayesian optimization problems and proposes a new algorithm that leverages variable selection. The idea is to embed high-dimensional space in low-dimensional space with a novel technique. The paper also analyzes the computational complexity and presents the empirical evaluations with results.\noverall_review: The proposed algorithm and the idea is simple and clear. The contribution is declared in a good way. However, there are some notations as I mentioned above should be corrected for publishing the paper. Also, in line 128, it is declared that \"The algorithm uses different strategies to choose new queries of the variables from two different sets.\" It is not clear for the rest of the paper.\nreview_summary: Except for minor corrections, the paper is well-written and contains novelties. There are some parts that are not clear enough as I declared above, and my score will be weak accept for consideration of the other reviewer's opinion.\nactions_required_to_increase_overall_recommendation: The definitions could be in another section instead of the Background. Especially, since the equations in the background are unexpected, preliminaries would be better. Mathematical notations need a double check as I mentioned some of them.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to the reviewer's comments \nThank you for reviewing our work, we have revised our manuscript based on your comments. Here are some further responses:\n\n`i.e. the numbers representing the sequential should be down instead of up in line 70. It may cause confusion about the root of a number.`\n\nWe use the superscript instead of subscript here because $x_i$ means the i-th entry of the vector $x$. Therefore, we use $x^{i}$ in line 70. \n\n`Also, in line 128, it is declared that \"The algorithm uses different strategies to choose new queries of the variables from two different sets.\" It is not clear for the rest of the paper.`\n\nAs we explained in section 3 of our manuscript, we use the Bayesian Optimization framework to choose new queries for important variables, while for unimportant variables, we use CMA-ES to sample their values. We have revised our manuscript to provide further clarity on this point.\n\n Global/general Response:\n```\nWe would like to express our sincere gratitude to all the reviewers for their time and effort in reviewing our manuscript and providing valuable feedback. We have carefully considered the feedback provided by the reviewers and revised our manuscript accordingly, particularly in the experimental section. We have also responded to each reviewer individually for addressing the comments and questions.\n```\n", "role": "user" }, { "content": "Thanks to the authors for their clarification. The response covers nearly all of the issues I raised. I will continue to recommend accepting this work.", "role": "assistant" } ]
QXKWSM0rFCK1	AutoML 2023 Conference	Reviewer_ctFh	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_ctFh, and the following are the paper you need to review:\n```<<QXKWSM0rFCK1>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions:  \n\nThe authors introduce a new method for high-dimensional Bayesian optimisation based on variable selection.\n\n \noverall_review:  \n\n__MAJOR POINTS__\n\n \n\n1. In the section on related work, the literature on VAE-based embedding methods is relevant but is not described. While I believe it is beyond the scope of the paper to compare against VAE-based embedding methods, references such as [2-4] could be briefly discussed. In terms of the literature on high-D BO, TuRBO [5] could also be discussed.\n\n2. In Algorithm 2, line 7, the stopping criterion is defined as the iteration when the difference between successive values of the negative log marginal likelihood is more less than a factor of 10 of the difference between successive values at the previous iteration. This is a hyperparameter of the algorithm and I am wondering if the authors performed any sensitivity analysis on this stopping criterion?\n\n3. For the experiments, why do the authors report the results with 1/8 standard deviation in place of the standard error? The latter may be preferred if the objective is to establish the relative ranking of algorithms. The former may be preferred if the goal is to showcase the variability in performance of the different algorithms.\n\n4. On line 283, in terms of the claim that VSBO outperforms VSBO-mix, would the errorbars be overlapping in the case that the standard error was plotted instead of 1/8 standard deviations? In other words can this claim be definitely supported? If not, perhaps it should be moderated?\n\n \n\n__MINOR POINTS__\n\n \n\n1. On line 2, the authors state that the input domain $\\mathcal{X}$ is in $[0, 1]^D$. The algorithm introduced, however, is more broadly applicable for bounded input domains in $\\mathbb{R}^D$. As such, it may be worth stating that for the purposes of the paper the authors restrict their study \"without loss of generality\" to input domains in $[0, 1]^D$.\n\n2. Line 23, \"maximizes the function f\" assumes a maximization problem. While the intention is understood perhaps \"optimization\" would be more appropriate before the black-box optimization problem is formerly defined as a maximization problem. This is a very minor nit and one which is a matter of taste!\n\n3. On line 28, the reference Griffiths and Hernández-Lobato was published in Chemical Science [6].\n\n4. Line 49, \"Letham et al. also points out that when projecting the optimal point from the embedding space back to the original space, there is no guarantee that the resulting point is within X , and thus the algorithm may fail to find an optimum within the input domain.\" It is worth mentioning here that this problem can be partially addressed (at least in the case of linear embedding methods such as ALEBO) by mapping the input bounds from the original space to the embedding space. In the case of PCA for example, the embedding-based bounds would take the form of linear inequality constraints.\n\n5. In Equation 1, the mean vector is set to zero. This implicitly assumes that the observations have been standardized and this would be worth mentioning.\n\n6. In Equation 5, the motivation for randomly sampling points $x^{(k)}$ as opposed to grid sampling them, in addition to the expense of grid sampling for high dimensions, is presumably so that computation of the IS is robust to uninformative dimensions as in [1] for example?\n\n7. Moriconi et al. Was published in the machine learning journal.\n\n8. BoTorch was published in NeurIPS 2020.\n\n9. Line 203, typo, \"Matérn\". This is a major crime against the French language.\n\n10. Line 326, in terms of the applications for computational chemistry it would be particularly exciting to see applications of VS-BO to descriptor-based molecular property prediction and optimization [7].\n\n11. In the references there are some missing capitalizations such as \"Bayesian\", \"Gaussian\", \"Monte Carlo\".\n\n12. Line 577, an appropriate citation for L-BFGS could be included [8].\n\n13. In Figure 6b) are the confidence bands really 1/8 standard deviations? The uncertainty bands would be very large if that is the case!\n\n \n\n__REFERENCES__\n\n \n\n[1] Bergstra and Bengio, [Random Search for Hyper-parameter Optimization](https://www.jmlr.org/papers/volume13/bergstra12a/bergstra12a.pdf), JMLR, 2012.\n\n[2] Grosnit et al., [High-Dimensional Bayesian Optimization with Variational Autoencoders and Deep Metric Learning](https://arxiv.org/abs/2106.03609), arXiv, 2021.\n\n[3] Maus et al., [Local latent Space Bayesian Optimization over Structured Inputs](https://proceedings.neurips.cc/paper_files/paper/2022/file/ded98d28f82342a39f371c013dfb3058-Paper-Conference.pdf), NeurIPS 2022.\n\n[4] Notin et al., [Improving black-box optimization in VAE latent space using decoder uncertainty](https://openreview.net/forum?id=F7LYy9FnK2x), NeurIPS 2021.\n\n[5] Eriksson et al., [Scalable Global Optimization via Local Bayesian Optimization](https://proceedings.neurips.cc/paper_files/paper/2019/file/6c990b7aca7bc7058f5e98ea909e924b-Paper.pdf)\n, NeurIPS 2019.\n\n[6] Griffiths and Hernández-Lobato, [Constrained Bayesian Optimization for Automatic Chemical Design using Variational Autoencoders](https://pubs.rsc.org/en/content/articlehtml/2019/sc/c9sc04026a), Chemical Science, 2020.\n\n[7] Griffiths et al., [GAUCHE: A Library for Gaussian Processes in Chemistry](https://arxiv.org/abs/2212.04450), arXiv, 2023.\n\n[8] Liu and Nocedal, [On the limited memory BFGS method for large scale optimization](https://link.springer.com/article/10.1007/BF01589116). 330 Math. Program., 45(1-3):503–528, 1989.\n\n \nreview_summary:  \n\nA new algorithm is introduced for high-dimensional Bayesian optimization. The algorithm is simple and straightforward to implement, the empirical and theoretical analyses are thorough and the paper is very well-written. As such I recommend acceptance.\n\n \nactions_required_to_increase_overall_recommendation:  \n\nI am willing to increase my score by a further point if the minor issues raised above are addressed.\n\n \nRating: 9 strong accept, excellent work\nConfidence: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\n", "role": "assistant" }, { "content": "Title: Response to the reviewer's comments\nThank you for reviewing our work. Here are our responses to your comments: \n\n`In Equation 2, there is a minus sign in front of the expression for the posterior predictive variance. Is this a typo?`\n\nThis is not a typo, it looks weird because we place the term $k(x,x,\\Theta)$ as the second term, we have revised its placement to make the notation more clearer. \n\n`Line 68, \"a GP\". Grammatically, it would be great to include definite/indefinite articles and pluralisation where appropriate e.g. \"a GP\", \"the GP\", \"GPs\".`\n\nWe have revised our manuscript based on this comment. \n\n`Line 70, It may be worth defining $y^{1:n}$ as ... to simplify Equation 2 later on. From the reviewer's (potentially opinionated!) perspective this tends to be the more common means of defining observation vectors in the literature.`\n\nWe have revised our manuscript based on this comment. \n\n`Line 73, perhaps it would be preferable to refer to \\Theta and \\sigma_0 as hyperparameters of the GP rather than parameters.`\n\nWe have revised our manuscript based on this comment. \n\n`Line 103, \"which is similar to an axis-aligned subspace embedding\". Is this not exactly an axis-aligned subspace embedding?`\n\nYes, our method is an axis-aligned subspace embedding, we have revised our manuscript based on this comment. \n\n`Line 111, \"hyperparameter in the kernel function\".`\n\nWe have revised our manuscript based on this comment. \n\n`Algorithm 1, line 1, f(x) is given as an input to the algorithm. This may imply that the black-box function is given directly (analytic form) rather than simply the ability to query it.`\n\nWe have revised our manuscript based on this comment. \n\n`Algorithm 1, line 4, as written, the observations are sampled without noise`\n\nWe have revised our manuscript based on this comment. \n\n`Algorithm 1, line 16, why is x^{max} chosen to be that with maximal...as opposed to that with the maximal GP posterior mean?`\n\nThis is a common choice for Bayesian Optimization, i.e. choose an input with a known high y value, rather than relying solely on a predicted high value. \n\n`Algorithm 2, line 3, IS[I]. Given that i is used as an index for the data points, it may be worth using a second, distinct index for the dimensions.`\n\nWe have revised our manuscript based on this comment. \n\n`Algorithm 2, line 4, It may be worth explicitly defining D as the full dimension of the input space.`\n\nWe have defined $D$ at the beginning of section 3. \n\n`Algorithm 2, line 4, What is the meaning of the subscript s? Is it necessary?`\n\nYes it is necessary. We use $x_{s(i)}$ to represent a permutation of variables in the input $x$ that are sorted by their importance scores, if we do not use $x_{s(i)}$, then $x_i$ should represent the i-th variable in $x$ originally. \n\n`Algorithm 1, line 5, p(x\|D) is introduced abstractly here while it is later explained that it is chosen to be a multivariate Gaussian distribution. It may be worth explaining that when introducing the algorithm.`\n\nWe have revised our manuscript based on this comment. \n\n`Line 184, typo, \"Because of the properties of Gaussian distributions\".`\n\nWe have revised our manuscript based on this comment. \n\n`Line 503, hyper-parameters are not defined as the kernels signal amplitude and the inverse lengthscales respectively.`\n\nWe have revised our manuscript based on this comment. \n\n`Line 527, should K() be lower-case k here as a vector quantity? This would be consistent with the main paper.`\n\nYes, it should be lower-case. We have revised our manuscript based on this comment. \n\n`Line 535, in the expression for the UCB acquisition, what is the reason for taking the square root of the dimensionless exploration parameter Beta?`\n\nThere is no specific reason. I think the only reason is to tell people Beta should be non-negative. We use the square root because previous work usually has this UCB formulation, such as https://arxiv.org/pdf/0912.3995.pdf.\n\n`In the section on related work, the literature on VAE-based embedding methods is relevant but is not described. While I believe it is beyond the scope of the paper to compare against VAE-based embedding methods, references such as [2-4] could be briefly discussed. In terms of the literature on high-D BO, TuRBO [5] could also be discussed.`\n\nWe have revised our manuscript based on this comment. \n\n`In Algorithm 2, line 7, the stopping criterion is defined as the iteration when the difference between successive values of the negative log marginal likelihood is more less than a factor of 10 of the difference between successive values at the previous iteration. This is a hyperparameter of the algorithm and I am wondering if the authors performed any sensitivity analysis on this stopping criterion?`\n\nWe have added ablation studies on hyper-parameters, stopping criterion $r_{stop}$ and the number of Monte Carlo samples $N_{is}$ for estimating the importance score. The results are in section D and Figure 9. We find that these two hyper-parameters are highly robust across different values. \n\n`For the experiments, why do the authors report the results with 1/8 standard deviation in place of the standard error? The latter may be preferred if the objective is to establish the relative ranking of algorithms. The former may be preferred if the goal is to showcase the variability in performance of the different algorithms.`\n\n`On line 283, in terms of the claim that VSBO outperforms VSBO-mix, would the errorbars be overlapping in the case that the standard error was plotted instead of 1/8 standard deviations? In other words, can this claim be definitely supported? If not, perhaps it should be moderated?`\n\nWe have revised our figures and replaced all the standard deviations with standard errors. When comparing VSBO with VSBO-mix, we find that the errorbars do not overlap in the Hartmann6 case and Styblinski-Tang4 case, so we think that the claim that VSBO outperforms VSBO-mix still holds. \n\n`On line 2, the authors state that the input domain...As such, it may be worth stating that for the purposes of the paper the authors restrict their study \"without loss of generality\" to input domains in [0,1]^{D}`\n\nWe have revised our manuscript based on this comment.\n\n`On line 28, the reference Griffiths and Hernández-Lobato was published in Chemical Science [6].`\n\nWe have revised our manuscript based on this comment.\n\n`Line 49, \"Letham et al. also points out that when projecting the optimal point from the embedding space back to the original space, there is no guarantee that the resulting point is within X , and thus the algorithm may fail to find an optimum within the input domain.\" It is worth mentioning here that this problem can be partially addressed (at least in the case of linear embedding methods such as ALEBO) by mapping the input bounds from the original space to the embedding space. In the case of PCA for example, the embedding-based bounds would take the form of linear inequality constraints.`\n\nWe have revised our manuscript based on this comment.\n\n`In Equation 1, the mean vector is set to zero. This implicitly assumes that the observations have been standardized and this would be worth mentioning.`\n\nWe have revised our manuscript based on this comment.\n\n`In Equation 5, the motivation for randomly sampling points... as opposed to grid sampling them, in addition to the expense of grid sampling for high dimensions, is presumably so that computation of the IS is robust to uninformative dimensions as in [1] for example?`\n\nWe agree with this opinion, although we didn't point out it explicitly in our manuscript since it is not a major point in our work. \n\n`Moriconi et al. Was published in the machine learning journal.`\n\nWe have revised our manuscript based on this comment.\n\n`BoTorch was published in NeurIPS 2020.`\n\nWe have revised our manuscript based on this comment.\n\n`Line 203, typo, \"Matérn\". This is a major crime against the French language.`\n\nWe have revised our manuscript based on this comment.\n\n`Line 326, in terms of the applications for computational chemistry it would be particularly exciting to see applications of VS-BO to descriptor-based molecular property prediction and optimization [7].`\n\nWe have cited [7] in the introduction section now. \n\n`Line 577, an appropriate citation for L-BFGS could be included [8].`\n\nWe have cited [8] now. \n\n`In Figure 6b) are the confidence bands really 1/8 standard deviations? The uncertainty bands would be very large if that is the case!`\n\nWe have replaced the standard deviation with the standard error. \n Global/general Response:\n```\nWe would like to express our sincere gratitude to all the reviewers for their time and effort in reviewing our manuscript and providing valuable feedback. We have carefully considered the feedback provided by the reviewers and revised our manuscript accordingly, particularly in the experimental section. We have also responded to each reviewer individually for addressing the comments and questions.\n```\n", "role": "user" }, { "content": "Title: All of my Points have been Addressed\nAll of my points have been addressed and as promised, I increase my score. There would appear to be a typo in Section D of the appendix, \"Sensitivity Analysis\". I appreciate the additional experiments provided by the authors as well as the clarifications in the author response!", "role": "assistant" } ]
QXKWSM0rFCK1	AutoML 2023 Conference	Reviewer_cR9i	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_cR9i, and the following are the paper you need to review:\n```<<QXKWSM0rFCK1>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: - The introduction of a novel, adaptive variable selection scheme for high-dimensional Bayesian Optimization\n- A novel treatment of important versus unimportant variables through the dual use of BO and CMA-ES\n- The BO that results from the proposed variable selection is cheap, both in theory and in practice\n\noverall_review: # Positive\n- __The relevance is high__: High-dim BO does feature a number of methods which are expensive to the point where even benchmarking against them is painful. Having computationally feasible methods is important, particulairly since the function is likely going to need 100's of evaluations to be optimized, anyway.\n- __The idea is good__: Identifying relevant variables and optimizing solely on these is a simple and intuitive idea, which I appreciate. Combining with CMA-ES also seems sensible from a practical perspective. \n- __The proposed algorithm is provably cheap__: Given the conventional BO setup, substeps of the algorithm scales computationally with the dimensionality of $x_{ipt}$ \n- __The evaluation protocol is novel (?) and very practial__: Adding semi- and mini-relevant dimensions to synthetic test functions is a great idea - one which seems substantially more realistic than the current black-and-white regime.\n- __Good empirical performance__: Clearly, the method performs well, which is even more impressive given its compute time. However, I _do not_ view this as supremely important. \n- __Clear and well-explained__. The paper is mostly a pleasant read, with a high degree of correctness and clarity. \n\n\n# Negative:\n- __Hyperparameters of variable selection__: Algo. 2, Line 7 contains the hyperparameters 3 and 10. These are not mentioned as such, nor ablated. I would greatly appreciate if the authors could shed some light on how they arrived at these values, and provide an ablation on how these impact performance. Moreover, I believe they should be parametrized so that their scalar values are not part of the algorithm. \n- __Computational expense & Large number of GPs__: The method involves fitting $D$ GPs, making it scale as $\\mathcal{O}(Dn^3)$. To my knowledge, is a uniquely large factor that other methods do not share (rather, they are $\\mathcal{O}($D^2 + Dn^2 + n^3)$ as pointed out by the authors. This puts the main claimed benefit of the paper into question, as there should be tasks (pref. those with some multiple on the number of evaluations) where the proposed method is in fact one of the slower ones.\n- __Inability to pick up semi-relevant dimensions__: VSBO does not seem to pick up on the semi-important dimensions very well based on the frequency plots. In particular, it struggles with the semi-active dimensions on Styb-Tang, despite the function's extreme steepness along the borders which should simplify matters. Does it pick up any semi-active dimensions (with the current hyperparameter values) for a function with a flatter surface like, say, Hartmann-3? . This could probably be addressed by varying the aforementioned hyperparameters, but I would ultimately want to see that it works as intended out of the box.\n- __The algorithm is not digestible__: There is _a lot_ of material, including 5 algorithms totaling over two pages, to go through to get to VS-momentum (i.e. the actual algorithm that is used). To me, this appears to be a design flaw, and _hints_ at over-fitting the algorithm to the problem(s) at hand. Moreover, I want to see the performance of the non-momentum variant.\n\n\n- __Missing comparison against relevant methods__: Two competing methods from NeurIPS-22, which both work with the (adaptive) axis-aligned subspace assumption, are missing (BaXUS [1], MCTS-VS [2]). In particular, they are both very relevant with the chosen evaluation protocol. I recognize that these are relatively recent (more than 3 months since publication), but I ultimately do not they are recent enough to warrant exclusion. Since we on the topic of computational efficiency/scalability, TuRBO [3] should probably be in there as well. At the least, these should be included as related work. Having [1] (which uses some of the same benchmarks) and [2] included would improve my perception of the paper.\n\n- __Relation of IS to lengthscales__:The proposed variable selection criterion _seems_ to be very correlated to the lengthscales of the model (except if short lengthscales & little data), and my intuition tells me that the lengthscales should suffice about as well for variable selection. Can the authors provide evidence to the contrary?\n- __The MC approximation is not ablated__: The Monte Carlo approximation seems fairly sample-intensive, considering the large dimensionality of the problem. How many samples were used, and does the IS suffer if (relatively) few samples are used? \n\n# General feedback:\n- __Improving the motivation for the method__: I greatly appreciate Fig. 4 and what it communicates (although I think a. and c. are unnecessary). I was not aware that the cost of optimizing the GP in high-dim is that significant, and I think it should be highlighted more. In fact, I think it should serve as a motivating example for the entire paper and be featured in the introduction. \n- __Excessive time-related plots__: Both wall-clock and CPU time seems extreme, and they are (naturally) just about identical. Please remove wall-clock time to make room for Styb-Tang. \n- __Plot the regret__: Please, plot the regret (best possible - current best), especially when you have introduced test new functions where readers may not know the optimal value by heart. \n-__CMA-ES blurs the evaluation: The inclusion of CMA-ES blurs the evaluation, especially under the proposed evaluation protocol. In fact, it appears that CMA-ES (and not the variable selection) is responsible for the optimization of the semi-active dimensions in 2/3 synthetic cases, since VSBO doesn't indenify these dimensions.\n\n\n# Questions\n_No, I don't expect experiments, but would greatly appreciate some form of clarity_!\n- How susceptible is the proposed variable selection procedure to observation noise of various magnitudes?\n\n\n# Minor stuff\n\n- I'm missing the boundary issue (Swersky 2016) as a reason for high-dimensional BO being challenging. Would the authors agree?\n- Plotting 1/8 standard __deviation__ seems odd. Why not plot the standard __error of the mean__?\n- Fig. 4: The y-axis label \"Wall clock time __/__ second\" reads \"Wall clock time __per__ second\" and is very is unconventional, please go with \"Wall clock time (seconds)\"\n- __the__ GP is missing almost everywhere, such as in line 73, 74 & 75.\n- Algo 5, row 28: You've written () instead of {} in LaTeX.\n\n\n1. Increasing the Scope as You Learn: Adaptive Bayesian Optimization in Nested Subspaces. Leonard Papenmeier, Luigi Nardi, Matthias Poloczek. In Advances in Neural Information Processing Systems (NeurIPS), 2022.\n\n2. Monte Carlo Tree Search based Variable Selection for High Dimensional Bayesian Optimization. Lei Song, Ke Xue, Xiaobin Huang, Chao Qian. In Advances in Neural Information Processing Systems (NeurIPS), 2022.\n\n3. D. Eriksson, M. Pearce, J. Gardner, R. D. Turner, and M. Poloczek. Scalable Global Optimization via Local Bayesian Optimization. In Advances in Neural Information Processing Systems (NeurIPS), 2019.\nreview_summary: The idea of variable selection is good, the evaluation protocol is novel and practical, and the paper could potentially fill a need for computationally efficient high-dimensional BO with consistent performance. However, the algorithm does not appear intrinsically scalable $(\\mathcal{O}(Dn^3))$ and it is very difficult to gauge how well the proposed variable selection scheme works, given:\n- The use of a modified method, VS-momentum, in the results\n- The hyperparameters of the method\n- The functions that are benchmarked against\n\nbut the total list is even longer. Thus, I am not convinced that the proposed variable selection scheme has merit.\n\nI recommend to reject the paper in its current format. I am confident in my assessment given the current evidence, but I am very open to changing my opinion if my critique is addressed.\nactions_required_to_increase_overall_recommendation: I am very open to changing my score, and it greatly hinges on further assessing the quality of the variable selection. In an ideal world, that would entail:\n- Showing the results (and frequency plots) of the non-momentum method, preferably with & without CMA-ES (doing RS instead)\n- Showing how sensitive the (non-momentum) method is to the hyperparameters\n- Ablate at which level (i.e. $0.1f, 0.2f, 0.3f$) variables start to become important for Hartmann-6 and Styb-Tang\n \nHere, I view the first two as must-haves, and the rest being nice-to-haves. I don't expect it to perfectly pick up all relevant dimensions, but there is currently far too many confounding factors.\n\nFurther, I would like to see that the computational expense does not blow up (relative to the competition, given the unique (?) scaling factor) when the number of evaluations is increased by a factor $2-5$ - a common high-dimensional BO budget. I recognize that this is unrealistic for ALEBO and SAASBO.\n\nMoreover, the inclusion of the aforementioned NeurIPS works would be nice to have.\nRating: 4 possibly reject, but has redeeming facets\nConfidence: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\n", "role": "assistant" }, { "content": "Title: Response to the reviewer's comments\nThank you for reviewing our work, we have revised our manuscript based on your comments. Here are some further responses:\n\n`Hyperparameters of variable selection: Algo. 2, Line 7 contains the hyperparameters 3 and 10. These are not mentioned as such, nor ablated. I would greatly appreciate if the authors could shed some light on how they arrived at these values, and provide an ablation on how these impact performance. Moreover, I believe they should be parametrized so that their scalar values are not part of the algorithm.`\n\n3 is actually not a hyper-parameter. Since the stopping criterion is defined as the iteration when the difference between successive values of the negative log marginal likelihood is less than a factor of 10 of the difference between successive values at the previous iteration, the number here is at least 3 (we need $L_{m-2}$, $L_{m-1}$, and $L_{m}$). \n\nWe have added some new ablation studies about hyper-parameter 10. Please see Figure 9 of the updated manuscript. We find that this hyper-parameter is highly robust across different values. \n\n`Computational expense & Large number of GPs: The method involves fitting...This puts the main claimed benefit of the paper into question, as there should be tasks (pref. those with some multiple on the number of evaluations) where the proposed method is in fact one of the slower ones.`\n\nWe think that the reviewers may have misunderstood the computational complexity of our method. Our approach only requires fitting one Gaussian Process (GP) per iteration, which is the same as the standard Bayesian Optimization (BO) framework, instead of D GPs.\n\nAlthough our method does fit multiple GPs during the variable selection step, two factors mitigate the computational cost. Firstly, assuming that the dimension of $x_{ipt}$, p, is significantly smaller than D, stepwise forward-based variable selection only requires fitting p GPs instead of D GPs. Secondly, we perform variable selection every 20 iterations rather than at each iteration in our experiments.\n\nWe acknowledge that in extreme cases where every variable is crucial, the variable selection step may require a longer computation time.\n\n`The MC approximation is not ablated: The Monte Carlo approximation seems fairly sample-intensive, considering the large dimensionality of the problem. How many samples were used, and does the IS suffer if (relatively) few samples are used?`\n\nWe use 10000 samples. We have added some new ablation studies about this number. Please see Figure 9 of the updated manuscript.\n\n`Showing the results (and frequency plots) of the non-momentum method, preferably with & without CMA-ES (doing RS instead)`\n\nWe have added new results on the comparison among VS-BO, VS-BO-nonmomentum, and VS-BO-nonmomentum without CMA-ES. Please see Figure 11 of the updated manuscript. We find that although the regret curves of VS-BO-nomomentum are highly similar to those of VS-BO, VS-BO-nomomemtum appears to be less effective in selecting secondary important variables, particularly in the Branin case. Therefore, we decide to keep VS-BO (with momentum) in our manuscript. \n\n`Excessive time-related plots: Both wall-clock and CPU time seems extreme, and they are (naturally) just about identical. Please remove wall-clock time to make room for Styb-Tang.`\n\nWe have put the Styb-Tang into the main manuscript. \n\n`Plot the regret: Please, plot the regret`\n\nWe have plotted the regret for synthetic functions. We don't plot the regret for real-world problems because it is assumed that we don't know the optimal value. \n\n`Plotting 1/8 standard deviation seems odd. Why not plot the standard error of the mean?`\n\nWe have used standard error to replace all the standard deviation. \n\n`Fig. 4: The y-axis label \"Wall clock time / second\" reads \"Wall clock time per second\" and is very is unconventional, please go with \"Wall clock time (seconds)\"`\n\nWe have revised our manuscript based on this comment. \n\n`the GP is missing almost everywhere, such as in line 73, 74 & 75.`\n\nWe have revised our manuscript based on this comment. \n\n`Showing how sensitive the (non-momentum) method is to the hyperparameters`\n\nAs we describe above, we have added ablation studies on the variable selection stop criteria and the number of Monte Carlo samplings. Please see Figure 9 of the updated manuscript. We find that both hyper-parameters are highly robust across different values. \n\n`Further, I would like to see that the computational expense does not blow up (relative to the competition, given the unique (?) scaling factor) when the number of evaluations is increased by a factor 2-5`\n\nWe have found that our method requires approximately 20 minutes to run 500 iterations for the Branin function, 15 minutes for the Hartmann6 function, 35 minutes for the StyblinskiTang4 function, 60 minutes for the Rover function, and 100 minutes for the MOPTA function. Therefore, we consider our method to be practical and feasible for these tasks.\n\nWe agree that ALEBO or SAASBO cannot handle such a high number of iterations. For example, SAASBO requires approximately 9 hours to complete 200 iterations on the Branin function. We have pointed out this in our updated manuscript. \n\n`Relation of IS to lengthscales:The proposed variable selection criterion seems to be very correlated to the lengthscales of the model (except if short lengthscales & little data), and my intuition tells me that the lengthscales should suffice about as well for variable selection. Can the authors provide evidence to the contrary?`\n\nFirstly, the lengthscale and variable selection serve different purposes. The lengthscale measures the importance of variables, while variable selection is a method to choose a subset of variables. Hence, the lengthscale cannot replace variable selection.\n\nSecondly, we chose not to use the lengthscale for computing the importance score in our work. This is because previous research, specifically \"Variable selection for Gaussian processes via sensitivity analysis of the posterior predictive distribution,\" has demonstrated the disadvantages of the lengthscale and shown that the method we use for computing the importance score is better. \n\n`Ablate at which level variables start to become important for Hartmann-6 and Styb-Tang`\n\nWe have added some new experimental results. We test our method on three new synthetic functions with importance weight [1,0.5,0.1] and [1,0.9,0.8], and the results in Figures 7&8 show that our method can detect correct important variables. Please refer to the updated manuscript for further details. \n\n", "role": "user" }, { "content": "Title: Further response to rebuttal\nThanks to the reviewers for addressing the concerns regarding the ablations. It is nice to see that the method is indeed robust to the likelihood-related hyperparameter. I appreciate the clarification on $m$, and I consequently agree with the authors. Further, it's nice to see that the \"vanilla\" (non-momentum, non-CMA-ES) version of the method appears effective. As such, I have increased my score one step. However, I am still not satisfied with regard to the discussion of scalability, which is the central claim of the paper. As a reported strength of the method, it is concerning that experiments are run with so few evaluations (200, TurBO does thousands for comparison) and that it possesses a uniquely large scaling factor. Anecdotal evidence is not sufficient in this regard.\n\n##### Although our method does fit multiple GPs during the variable selection step, two factors mitigate the computational cost. Firstly, assuming that the dimension of $x_{ipt}$, p, is significantly smaller than D, stepwise forward-based variable selection only requires fitting p GPs instead of D GPs. \n\n__In Algorithm 2, Line 5 & 6:__\n\n__for__ $ m = 1:D$\n\nFit GP to $\\mathcal{D_n}$\n\nThe subsequent if-condition does not guarantee a smaller computational cost. \n\nThe scaling is clearly $\\mathcal{O}(Dn^3)$ as long as the if-statement does not guarantee an early stop - I do not believe I have misunderstood this. To this end, it is unimportant how frequent the variable selection step takes place as long as it's done throughout. I agree that there are mitigating factors which affect the practical computationsl expense, but again - this does not affect the _theoretical_ scalability of the method. \n\n##### We acknowledge that in extreme cases where every variable is crucial, the variable selection step may require a longer computation time.\nYes, but $\\mathcal{O}(pn^3)$, for some $p < D$ would also be a uniquely large factor that could cause scalability issues.\n\n\nThanks to the authors for addressing some of my concerns - I am still open to further raising my score if the questions regarding scalability is properly addressed in the experiments.", "role": "assistant" }, { "content": "Title: Response to the reviewer's comments\nThank you for your further response. You are correct that from a theoretical perspective, the scaling for the variable selection step should be $O(Dn^3)$ because we don't provide a theoretical guarantee of the stepwise forward selection (which might be an interesting question we need to think of). We apologize for not being precise enough in our previous response.\n\nPractically, we have some evidence indicating that the variable selection step is not very time-consuming. First, we added new figures (Figure 6) showing the number of important variables chosen versus iterations. The number of important variables is the same as the number of GPs we need to fit. We find that they are all less than 10. These results show that we will do less than 10 GPs for each variable selection step. \n\nSecond, as we describe in our responses below, we test our method on all five function cases with max iteration 500. We find that the wall clock time of each case is still manageable and significantly smaller than that of ALEBO and SAASBO. The wall clock time here includes the time to do the variable selection. Therefore, we think our method is practical for larger iterations. We will try to include these results in the manuscript, although it might be challenging due to the page limit.\n", "role": "user" }, { "content": "Title: Response\nThank you for these additional experiments and explanations. I will make sure to give them the proper consideration, and reconsider my rating of the paper.", "role": "assistant" }, { "content": "Title: Response to the reviewer's comments \nThank you for the reviewer's further response. \n\nWe did some new tests on three synthetic functions with maximal iteration 1000. In this situation, we find that our method requires approximately 90 minutes for the Branin function and the Hartmann6 function, and 2 hours for the StyblinskiTang4 function (wall clock time). We also observed that as the maximum iteration increases, the time usage of different runs of the same function exhibits higher variance. For instance, for StyblinskiTang4, the minimum time is 6542 seconds, while the maximum time is 10839 seconds.\n Global/general Response:\n```\nWe would like to express our sincere gratitude to all the reviewers for their time and effort in reviewing our manuscript and providing valuable feedback. We have carefully considered the feedback provided by the reviewers and revised our manuscript accordingly, particularly in the experimental section. We have also responded to each reviewer individually for addressing the comments and questions.\n```\n", "role": "user" }, { "content": "I am grateful for the opportunity to discuss these points with the author, which has positively impacted the strength of this work.", "role": "assistant" } ]
XM_v85teqN	AutoML 2023 Conference	Reviewer_FmZd	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_FmZd, and the following are the paper you need to review:\n```<<XM_v85teqN>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper is about methods for training supernets, ie network architecture selection. It is mostly a detailed comparison of existing methods across existing benchmarks, with key parameters varied in order to establish performance ranking of methods and relate this to the underlying setting. Some rather subtle findings emerge and there is a reasonable attempt to relate this to salient features of the methods compared.\noverall_review: The paper is a large and thorough comparative assessment of some techniques that certainly fall within AutoML.\n\nThe paper is somewhat justified by a discussion of the context of the methods and setting, which go some way to motivating the study.\n\nExperiments are presented clearly and code is provided in support.\n\nOn the other hand,\n\nThe findings of the paper appear to be subtle.\n\nThe abstract does not highlight a clear finding.\n\nThe experiments would be best described as explorative, rather than designed carefully to assess a distinct hypothesis. This means that conclusions (where attempted) tend to be 'just-so', and are not then tested further. This makes the paper less valuable.\nreview_summary: My recommendation is based on decades of experience in ML and related areas, and an assessment largely of how clearly the authors have justified and presented their findings, and my estimation of how much interest/impact that will have. I accept that I could be wrong because this specific area of AutoML is not my area.\nactions_required_to_increase_overall_recommendation: The authors should improve the abstract to deliver a clearer finding / improve the chances of impact.\n\nThe authors should more clearly present testable hypotheses, or if this is not possible, then suggest ways in which findings could be tested in future.\nRating: 4 possibly reject, but has redeeming facets\nConfidence: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper\n", "role": "assistant" }, { "content": "Title: Response to Reviewer FmZd\nThank you for your thorough and informative review. Your comments on how to maximize the impact of this paper are much appreciated. We appreciate your positive view of our technical quality and have attempted to improve the clarity of the framing and, as you point out, thus also the potential impact of the manuscript. In response to your required actions to increase the overall recommendation:\n\nThe authors should improve the abstract to deliver a clearer finding / improve the chances of impact.\n\n* We agree that the previous abstract did not effectively highlight the contributions of this paper and we have revised the abstract, main text, and added a bulleted list of contributions to better highlight the noteworthy findings. We believe your assessment of the findings of our paper as subtle is accurate, there is not a single big splashy takeaway from this paper. However, the findings of this paper do address some key issues and ongoing debates within the NAS literature and we believe them very important to advancing our collective understanding of NAS. We have improved the writing to make clearer these prior debates and our intervention into them. \n\nThe authors should more clearly present testable hypotheses, or if this is not possible, then suggest ways in which findings could be tested in future.\n\n* The core motivating hypothesis of this paper can be stated as: magnitude-based pruning in NAS will often select suboptimal architectures as the result of the interacting phenomena of operation bias and discretization error. Of course, it would be nice if we could present a simpler, punchier hypothesis of “magnitude-based pruning is often ineffective for architecture selection in NAS” however prior studies have already been published framed around this hypothesis. Instead, we modify our hypothesis to adopt a more subtle variant to highlight what our study can contribute that prior work cannot: a holistic explanation of why this is the case. Again we hope that the new contributions section at the end of the Introduction helps to highlight these key takeaways and we appreciate your suggestion to better articulate them.\n", "role": "user" }, { "content": "Title: Improved abstract\nI agree the authors improved the abstract and made contributions clearer. \nMy grade is 5, (previously 4.)", "role": "assistant" } ]
XM_v85teqN	AutoML 2023 Conference	Reviewer_64dj	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_64dj, and the following are the paper you need to review:\n```<<XM_v85teqN>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The authors of this paper conducted an analysis of NAS algorithms to better understand the conceptual structure of supernet NAS algorithms and their performance. They conducted a literature review and an experimental case study to highlight the need for architecture selection and supernet training to be studied as interdependent processes. They then performed comparative evaluations of these algorithms. The authors stress the importance of systematic evaluations of NAS algorithms using equivalent amounts of training for compared algorithms and across multiple search spaces. The insights gained from this study can guide the development of future differentiable search methods.\noverall_review: NAS-Bench-201 is small and inexpressive as it contains only 15k architectures. The results on this benchmark are not generalizable and may not transfer to larger benchmarks. Therefore, larger NAS Benchmarks should be considered for evaluation. Or else, the authors should give justification for choosing only this benchmark. \n\nThere are many similar works in the past targeting architecture selection methods. It is difficult to understand the differences with other papers. A comparison table in Section 2.3 (or in the appendix) could provide a better understanding of the main contribution of this paper over others. \n\nIt may be good to plot the searched models for a few cases and compare the architectural specifications for different methods. \n\nThe size of the sampled/searched model matters. So, a deeper analysis of how model size varies with the architecture selection method may be potentially useful along with accuracy. \n\nThe authors considered only prune, valid, synflow, jacob, perturb for evaluation. What is the reason for choosing only these five? Are these the only methods available? \n\nThis work says “We did these experiments and here are the results with the interpretation of results” It may be good to provide any further guidelines for future researchers to improve the selection methods. \n\nreview_summary: The reviewer mainly focused on interpreting the results of benchmarking different architecture selection methods. The reviewer has gone through the methods chosen as a part of the study and provided a few steps to improve the article. \n\n\nactions_required_to_increase_overall_recommendation: As suggested before, the paper can be enhanced by conducting experiments on larger benchmarks and more architecture selection methods. The presented insights from the results on NAS-Bench-201 are confusing in a sense. It may be good separate zero-shot methods and give an in-depth analysis separately from the DARTS-based methods. \n\n\nThe following latest papers should be considered in references to support the literature review: \nWeight-sharing neural architecture search: A battle to shrink the optimization gap\nSuperNet in Neural Architecture Search: A Taxonomic Survey \nNeural Architecture Search Survey: A Hardware Perspective\nNeural Architecture Search: Insights from 1000 Papers\n\nMinor issues: \n\nReference to Table 3 is missing in lines 548 and 549\nLine 187: Here we focus in on the -> Here we focus on the\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer 64dj\nThank you for your highly informative and well researched review, your feedback is greatly appreciated. We are glad that, despite clarity issues in the original submission addressed in other reviews, you were able to find the core contributions of this paper and find them to be of great potential impact on the field. In response to your required actions to increase the overall recommendation:\n\nThe following latest papers should be considered in references to support the literature review: Weight-sharing neural architecture search: A battle to shrink the optimization gap SuperNet in Neural Architecture Search: A Taxonomic Survey Neural Architecture Search Survey: A Hardware Perspective Neural Architecture Search: Insights from 1000 Papers\n\n* We definitely agree that in the originally submitted version of this paper, we did not do a sufficient job distinguishing our work from past meta-analyses of NAS. We have since revised the writing to make our contributions more clear, denoted explicitly at the end of our introduction. We have additionally incorporated the perspectives of the suggested surveys, clear in our new discussion of operation bias and discretization loss.\n\nThe paper can be enhanced by conducting experiments on larger benchmarks and more architecture selection methods. The presented insights from the results on NAS-Bench-201 are confusing in a sense. It may be good separate zero-shot methods and give an in-depth analysis separately from the DARTS-based methods\n\n* We certainly agree that NAS-Bench-201 is smaller in size and that generalizable conclusions require the use of larger search spaces, which is the reason that we also utilized the DARTS search space (which is the same search space of NAS-Bench-301) for our combined results. Based on the comparison of the NAS-Bench-201 results and DARTS results revealing very little consistency in rankings of architecture-selection/supernet-training algorithm combinations across search spaces, we believed adding an additional search space would require considerable additional computational expenditure and be unlikely to clarify the core results of this paper. We have incorporated an additional section in Appendix C explaining our choice of NAS benchmarks. However, we do think expanding this study to other search spaces, especially for models other than computer vision, is a promising direction for future work. As for our selection of architecture search methods, we have added additional discussion of this experimental design in Section 4. While the approaches we selected are not exhaustive of all published architecture selection approaches, we do believe that they meaningfully represent the range of adopted approaches.\n", "role": "user" }, { "content": "Title: Reply to Authors\nThe authors did not provide a detailed response to many of the raised concerns or revised the manuscript based on the reviews provided. The paper still has flaws which are yet to be addressed. As suggested by other reviewers as well, the lack of diversity in terms of experiments and benchmarks makes it difficult to believe the true value of this work. The authors provided the results in a good way, however, the interpretation is confusing. Also, it is still unclear for future researchers to effectively utilize this study to design novel sampling and Supernetwork based methods. The reviewer is still skeptical about the limited study and difference between the previous meta-analysis works. However, this study is still important for the existing body of literature. Hence, the reviewer rating is still 6 (Borderline Leaning Accept). ", "role": "assistant" } ]
XM_v85teqN	AutoML 2023 Conference	Reviewer_B96n	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_B96n, and the following are the paper you need to review:\n```<<XM_v85teqN>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper broadly considers the task of architecture selection from supernet based methods for neural architecture search. The authors study DARTS- and DARTS -PT to claim that search space design is the most significant factor in NAS, with epochs required to train the supernet being more important than the data-set used. The paper makes a case for the importance of systemic evaluation of NAS algorithms. \noverall_review: This is a timely, well motivated paper with interesting insights. The case study provides a clear understanding of the limitations of DARTS as well as the differences in methods of addressing the instability of architecture weights in DARTS. The results are well presented and conclusions are well supported by available evidence. However, the diversity of data-sets and search spaces brings into question the quality of available evidence. The main negative aspect of the paper is the lack of diversity in the architectural search space, making their comments on the importance and trade-offs of architecture design, architecture selection methods and training lengths inadequately supported. \nreview_summary: I recommend a weak reject. I do not believe there are any technical flaws in the study, but its impact would be limited due to the weak evaluation. The paper addresses an important problem in differentiable NAS and conducts an effective literature survey to identify aspects of differentiable NAS that requires further study. However, a greater diversity of search spaces and data-sets are required to convince me of the conclusion and contributions.\nactions_required_to_increase_overall_recommendation: 1. I would appreciate the authors insights on whether their selection of search spaces and data-sets are sufficient to support the claims made in the paper.\n2. Why are full-ImageNet scale evaluations not included in the paper? I believe they would make the paper more insightful.\nRating: 5 marginally below the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer B96n\nThank you for your thoughtful review of our work. We appreciate your comments on the potential impacts of the ideas behind the paper and our evaluation of supernet training and architecture selection combinations in Table 2. In response to your required actions to increase the overall recommendation:\n\nI would appreciate the authors insights on whether their selection of search spaces and data-sets are sufficient to support the claims made in the paper.\n\n* We have also added a section discussing our choice of search spaces to Appendix C. While NAS-Bench-201 does represent a fairly limited search space, we have also incorporated results from the search space used by DARTS (as well as NAS-Bench-301), which is a much larger search space. Given the observed significant variation in results across search spaces, it did not seem that presenting results in a third search space would clarify the core message of our work. \n\nWhy are full-ImageNet scale evaluations not included in the paper? I believe they would make the paper more insightful.\n\n* Full-ImageNet results are not used for NAS-Bench-201 because the full version of the data set is not part of the benchmark. Full ImageNet results would certainly provide an interesting data point, and a scaled up version of this study would be a promising direction of future work. However, given that our results, detailed in Appendix F, indicate that the ranking of supernet training/architecture selection algorithm combinations varies more across different numbers of epochs of training than different image data sets, and that CIFAR 100 remains somewhat difficult for DARTS-space architectures, suggesting the difficulty of our existing settings. Further confidence in these findings through other difficult and different datasets in the future will certainly add to the concepts highlighted here. \n", "role": "user" }, { "content": "Title: Resposne To Authors\nThank you for your response. \n\nI understand that NASBench-201 does not have a full imagenet version of the data-set, but several existing papers on super-networks focus on full-scale ImageNet, going as far back as 2020 (DSNAS, SNAS etc.). However, I can understand that it is more reasonable to present results on CIFAR-10 and CIFAR-100, given that DARTS has a search space of 10^8 architectures. \n\nI find the idea of using zero cost methods for selecting architectures an interesting thought, and I am not surprised that it works. Whether the inherent FLOPs/Params bias is propagated by such zero cost selection methodologies is also an important question that I am not sure is evaluated here. Kindly correct me if I missed it. \n\nI completely agree with the authors remark on Line 331:\n'This suggests that evaluating NAS algorithms across different search spaces should be a significantly higher priority than evaluating across different data sets, and a fair comparison between algorithms'\n\nI think the ideas in the paper are interesting, but evaluating the proposed methods on more design spaces is central to such a paper which focuses on empirical evidence for the statements made. It may be a personal opinion, but I would suggest making Table 2 more succinct, and to separate out the conclusions drawn from epochs and search space (NASBench-201 and DARTS-space), as it is a bit difficult to register the underlined observations on the table. \n\nI have raised my score to 5, as I am still concerned about overall novelty and the number of search spaces the test is conducted on. I understand full-ImageNet may not be compulsory to support the statements made in the paper. \n", "role": "assistant" } ]
XM_v85teqN	AutoML 2023 Conference	Reviewer_QfjP	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_QfjP, and the following are the paper you need to review:\n```<<XM_v85teqN>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper presents a comprehensive assessment of architecture selection techniques along with various supernet training methods, highlighting the previously overlooked interdependence among them. It also showcases the effective integration of zero-shot NAS methods as new architecture selection mechanisms in supernet NAS training while evaluating the aforementioned techniques\noverall_review: I want to first of all highlight the effort for designing the framework by taking the search space, the datasets and the number of training epochs to evaluate different pairs of supernet training and architecture selection mechanisms. Running all possible combinations resulted in a very high computational cost (25400 GPU hours). This very high cost is the actual reason that pushes practitioners to focus on a single aspect which is mostly the supernet training. Obviously, they are aware of the issues related to the architecture selection mechanisms and the interplay between those and the supernet training strategies and not as claimed by the authors. The whole point of AutoML is not only to reduce the manual workload but also to keep the computational load under control. Apart from this major critisism, the take home-messages of the paper regarding the role of the search space, the training length and the dataset are not novel. \nreview_summary: Given the lack of novelty and the minor expected impact on the field of AutoML, this paper is not eligible for acceptance at this venue. \nactions_required_to_increase_overall_recommendation: it is hard to improve the paper since it is an empirical study of a reasonably large search space. The lack of consistent outcomes and generic take home messages across setups makes it just not worth acceptance.\nRating: 2 reject, significant issues present\nConfidence: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\n", "role": "assistant" }, { "content": "Title: Response to Reviewer QfjP\nThank you for your review. We are glad that you found the empirical results to be strong and the paper clearly written. In response to your comments:\n\nThe take-home messages are neither consistent across different setups nor novel.\n\n* We agree that the initially submitted version of this paper did not clearly present the key-takeaways of our study. We have revised the paper to make the core message much more clear (including a bulleted list of contributions at the end of the Introduction) and better demonstrate the novelty of this finding compared to prior work. \n\nThe paper provides an empirical verification of common intuitions.\n\n* We find value in empirical verification of common intuition, and view it as a hallmark of a sound scientific process. Though we certainly admit that it is not as flashy as a surprising finding. \n\nThe whole point of AutoML is not only to reduce the manual workload but also to keep the computational load under control\n\n* The intent of this work is to leverage toy problems and existing benchmarks to try and gain understanding of the relationship between supernet training and architecture selection. While we hope that the understanding from this paper will help to inform future studies, we do not think it is necessary, nor wise, for future authors or practitioners of NAS to necessarily replicate the immense computational study we undertake here – though perhaps the lesson learned from this work might persuade them to present the results of a novel supernet training algorithm in the context of multiple architecture selection methods for a broader picture of its potential success (and maximum performance), as this is not commonly done in the field right now. \n\nI want on the other side to highlight one major claim of this paper: \"demonstrate the previously underappreciated interdependence between various methods for supernet training and architecture selection mechanisms\". This is a controversial claim that is not well supported. \n\n* We apologize for the poor phrasing, and are absolutely not suggesting that other NAS authors are ignorant to the potential issues with architecture selection or have to always publish broad comparative studies like this one. Our assertion was not that studies which focus on a single aspect of the NAS process in the course of presenting a new algorithm are flawed, but merely that broad comparative studies such as this one are also necessary to develop generalizable explanations of the NAS process, and are less representative of the currently published works in the field (perhaps in part due to your observation that unintuitive and concise take-home messages are not typically highly valued). \n", "role": "user" }, { "content": "Thanks to the authors for their thoughtful response. It satisfactorily addresses most of my concerns, and I am considering revising the rating.", "role": "assistant" } ]
j5IKxqbpE2w	AutoML 2023 Conference	Reviewer_ZKBW	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_ZKBW, and the following are the paper you need to review:\n```<<j5IKxqbpE2w>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: In this work, the authors investigate how the makeup of an algorithm portfolio affects automated algorithm selection (AAS) from that portfolio. In particular, they explore meta-representation (SHAP and performance2vec) driven selection techniques for building such a portfolio. They ultimately show discrepancies in portfolios selected based on each meta-representation approach, and qualitative differences in AAS performance.\noverall_review: Strengths\n- New and interesting perspective on data-driven portfolio selection\n- Personalized version of their algorithm demonstrates better performance over existing greedy selection paradigm\n\nWeaknesses\n- No novel contribution on the meta-representation learning side\n- P2V and SHAP serve as reasonable baselines, but it's been shown many times that SHAP is not a robust explainability method -- I'm confident that its feature attributions serve as a grounded source of truth here either\n- Non-personalized versions of their method don't improve upon baseline\nreview_summary: Overall, this work presents an interesting approach to portfolio selection. While some of the experimental results are not very convincing, and the choice of meta-representation learning methods is not perfect (and there is also no work done on novel meta-representation learning methods), the idea itself is worth introducing to the community.\nactions_required_to_increase_overall_recommendation: Ideally, develop a new meta-representation learning method that avoids the need to learn an instance-specific portfolio.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer ZKBW\nWe thank you for your helpful and valuable comments. In the following, we address your individual comments. Please refer to the newly updated version of the paper for modifications.\n\n### No novelty ###\n\nThe design of meta-representations of optimization algorithms is not within the scope of this paper. It is a complex task that merits its own research; we plan to focus substantial research efforts towards this goal. The designed meta-representations must be evaluated in detail and demonstrated to work well before being used in downstream applications such as ours. For this reason, we have chosen to rely on previously proposed meta-representations and demonstrate how they can be used to select complementary algorithms for an algorithm portfolio. However, the suggested approach is general and can be applied using any meta-representations developed in the future.\n\n### Shapley values ###\n\nShapley values quantify individual contribution of input features to the overall performance of the ML model. These contributions capture essential information about what can be thought of algorithm-instance interaction, both on a local and a global level. For these reasons, we believe that the SHAP method can be effectively used as a possible algorithm meta-representation, despite being instance-specific.\n\n### Figures ###\n\nBottom row of Figure 1 unfortunately depends on the visualization library we used (networkx, v2.5). We keep this in mind for the camera-ready version, until which we will have more time to explore different visualization options and choose one that works better. Thank you for pointing this out!", "role": "user" }, { "content": "Thanks for the authors' response. The authors have answered almost all of my concerns. I remain supportive of my positive recommendation for this paper.", "role": "assistant" } ]
j5IKxqbpE2w	AutoML 2023 Conference	Reviewer_i5Kx	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_i5Kx, and the following are the paper you need to review:\n```<<j5IKxqbpE2w>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The authors evaluate how the performance of automated algorithm selection (AAS) depends on the portfolio of algorithms to choose from. There exists a tradeoff in the size of the portfolio: a larger portfolio contains the best algorithm with higher probability, but selecting the best algorithm out of a large number of candidates is also harder. The authors propose constructing a similarity graph from a set of algorithms based on their meta-representations, computed using performance2vec or Shapley values (SHAP). Then, given a certain constraint on the portfolio size, graph algorithms like Maximal Independent Sets (MS) and Dominating Sets (DS) can be leveraged to select a portfolio that is representative and nonredundant. \n\nThe authors conclude that performane2vec selects a smaller portfolio but achieves better AAS performance relative to the true optimal algorithm, while SHAP selects a larger portfolio but achieves worse AAS performance. The authors further conclude that the personalized method, where portfolios are constructed for each problem, generally outperforms other baselines. \n\nAll of the proposed methods outperform the case of having a full portfolio.\n\noverall_review: The methodology is valid and innovative. The experiments are well-designed and thoroughly conducted and the insights drawn from the experiments are explained well (e.g., Figure 4).\n\nThe methodology could be improved in terms of clarity of explanation. The paper may also benefit from adding a section on related work. \n\nreview_summary: The methodology is valid and innovative, and the experimental results demonstrate better behavior than other baselines. The authors also provide insight into interpreting the experimental results.There are some issues with clarity mentioned above. Therefore I recommend weak accept.\n\nactions_required_to_increase_overall_recommendation: - clarity issues mentioned above\n- adding a section on related work\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer i5Kx\nWe thank you for your helpful and valuable comments. In the following, we address your individual comments. Please refer to the newly updated version of the paper for modifications.\n\n### Question on Section 4 ###\n\nWe do agree there is no significant difference between the personalized portfolios boxplot and the greedy and full portfolio ones; we removed this claim to steer clear of further confusion.. However, on some problems we actually do observe a slight edge in favor of personalized portfolios across all considered dimensions and budgets (please refer to our Zenodo repo for all figures).\n\n### Clarity issues and related work ###\n\nWe revised the methodology section by explaining everything more clearly and by introducing mathematical notation to support the storyline. The “nested loop” of the p2v paragraph is now rewritten to increase accessibility. We added a reference for the empirical cumulative distribution function (ECDF); by definition, the value of ECDF at any specified value of the measured variable is the fraction of observations of the measured variable that are less than or equal to the specified value. In our experiments, area under ECDF curve is the performance of the algorithm up to a certain point. We also added a paragraph on related work in the Introduction section.\n", "role": "user" }, { "content": "Thanks for the authors' response. It adequately addresses the points I raised. I continue to recommend this work positively.", "role": "assistant" } ]
j5IKxqbpE2w	AutoML 2023 Conference	Reviewer_C1SQ	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_C1SQ, and the following are the paper you need to review:\n```<<j5IKxqbpE2w>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: It is well known that the effectiveness of automated algorithm selection heavily relies on the selection of a suitable algorithm portfolio. This task involves finding a balance between the flexibility of a large portfolio and the increased complexity of the AAS process. Typically, algorithms for the portfolio are chosen using a greedy selection approach based on their performance in some reference tasks. \n\nThe purpose of the paper is to explore alternative, data-driven techniques for algorithm portfolio selection. Tha authors introduce a method that involves learning meta-representations of algorithm behavior, constructing a graph based on the similarity of these representations, and applying a graph algorithm to select a final portfolio of diverse and non-redundant algorithms. \n\nThe effectiveness of two distinct meta-representation techniques is assessed by selecting complementary portfolios from a set of 324 CMA-ES algorithm variants for optimizing classic BBOB problems. The experimental results show that the performance based on performance2vec representations tend to favor small portfolios with negligible error in the AAS task, while the SHAP-based representations offer more flexibility but decreased AAS performance. Personalized portfolios perform comparably or slightly better than the greedy approach and outperform the full portfolio in all scenarios.\n\noverall_review: Pro : \n\nThe study's examination of the trade-off between the flexibility provided by a vast portfolio and the difficulty of selecting a subset of relevant algorithms is intriguing, and the proposed approach is relevant, even if not entirely novel.\n\nThe use of meta-representation of algorithm performance provides an innovative way of achieving AAS.\n\nThe experiments are thoroughly detailed, and the results are comprehensively analyzed, providing valuable insights for practitioners.\n\nThe graph representation utilized in this study provides interesting visualizations of the resulting portfolios, offering opportunities for a better understanding of the portfolio's building process.\n\nThe discovery that portfolios customized to individual preferences using Shapley meta-representations offer superior performance and greater flexibility is a noteworthy finding.\n\n\nCons : \n\nThe initial presentation of the problem could have been more comprehensive, providing more detailed background information and context.\n\nWhile the state of the art is briefly discussed, a more thorough analysis could have been beneficial to contextualize the proposed approach further.\n\nDespite being addressed, the computational cost induced by the proposed approach remains a significant limitation that requires careful consideration.\n\nExperiments are limited to one kind of algorithm. The performance scope could be biased by this restriction. \n\nreview_summary: The main contribution of this paper is the assessment of two meta-representation techniques, namely SHAP and performance2vec, to choose portfolios from CMA-ES variations for BBOB. Two types of portfolios have been considered: those related to the algorithm's overall behavior and those personalized to the algorithm's behavior for each problem. The experiments show that Performance2vec-based portfolios favor smaller sizes with minimal error, while SHAP-based portfolios are more adaptable but have lower performance. Personalized portfolios offer comparable or marginally better outcomes in nearly all situations compared to the classical greedy approach of selecting the best-performing algorithms based on reference tasks. \n\nThe paper focuses on BBO and specifically addresses the AAS problem for optimization, which is a well-known and extensively researched topic. The proposed methodology is highly pertinent, and the results obtained are intriguing for the BBO optimization community. Although some sections could have benefited from additional elaboration, the paper is well-written overall.\n\nactions_required_to_increase_overall_recommendation: The impact is limited. The possible use for wider and more different portfolios could have been examined.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer C1SQ\nWe thank you for your helpful and valuable comments. In the following, we address your individual comments. Please refer to the newly updated version of the paper for modifications.\n\n### Limited impact ###\n\nInvestigating the use of our approach on a wider range of inherently different algorithms is certainly one of the highest-priority TODOs in the future. However, the algorithm selection task itself is complex enough when we deal with a large family of similar algorithms, such as the modular CMA-ES framework. To this end, we believe that the essence of the observations made in this paper will propagate across different considered portfolios, meaning that the data-driven techniques for portfolio selection carry untapped potential for further performance gains.\n\n### Initial problem presentation/SOTA ###\n\nWe revised the introduction and added a paragraph on related work. Please note there is a surprising lack of works tackling this topic.\n\n### Computational cost ###\n\nWe do agree with the comment about the computational cost, which is one of the drawbacks listed in Section 5. We believe that, going forward, we can keep the computational cost low by reusing the explanations stemming from other studies on automated algorithm selection/configuration and making them more widely interpretable by including them in a semantic model such as the OPTION ontology.\n", "role": "user" }, { "content": "Thanks to the authors for their clarification. The response addresses most of my concerns. I continue to recommend this work positively.", "role": "assistant" } ]
j5IKxqbpE2w	AutoML 2023 Conference	Reviewer_BXUD	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_BXUD, and the following are the paper you need to review:\n```<<j5IKxqbpE2w>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper investigates \"data-driven\" approach for portfolio selection in the context of black-box optimization. The authors evaluated two ways to characterize algorithms -- performance2vec and Shapley based characterization. The paper also explores the idea of \"personalized\" portfolio building where a portfolio is built for each problem type. The experiments show that using algorithm meta-representation to build portfolios doesn't usually help with algorithm selection. However personalized portfolios yield the best performance.\noverall_review: Negative: \n- mostly exploratory work \n- I'm not convinced that Shapley values can serve as meta-representation of algorithms \n- No clear motivation (why is this a problem to begin with?)\n\nPositive: \n- Nice experimental setup \n\nreview_summary: Overall, I think the paper is a nice exploratory work of portfolio selection but there are a couple of comments I have about the paper (see in other sections). I'm not convinced that it has a large impact in the AutoML.\nactions_required_to_increase_overall_recommendation: I think the paper would benefit from better motivation for the work. For example, you mention that you used 324 algorithms in your study and the introduction allures to the problem of large number of algorithms in full portfolio. Maybe mention the number of algorithms in the beginning? I think that can motivate the paper better. \nI'm not convinced that Shapley values can serve as meta-representation of algorithms. Can you provide your reasoning why you chose to use Shapley values?\nRating: 4 possibly reject, but has redeeming facets\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer BXUD\nWe thank you for your helpful and valuable comments. In the following, we address your individual comments. Please refer to the newly updated version of the paper for modifications.\n\n### Better motivation ###\n\nWe made sure to better motivate the problem by mentioning the importance of the algorithm selection (AS) task when we deal with hundreds of possible algorithms. However, we strongly believe that investigating the trade-off between the flexibility of the portfolio and the complexity of the AS task is a research question that merits zooming into, what we have attempted to do with this work. Understanding these relations could further help practitioners to position their use-cases and employ the most adequate strategy, whether it is having more complementary algorithms in the portfolio to begin with, or reducing the AS complexity to a minimum.\n\n### Clarity ###\n\nWe added more details regarding Maximal Independent Sets (MIS) and Dominating Sets (DS).\n\n### Shapley values ###\n\nShapley values quantify individual contribution of input features to the overall performance of the ML model. These contributions capture essential information about what can be thought of algorithm-instance interaction, both on a local and a global level. For these reasons, we believe that the SHAP method can be effectively used as a possible algorithm meta-representation, despite being instance-specific.\n", "role": "user" }, { "content": "Thanks for the response provided by the authors. It answers most of my concerns, and I will consider raising the rating score.", "role": "assistant" } ]
lRFxsSgeYdE	AutoML 2023 Conference	Reviewer_DuT1	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_DuT1, and the following are the paper you need to review:\n```<<lRFxsSgeYdE>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: - The authors propose a solution that offers a speed-up during inference time by using multi-stage inference, switching between a simpler model and a more complex model such as XGBoost based on a certain fraction of the total points. The authors use LR by considering the application at only a subspace at a time, where, only the most accurate subspaces are combined for the first-stage inference.\noverall_review: Questions:\n\n- Line 123: Creating $b^n$ subsets\n\n If I understand correctly, for each numerical feature we have b bins, for a binary categorical feature we have 2 bins, while for categorical features we have n_unique_cat bins, so $b^n$ is not the correct term, since for a categorical feature we can have either fewer or more bins, making the term neither a minimum, neither a maximum constraint.\n\n- Could the authors describe the number of features and bins that were discovered for the considered datasets in Table 1?\n\n- Is there any reason for why standard deviations were not provided for Cases 1-4 in Table 1?\n\n- \"Other datasets are a representative subset of the 20+ public datasets from [31] that we used for evaluation.\"\n\n What was the criteria for filtering the datasets and only including a subset?\n\n- I would like the authors to add a few more datasets to their experiment. In particular, I would request 3 interesting public datasets from the AutoML Benchmark[1], KDDCup09_appetency, jasmine, Amazon_employee_access, and Higgs from [31] (binary classification problems that fit the experimental framework of the work).\n\n\n[1] Gijsbers, P., LeDell, E., Thomas, J., Poirier, S., Bischl, B., & Vanschoren, J. (2019). An open source AutoML benchmark. arXiv preprint arXiv:1907.00909.\n\n[2] The work.\n\nCorrections:\n\nLine 38: decision trees still outperform on tabular datasets\n\nI would correct decision trees to gradient boosted decision trees since by themselves decision trees are weak learners.\n\nLine 35: ML competitions with tabular data have been dominated not by deep learning models, but by gradient boosting models\n\n I believe ML competitions quite a few times include an ensembling of DNNs and GBDTs for the top solutions.\n\nLine 47, Deep learning tends to lose out to XGBoost on structured data and this trend is stronger when training data is limited in size.\n\nThe related work section seems to be lacking, since there have been several works that argue the opposite, that deep learning outperforms tabular data and the works do not mention any trends that deep learning does not perform well when training data is limited [1][2]. From personal experience, I have not observed the trend either. \n\nKadra, A., Lindauer, M., Hutter, F., & Grabocka, J. (2021). Well-tuned simple nets excel on tabular datasets. Advances in neural information processing systems, 34, 23928-23941.\n\nGorishniy, Y., Rubachev, I., Khrulkov, V., & Babenko, A. (2021). Revisiting deep learning models for tabular data. Advances in Neural Information Processing Systems, 34, 18932-18943.\nreview_summary: Pros: \n\n- The paper proposes an interesting solution on how to handle multi-stage inference and provides a motivating example (Figure 1) on the usage of LR by putting the instances into diverse bins.\n- The paper is clear.\n\nCons:\n\n- The method proposed by the authors affects a specific niche, that is an organization that needs to perform a large number of predictions continuously on an already trained model. I do not believe the proposed method affects the majority of researchers in the field of AutoML, since the time taken even without using the multi-stage inference is in the range of milliseconds. Additionally, the latency reduction comes with a performance deterioration. Moreover, preparing the method has a significant overhead on the practitioner since one needs to detect the feature importances (the authors also suggest XGBoost), then organize the data into bins, run LRwithBins and detect the correct hyperparameters based on the hyperparameter n for the number of important features and b number of bins via HPO. Lastly, the practitioner would need to detect the threshold that is also dataset specific for where the performance degrades from using LRBins and XGBoost.\n\n- There is a larger space overhead that would be required for every dataset depending on the number of features and number of bins.\nactions_required_to_increase_overall_recommendation: Addressing the questions and corrections.\nRating: 4 possibly reject, but has redeeming facets\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Addressing review\nDear Reviewer,\n\nThank you for your thoughtful review and valuable feedback on our paper. We have carefully addressed your comments and made the necessary changes to improve our manuscript. Please find our detailed responses below.\n\nPotential Impact On The Field Of AutoML:\nWe appreciate your insights on the potential impact of our work. We acknowledge that our method solves a specific practical challenge, but it is an important one. Industrial applications of ML often have significant resource demands for inference to support products, and our method aims to optimize the tradeoffs between performance, speed and resource usage.\n\nFigure 7 colors:\nThank you for pointing this out. We have updated the caption to include the information about the colors and datasets in Figure 7.\n\nTable 1 formatting:\nWe have fixed the formatting issue with Table 1 as you suggested.\n\nReferencing style:\nWe have carefully considered your suggestion, but have decided to maintain our current referencing style for consistency throughout the paper.\n\nLine 123: Creating subsets:\nYou are correct in your understanding of the binning process. We have revised the terminology in the manuscript to more accurately reflect the constraints of our binning approach.\n\nFeatures and bins in Table 1:\nDue to page limitations, we have not included these details in the paper. However, we have provided this information in the hyperparameters file of our code supplement, as you suggested.\n\nDataset selection criteria:\nWe chose a representative subset of the datasets showing a few that worked well with our approach and a few that did not work as well.\n\nAdditional datasets:\nThank you for suggesting additional datasets for our experiment. We have included 2 of the suggested datasets and cited the AutoML benchmark paper. Unfortunately, we were unable to include the other two datasets due to technical limitations.\n\nLine 38, 35, 47 corrections:\nWe appreciate your suggestions and have made the necessary corrections in our manuscript.\nWe have expanded our related work section to include the papers you mentioned and to better reflect the current state of research in the field.\n\nMethod's applicability and overhead:\nIn a major difference from academic ML research, industrial applications of ML spend much greater resources on inference, to support products. This can be illustrated by ChatGPT, which reportedly costs \\$700M/day to run, but only tens of millions USD total to train. The cost of inference is determined by CPU/GPU resources, which are closely related to mean latency. Whether the mean CPU latency is in seconds or milliseconds, it is multiplied by the large number of queries served to obtain total resources. Additionally, user-observed latency includes network latency and that of database lookups. Many online applications require latency below the cognitive threshold of 300ms, to appear instantaneous. We also clarifiy that the overhead that would be required for every dataset depending on the number of features and number of bins is small in general.\n\nWe hope that our revisions have addressed your concerns and improved the quality of our paper. Please also see our general comment to reviewers on our submission page. Once again, we appreciate your valuable feedback and the opportunity to enhance our work.\n Global/general Response:\n```\nDear Reviewers,\n\nThank you for the time and effort you have invested in evaluating our work. We have carefully addressed the comments provided and improved our revised manuscript as a result.\n\nNotably, most reviewers found our paper to be clear (\"The authors provide a clear description of the problem they are trying to solve, and their proposed solution appears to be logically consistent with the stated problem\", \"The work is very well written and is very easy to read. However, it has some details that could improve clarity\", \"The technical aspects of the proposed approach are presented in a clear manner, but the section on the algorithm could benefit from more detail\").\nWe further enhanced the clarity based on constructive feedback received:\n- We added pseudocode, making it easier for readers to understand the details and follow the proposed method. \n- We provided requested details.\n- We fixed a variety of small glitches and omissions noted by reviewers, and improved formatting as requested.\n- We added two new datasets in the empirical evaluation.\n- Admittedly, we may have understated the importance of AutoML in our original manuscript. Therefore, our revised manuscript articulates the crucial role of AutoML in the success of multistage inference --- to facilitate key tradeoffs between the two stages of inference and to configure the first stage. \n- We are particularly grateful to the reviewer who provided comments on our code, and we addressed those as well, making the code easier to understand and reuse.\n\nWe also addressed several misconceptions raised by the reviewers.\n- Optimizing inference for real-time ML systems is not a niche area but rather a mainstream concern in the industry, where efficient ML inference impacts resource consumption and affects business profitability. This is critical in a growing number of real-time ML deployments --- within Web search, ad networks, many applications in social networks, various recommendation systems, etc.\n- Optimizing mean inference latency in terms of (tens and hundreds of) milliseconds is not a limitation, but rather a typical industry context. Optimizing mean CPU latency (without increasing the parallelism) helps reduce overall CPU resources.\n- The widespread impact of our proposed solution in real-time industry applications should be clearer when we disclose our affiliation and prior publications in the published paper (but we cannot do this during blind review). Our solution is deployed in a real-time industry ML platform that performs millions inferences per second for a variety of applications with billions (with a B) users worldwide.\n\nWe trust that our revisions address the reviewers' concerns and enhance the quality of the manuscript, making it suitable for publication.\n\nThank you once again for your valuable feedback and the opportunity to improve our work.\n\nThe Authors\n```\n", "role": "user" }, { "content": "Thanks for the response provided by the authors. It answers most of my concerns, and I may consider adjusting the rating upwards.", "role": "assistant" } ]
lRFxsSgeYdE	AutoML 2023 Conference	Reviewer_UffP	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_UffP, and the following are the paper you need to review:\n```<<lRFxsSgeYdE>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: In many machine learning applications, real-time inference can still present a bottleneck. Hence, it is common to separate ML code into services that are queried by product code using APIs. While this approach simplifies product code by abstracting away ML internals and clarifies software architecture, it also introduces network latency and additional CPU overhead.\n\nIn this paper, the authors propose simplifying inference algorithms and embedding them directly into the product code, thereby reducing network communication. Their approach can optimize over half of the inputs for public datasets and a high-performance real-time platform that deals with tabular data. By using AutoML for both training and inference, they achieved a reduction in inference latency, CPU resources, and network communication between the application front-end and the ML back-end.\n\noverall_review: Overall Review\n\nPro : \n\nThe empirical evaluation of the proposed method is based on experiments conducted on both public and company-internal datasets from a high-performance ML platform that serves millions of real-time decisions per second. \n\nThe evaluation shows that the proposed method, LRwBin, achieves competitive performance. In particular, the standalone performance of LRwBin demonstrates the efficiency of the approach.\n\nCode for this project is publicly available\n\nCons \n\nAlthough the proposed system is mainly based on previously existing techniques, it still makes a valuable contribution to the field by combining these techniques in a novel way. \n\nSome parts of the paper lack technical explanations and details, which could make it difficult for readers to fully understand the system. Moreover, while the AutoML part is mentioned, it could have been better highlighted as an important aspect of the proposed approach.\n\nAdditionally, the paper discusses the tradeoff between ML performance and efficiency of inference, but this aspect could have been given more attention and elaboration, especially since it is an important consideration for practical applications.\n\nreview_summary: The paper proposes a multi-stage inference approach that includes a simplified first stage that can be embedded into product code to reduce network communication and lower CPU overhead with negligible loss in ML performance. The proposed approach relies on AutoML for the optimal choice of hyperparameters, the composition of combined bins, and choosing the separation threshold between stages of inference. The approach improves the average resource efficiency of inference, which also improves energy efficiency, and is compatible with hardware acceleration. This work could lead to significant enhancements in resource efficiency and latency.\n\nThis is a technical paper that presents a system together with an extensive empirical evaluation. The outcomes may be important in terms of required resources and are relevant for handling real-time inference in large applications context. \n\nThe AutoML contribution is limited since AutoML is only used as a tool in the system. Moreover, this part lacks details. \n\nactions_required_to_increase_overall_recommendation: Better highlight AutoML part. \nTry to better highlight the contribution within the global system.\nRating: 4 possibly reject, but has redeeming facets\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Addressing review\nDear Reviewer,\n\nThank you for your valuable feedback on our paper. We appreciate your insights and have made several revisions to address your concerns. Please find our detailed responses below.\n\nHighlight impact:\nWe note that real-time inference dominates resource usage in industry applications by 1-2 orders of magnitude compared to training.\nIn a major difference from academic ML research, industrial applications of ML spend much greater resources on inference, to support products. This can be illustrated by ChatGPT, which reportedly costs \\$700M/day to run, but only tens of millions USD total to train. The cost of inference is determined by CPU/GPU resources, which are closely related to mean latency. Whether the mean CPU latency is in seconds or milliseconds, it is multiplied by the large number of queries served to obtain total resources. Additionally, user-observed latency includes network latency and that of database lookups. Many online applications require latency below the cognitive threshold of 300ms, to appear instantaneous.\n\nAlgorithm detail and clarity:\nWe have added pseudocode and more comments to provide a clearer and more detailed presentation of the algorithm, making it easier for readers to understand the system.\n\nTechnical explanations and AutoML integration:\nWe have expanded our discussion on using AutoML and its importance as an aspect of the proposed approach. This should provide a more comprehensive understanding of our system and its practical applications.\n\nTradeoff between ML performance and efficiency of inference:\nWhile we would have liked to provide more attention and elaboration on this tradeoff, we are limited by the space constraints of the paper. We believe our current discussion provides a reasonable balance between exploring this tradeoff and addressing other important aspects of our work.\n\nOnce again, we appreciate your valuable feedback and the opportunity to enhance our work. Please also see our general comment to reviewers on our submission page. Thanks!\n Global/general Response:\n```\nDear Reviewers,\n\nThank you for the time and effort you have invested in evaluating our work. We have carefully addressed the comments provided and improved our revised manuscript as a result.\n\nNotably, most reviewers found our paper to be clear (\"The authors provide a clear description of the problem they are trying to solve, and their proposed solution appears to be logically consistent with the stated problem\", \"The work is very well written and is very easy to read. However, it has some details that could improve clarity\", \"The technical aspects of the proposed approach are presented in a clear manner, but the section on the algorithm could benefit from more detail\").\nWe further enhanced the clarity based on constructive feedback received:\n- We added pseudocode, making it easier for readers to understand the details and follow the proposed method. \n- We provided requested details.\n- We fixed a variety of small glitches and omissions noted by reviewers, and improved formatting as requested.\n- We added two new datasets in the empirical evaluation.\n- Admittedly, we may have understated the importance of AutoML in our original manuscript. Therefore, our revised manuscript articulates the crucial role of AutoML in the success of multistage inference --- to facilitate key tradeoffs between the two stages of inference and to configure the first stage. \n- We are particularly grateful to the reviewer who provided comments on our code, and we addressed those as well, making the code easier to understand and reuse.\n\nWe also addressed several misconceptions raised by the reviewers.\n- Optimizing inference for real-time ML systems is not a niche area but rather a mainstream concern in the industry, where efficient ML inference impacts resource consumption and affects business profitability. This is critical in a growing number of real-time ML deployments --- within Web search, ad networks, many applications in social networks, various recommendation systems, etc.\n- Optimizing mean inference latency in terms of (tens and hundreds of) milliseconds is not a limitation, but rather a typical industry context. Optimizing mean CPU latency (without increasing the parallelism) helps reduce overall CPU resources.\n- The widespread impact of our proposed solution in real-time industry applications should be clearer when we disclose our affiliation and prior publications in the published paper (but we cannot do this during blind review). Our solution is deployed in a real-time industry ML platform that performs millions inferences per second for a variety of applications with billions (with a B) users worldwide.\n\nWe trust that our revisions address the reviewers' concerns and enhance the quality of the manuscript, making it suitable for publication.\n\nThank you once again for your valuable feedback and the opportunity to improve our work.\n\nThe Authors\n```\n", "role": "user" }, { "content": "Thank you for the authors' reply. It answers most of my concerns, and I may consider adjusting the rating upwards.", "role": "assistant" } ]
lRFxsSgeYdE	AutoML 2023 Conference	Reviewer_BtQb	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_BtQb, and the following are the paper you need to review:\n```<<lRFxsSgeYdE>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper discusses the challenges in developing robust high-performance machine learning (ML) systems and proposes multi-stage inference as a solution to the bottleneck of real-time inference. The authors argue that decision tree models, such as XGBoost, outperform deep learning models on structured data, and that a simple first-stage model embedded in the product code can reduce latency and CPU resource usage. The authors validate their proposed multi-stage inference on diverse public tabular datasets and in a high-performance production system, showing minor declines in accuracy and ROC AUC but significant latency and CPU resource consumption gains.\noverall_review: The paper offers several strengths:\n\n-The paper describes a well-defined problem of creating multi-stage inference systems and proposes a practical solution to automate the process using AutoML.\n-The proposed solution appears to be logically consistent with the stated problem and the experiments conducted to evaluate the approach are well-designed, producing results that support the authors' claims.\n-The work is very well written and is easy to read.\n-The multi-stage method proposed in the paper is a highlight, as it reduces the number of calls to the model's API, saving time and resources.\n\nThe authors provide a clear description of the infrastructure experiment setting, which is important for understanding the results and contributions of the work.\n\nThere are some concerns with the paper, including:\n\n-As noted in the Clarity section, comparison between XGBoost and a hybrid model is confusing, and it is unclear if the hybrid model is the same as LRwBins or something different. Clarifying this is important to relate and improve traceability between tables 1 and 2.\n-More information about the experimental setup could further clarify the contributions of the work. Specifically, information about the infrastructure for the remote model and product code.\n\nreview_summary: The paper propose a practical solution to automate the process of creating multi-stage inference systems for real-time inference problem. The experiments conducted to evaluate the proposed approach are well-designed, and the results support the authors' claims. The paper is well-written and easy to read. While there are some minor details that could be improved for clarity, overall, the work presents a valuable contribution to the field of AutoML and Multi-Stage inference.\nactions_required_to_increase_overall_recommendation: NA\nRating: 8 accept, good paper\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Addressing comments\nDear Reviewer,\n\nThank you for the insightful comments that help improve the manuscript.\nWe have clarified the hybrid model confusion in the caption of Table 1 and have also modified Table 1 according to your comments.\nIndeed, our work's contribution to AutoML is beyond hyperparameter tuning, with the main focus on configuring multistage inference to ensure proper tradeoff between efficiency and coverage. A key empirical result is that the first stage can typically handle over 50% of inferences, and this does not require hand tuning.\n\nWe appreciate your valuable feedback and the opportunity to enhance our work. Please also see our general comment to reviewers on our submission page. Thank you!\n Global/general Response:\n```\nDear Reviewers,\n\nThank you for the time and effort you have invested in evaluating our work. We have carefully addressed the comments provided and improved our revised manuscript as a result.\n\nNotably, most reviewers found our paper to be clear (\"The authors provide a clear description of the problem they are trying to solve, and their proposed solution appears to be logically consistent with the stated problem\", \"The work is very well written and is very easy to read. However, it has some details that could improve clarity\", \"The technical aspects of the proposed approach are presented in a clear manner, but the section on the algorithm could benefit from more detail\").\nWe further enhanced the clarity based on constructive feedback received:\n- We added pseudocode, making it easier for readers to understand the details and follow the proposed method. \n- We provided requested details.\n- We fixed a variety of small glitches and omissions noted by reviewers, and improved formatting as requested.\n- We added two new datasets in the empirical evaluation.\n- Admittedly, we may have understated the importance of AutoML in our original manuscript. Therefore, our revised manuscript articulates the crucial role of AutoML in the success of multistage inference --- to facilitate key tradeoffs between the two stages of inference and to configure the first stage. \n- We are particularly grateful to the reviewer who provided comments on our code, and we addressed those as well, making the code easier to understand and reuse.\n\nWe also addressed several misconceptions raised by the reviewers.\n- Optimizing inference for real-time ML systems is not a niche area but rather a mainstream concern in the industry, where efficient ML inference impacts resource consumption and affects business profitability. This is critical in a growing number of real-time ML deployments --- within Web search, ad networks, many applications in social networks, various recommendation systems, etc.\n- Optimizing mean inference latency in terms of (tens and hundreds of) milliseconds is not a limitation, but rather a typical industry context. Optimizing mean CPU latency (without increasing the parallelism) helps reduce overall CPU resources.\n- The widespread impact of our proposed solution in real-time industry applications should be clearer when we disclose our affiliation and prior publications in the published paper (but we cannot do this during blind review). Our solution is deployed in a real-time industry ML platform that performs millions inferences per second for a variety of applications with billions (with a B) users worldwide.\n\nWe trust that our revisions address the reviewers' concerns and enhance the quality of the manuscript, making it suitable for publication.\n\nThank you once again for your valuable feedback and the opportunity to improve our work.\n\nThe Authors\n```\n", "role": "user" }, { "content": "Thank you for the authors' feedback. The response covers nearly all of the issues I raised. I will continue to recommend accepting this work.", "role": "assistant" } ]
lRFxsSgeYdE	AutoML 2023 Conference	Reviewer_LbfZ	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_LbfZ, and the following are the paper you need to review:\n```<<lRFxsSgeYdE>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper presents an approach to speed up inference (on average) in certain deployed applications, through multi-stage inference. The core proposal is that instead of using remote calls to offload inference to servers, the customer facing or deployed component can perform inference on part of the data. \nThe approach is meant only for tabular data. The approach achieves this through the proposed Logistic Regression with Bins (LRwBins) algorithm. The authors present a detailed description of the approach, and some example figures to illustrtae the same. The autyhors also provide a section on implementation. Empirical evaluation is performed with both open source and proiprietary data sets. Additionally, figures are presented to show comparative performance degradation. Furthermore, authors present some additional details on limitations as well as future directions.\noverall_review: The paper is interesting, but is lacking in several aspects. The narrative clarity is poor and information is not well presented which makes it difficult to understand the details. The technical aspects while sound, authors do not present edge cases or generalization study. The empirical evaluation uses only a few datasets. While some of them are sizable, the applicability of this approach in a broader scenario and different data characteristics is not clear. \nOn the positive side, the authors provide adequate references. The analysis performed is also good, especially to compare loss of accuracy. The case of RPC causing delays in inference which is a valid problem the authors are proposing a solution for.\nreview_summary: I would recommend the paper not be accepted in the current format. \nactions_required_to_increase_overall_recommendation: 1. Improve clarity\n2. Improve empirical evaluation\nRating: 2 reject, significant issues present\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Addressing review\nDear Reviewer,\n\nThank you for your valuable feedback on our paper. We appreciate your insights and have made several revisions to address your concerns. Please find our detailed responses below.\n\nReal-world adoption and impact:\nBoth the importance and the widespread adoption of the problem we solve have been described in a recent KDD publication not shown during blind review. In particular, our ML platform is deployed for over a hundred real-time industry applications, with millions of inferences per second.\n\nAlgorithm clarity and design choices:\nWe have added pseudocode to improve the clarity of our algorithm presentation, making it easier for readers to understand the details. We have also fixed the long sentence on Line 150 and updated Figure 1 caption to better connect it to LRwBins. Additionally, we added to the Figure 7 caption clarifying the missing information.\n\nTechnical explanations and AutoML integration:\nWe have expanded our discussion on using AutoML and its importance as an aspect of the proposed approach. This should provide a more comprehensive understanding of our system and its practical applications.\n\nAdditional comments:\nWe have made efforts to improve the narrative clarity of our paper by revising sections and adding pseudocode for a better understanding of our proposed algorithm. We acknowledge that presenting edge cases and a generalization study on different data characteristics would be beneficial. However, space constraints make it difficult to include these details in the current version of our paper. \nWe plan to investigate these aspects in future work to provide a more comprehensive evaluation of our approach in the arxiv version. Regarding the empirical evaluation, we have added more datasets to demonstrate the applicability of our approach in various scenarios. \n\nOnce again, we appreciate your valuable feedback and the opportunity to enhance our work. Please also see our general comment to reviewers on our submission page. Thank you!\n Global/general Response:\n```\nDear Reviewers,\n\nThank you for the time and effort you have invested in evaluating our work. We have carefully addressed the comments provided and improved our revised manuscript as a result.\n\nNotably, most reviewers found our paper to be clear (\"The authors provide a clear description of the problem they are trying to solve, and their proposed solution appears to be logically consistent with the stated problem\", \"The work is very well written and is very easy to read. However, it has some details that could improve clarity\", \"The technical aspects of the proposed approach are presented in a clear manner, but the section on the algorithm could benefit from more detail\").\nWe further enhanced the clarity based on constructive feedback received:\n- We added pseudocode, making it easier for readers to understand the details and follow the proposed method. \n- We provided requested details.\n- We fixed a variety of small glitches and omissions noted by reviewers, and improved formatting as requested.\n- We added two new datasets in the empirical evaluation.\n- Admittedly, we may have understated the importance of AutoML in our original manuscript. Therefore, our revised manuscript articulates the crucial role of AutoML in the success of multistage inference --- to facilitate key tradeoffs between the two stages of inference and to configure the first stage. \n- We are particularly grateful to the reviewer who provided comments on our code, and we addressed those as well, making the code easier to understand and reuse.\n\nWe also addressed several misconceptions raised by the reviewers.\n- Optimizing inference for real-time ML systems is not a niche area but rather a mainstream concern in the industry, where efficient ML inference impacts resource consumption and affects business profitability. This is critical in a growing number of real-time ML deployments --- within Web search, ad networks, many applications in social networks, various recommendation systems, etc.\n- Optimizing mean inference latency in terms of (tens and hundreds of) milliseconds is not a limitation, but rather a typical industry context. Optimizing mean CPU latency (without increasing the parallelism) helps reduce overall CPU resources.\n- The widespread impact of our proposed solution in real-time industry applications should be clearer when we disclose our affiliation and prior publications in the published paper (but we cannot do this during blind review). Our solution is deployed in a real-time industry ML platform that performs millions inferences per second for a variety of applications with billions (with a B) users worldwide.\n\nWe trust that our revisions address the reviewers' concerns and enhance the quality of the manuscript, making it suitable for publication.\n\nThank you once again for your valuable feedback and the opportunity to improve our work.\n\nThe Authors\n```\n", "role": "user" }, { "content": "Title: Comments on Revisions\nA general suggestion is to highlight changes ( preferably in a different color ) to aid comprehensibility in an iterative review process.\n\nPositives\n- Adding pseudocode\n- Improved clarity to some extent. \nNegatives\n- In Alg 1, line 3 -- special handling of categorical variables is not elucidated upon\n- While the authors have added more datasets, it is not evident to the reader the composition of the features--- since that was one of the issues pointed out earlier that b^n can become large.\n\nIn summary, the authors have revised the paper to some degree and it is evident. The reviewer's recommendation however does not change owing to aforementioned points.", "role": "assistant" }, { "content": "Title: Addressing new comment\nDear Reviewer,\n\nThank you for taking the time to review the edits to our paper and provide feedback. We appreciate your suggestions and will address the concerns you raised.\n\nRegarding the special handling of categorical variables in Algorithm 1, the special handling is not explicitly elucidated in line 3. However, we do provide a more detailed description of this process on lines 125-127 of the paper. \n\nWe understand your concern about the potentially large size of b^n, particularly for categorical features with high cardinality. In our paper, we have observed that, for the datasets used in our experiments, the important features are not categorical with high cardinality. That is, high-cardinality categorical features, when present, are handled by the second stage. Consequently, we chose not to focus on this less critical aspect. To address this issue, one could either avoid using categorical features with excessively high cardinality when constructing the bins or opt not to use the method altogether in such cases.\n\nAdditionally, we would like to emphasize the robustness of our approach. By incorporating a second stage model, our technique ensures that even when applied to cases with important high-cardinality categorical features, the results do not deteriorate. This aspect further supports the applicability and flexibility of our method.\n\nThank you once again for your valuable feedback.", "role": "user" }, { "content": "I am delighted to have engaged in the above conversation with the author, which undoubtedly enhances the strength of this paper.", "role": "assistant" } ]
lmtNt--6dw	AutoML 2023 Conference	Reviewer_UhWx	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_UhWx, and the following are the paper you need to review:\n```<<lmtNt--6dw>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper performs an empirical study that compares early stopping-based NAS methods: HB, ASHA, and BOHB. The main question that it seeks to answer is how to best allocate resources, both for the NAS experiment overall (Q1) and for a single trial (Q2).\n\nFor Q1, the paper looks at parallel efficiency vs. number of GPUs allocated, showing that with up to 32 GPUS, the parallel efficiency is above 0.75, but for 64 GPUs, the efficiency drops to 0.6-0.65. For Q2, the paper suggest increasing parallelism in order to get a reduced runtime.\noverall_review: Strengths:\n* The main strength of this paper is an empirical comparison of the methods, which can help NAS practitioners. The answer to RQ1 was clear and conclusive and perhaps provides an insight that was not previously known\n\nWeaknesses:\n* For RQ2, is the main takeaway to increase parallelism as much as possible? This could be made clearer throughout. Although this is not too surprising, it is still nice for practitioners to know.\n* The plots should have error bars.\n* One question I expected to get an answer to is the following: suppose I had a total compute budget along with a instantaneous compute budget, how should I trade off adaptivity vs wall time?\nreview_summary: Although the paper gives a nice empirical comparison of three early stopping-based methods, I think the impact at this point is still relatively small.\nactions_required_to_increase_overall_recommendation: * I think if this paper performed a more systematic evaluation and provided clearer results and takeaways (esp for RQ2), then it would have increased impact.\n* (Copied from above) One question I expected to get an answer to is the following: suppose I had a total compute budget along with a instantaneous compute budget, how should I trade off adaptivity vs wall time?\n\nEdit I've increased the score after reading responses. Thank you to the authors for the prompt response.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Reply to Reviewer UhWx\nWe thank the reviewer for the assessment of our work and the provided feedback. We address all concerns raised by the reviewer below. \n\n> For RQ2, is the main takeaway to increase parallelism as much as possible? This could be made clearer throughout. Although this is not too surprising, it is still nice for practitioners to know.\n\nTo clarify: the main takeaway is to increase the parallelism of the NAS loop as much as possible at the cost of the parallelism of the single trials. We agree that this was not made clear enough in the paper and thank the reviewer for pointing this out. To make this more clear, we have modified Section 5 to answer the research questions from the introduction and present the main takeaway. Additionally, we highlight this finding also already in the introduction (Section 1) after formulating the research questions. \n\n> The plots should have error bars.\n\nWe agree that showing error bars is important for quantifying the reliability of the results and thank the reviewer for pointing this out. As the lines in the Figures are close together (see e.g. Fig. 2), we have decided for omitting them in the plots in the main paper. However, we do now include the untruncated plots with error bars in Appendix A1. \n\n> One question I expected to get an answer to is the following: suppose I had a total compute budget along with a instantaneous compute budget, how should I trade off adaptivity vs wall time?\n\nWe would like to ask the reviewer to elaborate on this question, especially on what is meant with an “instantaneous compute budget” and with “adaptivity”. We will try to answer the question and include the answer in the paper once we have received this.\n\n> I think if this paper performed a more systematic evaluation and provided clearer results and takeaways (esp for RQ2), then it would have increased impact.\n\nWe have extended our evaluation in a more systematic fashion to include larger models (Subsection 4.4) as well as different types of GPUs (Subsection 4.5, Figure 6). In addition to our modification of Section 5 to highlight our main takeaway, we hope to have addressed all concerns raised by the reviewer and would be happy to address any follow-up questions by the reviewer. \n\n\n Global/general Response:\n```\nWe would like to thank all reviewers for their valuable feedback and suggestions to improve the paper. We have significantly extended our experimental evaluations, now including runs on two more types of GPUs (NVIDIA V100 and AMD MI50). Moreover, another NAS search space, where sampled models are trained on the complete imagenet dataset (see Subsections 4.4 and 4.5 and Figure 6), is included to the revised manuscript. These evaluations are inline with the previous conclusions and main takeaway, which has been that an increase of the NAS parallelism is favorable, even at the cost of decreasing the parallelism of the single trials. Due to space constraints, the Figures with lines including standard deviations as uncertainty measurements are now included in Appendix A1. As it took us some time and a large amount of computing resources to run the new experiments, four days were needed for updating the paper. All major additions to the paper are highlighted in red. \n\nDear reviewers,\n\nas the end of the author-reviewer discussion period is approaching, we are reaching out to resolve any additional questions you may have regarding our responses to your comments and the revised paper. We want to ensure that all your concerns are thoroughly addressed before the deadline.\n\nWe thank you for your time and consideration. \nDear Area Chair,\ndear Reviewers,\n\nwe would kindly remind you that the second phase of the discussion period ends this Monday. We are in your service if any additional clarification to our responses is needed.\n\nWe thank you for your time and consideration.\n```\n", "role": "user" }, { "content": "Title: Response\n> We would like to ask the reviewer to elaborate on this question, especially on what is meant with an “instantaneous compute budget” and with “adaptivity”. We will try to answer the question and include the answer in the paper once we have received this.\n\nBy \"adaptivity\", I mean the parallelism of NAS (lower parallelism = higher adaptivity). By \"instantaneous compute budget\", I mean the maximum parallelism at any given time. By total compute budget, I mean total resources used overall across time (think total AWS cost, for instance).", "role": "assistant" }, { "content": "Title: Response to the Reviewers Question\nWe thank the reviewer for the clarification. From our understanding, the question is aimed at how to balance the total amount of computational resources (e.g. AWS cost) with the amount of resources used at the same time (e.g. number of GPUs allocated concurrently).\n\nThe answer to this question is a combination of RQ1 (\"How many resources should be allocated in total for a NAS?\") and RQ2 (\"How many resources should be allocated for each NAS trial?\") from our paper. Our findings in regard to RQ1 suggest that using a total number of 32 workers in parallel is favorable for maintaining a high computational efficiency (>0.75). Using more than 32 workers leads to lower computational efficiency and therefore to the waste of resources. Our findings in regard to RQ2 suggest that as much parallelism of the NAS loop as possible is favorable for reducing the runtime. The answer to the question is therefore: It would be best to allocate the total available compute budget in such a fashion that an instantaneous compute budget of 32 workers (i.e., GPUs) can be used in parallel and at the same time select the adaptivity in such a fashion that the parallelism of the NAS loop is high (as many trials as possible on those 32 workers in parallel).\n\nModifications to the paper are unfortunately no longer possible, but we believe to have addressed this already in our revised paper in the conclusion in lines 311 to 318 (new text marked in red).", "role": "user" }, { "content": "I appreciate the fruitful discussion with the author, which has positively impacted the strength of this work.", "role": "assistant" } ]
lmtNt--6dw	AutoML 2023 Conference	Reviewer_Kwnc	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_Kwnc, and the following are the paper you need to review:\n```<<lmtNt--6dw>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The authors experimentally investigate the impact of parallelism of the NAS loop on the runtime and quality of the outcome using three common early-stopping based methods: Hyperband, Asynchronous successive halving (ASHA), and Bayesian optimization hyperband (BOHB). In particular, two research questions were investigated: how many resources should be allocated to a single NAS task, and how many resources should be allocated to a single NAS trial (evaluating a single architecture configuration). The authors conclude that using up to 32 GPUs per NAS task is advisable for general early-stopping based NAS methods, and that having a smaller number of trials per GPU (e.g., 1-4 trials per GPU) gives stronger anytime performance in terms of validation accuracy than setting 8 trials per GPU.\n\n\noverall_review: The experiments are well-designed and thoroughly conducted and the conclusions drawn from the experiments are reasonable. The insights in this paper could provide useful guidance for ML practitioners to allocate resources for NAS. \n\nThe paper is also well-written. The motivation, related work, and experiments are clearly explained. \nreview_summary: The methodology seems sound, the explanations are clear, and the insights are beneficial to a wide range of practical situations. Therefore I recommend accept.\n\nactions_required_to_increase_overall_recommendation: It would be nice if the authors could address the three points mentioned in \"clarity\".\nRating: 8 accept, good paper\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Reply to Reviewer Kwnc\nWe thank the reviewer for the assessment of our work and the provided feedback. We address all concerns raised by the reviewer below. \n\n> How exactly does the global gradient \"degrade\" (line 129) with a smaller number of devices? Would appreciate it if the authors could clarify that more.\n\nWe agree with the reviewer that the sentence induces confusion and thank her/him for pointing this out. It is only possible to use the learning rate scaling technique for mitigation of the degradation of generalization performance if the number of devices is small. If the number of devices is large, even using this technique cannot help avoid the degradation of generalization performance. We have reformulated this sentence in the revised manuscript. \n\n> The fact that E_G in Equation (3) is bounded above by 1 implies that T_G is bounded below by 1/G T_1. Is this due to overhead created by communication and scheduling? Though it may be common sense for domain experts, it would still be nice if this could be explicitly explained.\n\nIndeed, T_G is bounded below by 1/G T_1. Since T_1 is the runtime of the problem on a single worker and T_G is the runtime of the problem on G workers, this lower bound represents the perfect scaling case. In some cases, this value can go below the threshold (which would mean an E_G value of larger than 1), which is referred to as superlinear speed-up. We have added a sentence on this in the revised paper.\n\n> Possible typo: \"substantial\" rather than \"substational\" on line 109.\n\nWe have fixed this typo in the revised paper. \n\n> It would be nice if the authors could address the three points mentioned in \"clarity\".\n\nWe hope to have addressed all concerns raised by the reviewer and would be happy to address any follow-up questions by the reviewer. \n\n\n\n Global/general Response:\n```\nWe would like to thank all reviewers for their valuable feedback and suggestions to improve the paper. We have significantly extended our experimental evaluations, now including runs on two more types of GPUs (NVIDIA V100 and AMD MI50). Moreover, another NAS search space, where sampled models are trained on the complete imagenet dataset (see Subsections 4.4 and 4.5 and Figure 6), is included to the revised manuscript. These evaluations are inline with the previous conclusions and main takeaway, which has been that an increase of the NAS parallelism is favorable, even at the cost of decreasing the parallelism of the single trials. Due to space constraints, the Figures with lines including standard deviations as uncertainty measurements are now included in Appendix A1. As it took us some time and a large amount of computing resources to run the new experiments, four days were needed for updating the paper. All major additions to the paper are highlighted in red. \n\nDear reviewers,\n\nas the end of the author-reviewer discussion period is approaching, we are reaching out to resolve any additional questions you may have regarding our responses to your comments and the revised paper. We want to ensure that all your concerns are thoroughly addressed before the deadline.\n\nWe thank you for your time and consideration. \nDear Area Chair,\ndear Reviewers,\n\nwe would kindly remind you that the second phase of the discussion period ends this Monday. We are in your service if any additional clarification to our responses is needed.\n\nWe thank you for your time and consideration.\n```\n", "role": "user" }, { "content": "I acknowledge the authors' response. Most of my concerns have been addressed. I stand by my favorable recommendation for this paper.", "role": "assistant" } ]
lmtNt--6dw	AutoML 2023 Conference	Reviewer_oBzf	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_oBzf, and the following are the paper you need to review:\n```<<lmtNt--6dw>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper tackles the problem of parallel compute resource allocation in the context of Neural Architecture Search (NAS) for early stopping-based algorithms. The authors distinguish two assignment choices in the presence of multiple GPUs: 1) assign more resources for speeding up individual configuration training, and 2) assign more resources for NAS procedure for increased parallel configuration sampling. Furthermore, using a unified benchmark, NATS-Bench, they investigate the optimal allocation trade-off on three datasets in a setup of varying numbers of GPUs available (from 1 to 64) of four baselines. Empirical results show that by optimizing the allocation, a speed-up of factor ~2-5 can be achieved. \noverall_review: Strengths:\n- The paper is well-motivated and bears potential practical significance.\n- The evaluation protocol is standardized.\n- The appropriate baselines are chosen for empirical evaluation. \n- The scaling experiment, measuring parallel efficiency is very interesting. \n- The insights about batch size scaling for better GPU utilization are valuable.\n- The future work direction, especially adaptive resource allocation, seems very promising. \n\nWeaknesses:\n- The checkpointing mechanisms (for a single GPU experiment) is not entirely clear, along with its overhead influencing overall runtime.\n- The presentation of some of the figures could be improved to better visualize discussed speed-ups. Maybe displaying initial speed-ups on one plot and the final on another would declutter them.\n- The resource allocation scheme has been optimized for a single search space parametrization, it would be recommended to demonstrate that this carries to different experiment setups and can be reused for different problems.\n- It would be great to see experiments for large models. \n- Clear takeaway messages are missing, to guide researchers in their experimental setup. \nreview_summary: The authors investigate the optimal resource allocation for early stopping-based algorithms. They motivate their research very well and pose interesting research questions. The evaluation protocol has some minor flaws when it comes to clarity and presentation. The empirical results and discussion do not present clear takeaway messages for future research which downplays the potential impact of the paper. As I opt for borderline leaning reject, I encourage to incorporate the feedback to improve the overall paper quality and increase the score. \nactions_required_to_increase_overall_recommendation: - Improve the clarity of the paper.\n- Improve results presentation.\n- Present paper contributions in a more convincing manner.\n- Show that optimized allocation is transferable.\n- Show large-scale experiments.\n- Present takeaway messages more clearly.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Reply to Reviewer oBzf\nWe thank the reviewer for this assessment of our paper and the feedback. We address all concerns raised by the reviewer below. \n\n> In the related work describing HyperBand (L78-79), \"HB waits for all trials to reach a certain threshold in time before applying Successive Halving\" which results in idling workers. Isn't it the case that if there are idling workers, another SH bracket will be spawned? E.g. let's say we have 64 workers, and the first bracket would have 27 configurations evaluated for one epoch, then the remaining 37 workers would not wait for the first 27 to finish, but another bracket with 9 configurations for 3 epochs will start, and so on.\n\nThe reviewer is correct. If there are enough free workers to start the computation of a new bracket, the HB algorithm would do so. However, HB still needs to wait for all trials of a bracket to reach a certain threshold before it can decide on which trials to stop. This is the main difference to the ASHA algorithm, which does so on a rolling basis. We improved the clarity by modifying the statement that HB leads to idling workers to reflect the point raised by the reviewer (see Section 2.1) in the revised manuscript.\n\n> The early stopping-based approaches revolve around stopping and continuing some of the configurations training, do you always train from scratch or checkpoint and load the model? E.g. in case you evaluated a configuration for 1 epoch already when we promote it for 3 epochs, is it trained for 3 (from scratch) or for 2 epochs?\n\nThere is a checkpoint that saves the current state of each training (including model weights) every 10 epochs, whereasdecisions on which trainings to continue and which to stop are also made in the same 10 epoch intervals. Therefore, when a configuration is promoted to the next stage, it would be reloaded from the checkpoint and, hence, is not trained from scratch. \n\n> Follow-up to the above, is this continuation cost reflected on the x-axes in paper figures?\n\nThe runtime plotted in the figures is the whole runtime of the NAS optimization process, so it includes the overhead from the checkpointing and loading of these checkpoints (note that as mentioned above, there is no re-training from scratch). \n\n> The experiments with parallel runs on a single GPU. Does that imply constant switching between configurations training as only one configuration can be trained at once? E.g. in case we have scheduled 9 configurations for 3 epochs, do we take one, train for one epoch, checkpoint, train another configuration for one epoch, checkpoint, and so on? It might be interesting to see what is the overhead of checkpointing and loading models in this scenario.\n\nIt is correct that on a low level, the GPU can just compute a single problem at a time. Therefore, when running parallel trials on a single GPU, “context switching” is used. This means that the GPU constantly switches between different threads, but at a much higher frequency (e.g., after the computation of a single batch of a model) than the employed checkpointing intervals (e.g., every 10 epochs). Unfortunately, information from the hardware providers (such as NVIDIA) on how this is performed under the hood is sparse. But, in our experiments, we measure the whole runtime of the NAS loop (including any overhead created by this context switching) and our conclusions suggest that running more trials in parallel on a single GPU leads to a faster runtime. We therefore assume these overheads to be small. This information, where context swichtig is used in the case of sharing a single GPU is included to Section 3.1 of the revised manuscript. \n\n> It is recommended to include MOBSTER or any multi-fidelity BO method that implements fantasizing the results of pending evaluations.\n\nWe thank the reviewer for this suggestion. However, the Ray Tune framework for the evaluations to avoid implementation-specific disturbance factors used in our work and the MOBSTER framework are unfortunately not compatible out of the box. Hence, we sincerely apologize for not being able to include this suggestion due to the limited time for this rebuttal.\n\n> The motivation is strong and the potential impact on the field is big. However, there are some points that need clarification (see Technical Quality and Correctness Rating). Moreover, the final take-away message is not entirely clear and gives no definitive answer to the posed research questions.\n\nWe thank the reviewer for this positive feedback, and have reformulated Section 5 to directly answer the posed research questions and also clearly formulate the main takeaway message.\n\n> Please consider moving the research questions from the introduction to the experiments section such that a reader does not need to scroll up to orient themself and replace them with clear paper contributions.\n\nAs suggested we have added our main contribution to the introduction of the revised manuscript. \n\n\n\n\n\n\n\n> Please consider working on the presentation of the plots and explain how they were truncated.\n\nThe plots were truncated in the sense that the x-axis was shortened towards the end (as the blue lines run for a long time) and the y-axis was truncated on the bottom. For clarity, we have added the untruncated Figures to Appendix A1 and mentioned the truncation more clearly in the revised manuscript (see Section 4.2). \n\n> The checkpointing mechanisms (for a single GPU experiment) is not entirely clear, along with its overhead influencing overall runtime.\n\nIn our experiments, we measure the whole runtime of the NAS optimization process, therefore, the runtime measurement of the trials on a single GPU includes any overhead. It is advisable from our conclusions to increase the parallelism of the NAS loop. \n\n> The presentation of some of the figures could be improved to better visualize discussed speed-ups. Maybe displaying initial speed-ups on one plot and the final on another would declutter them.\n\nWe thank the reviewer for suggesting to declutter the speed-up plots to help with readability. As requested, we have split the Figures into initial and final speed-ups and included them in the Appendix B of the revised manuscript, as space in the main paper is sparse. \n\n> The resource allocation scheme has been optimized for a single search space parametrization, it would be recommended to demonstrate that this carries to different experiment setups and can be reused for different problems.\n\nWe agree with the reviewer that demonstrating the resource allocations schemes in more experimental setups would be helpful to ensure the generalizability of the findings, and therefore, added an evaluation on a ResNet search space with the (full size) imagenet-1k dataset (see Section 4.4 and Figure 6). Additionally, we include evaluation on more HPC systems, equipped with NVIDIA V100 and AMD MI50 GPUs (see Section 4.5). \n\n> It would be great to see experiments for large models.\n\nAs mentioned in our previous answer, we have added a study on a ResNet search space. The models sampled from this search space are roughly two orders of magnitude larger than the models from NATS-Bench (e.g., 30M vs 0.3M parameters, respectively). \n\n> Clear takeaway messages are missing, to guide researchers in their experimental setup.\n\nWe have modified Section 5 to answer the research questions directly and state the main takeaway message (to increase the parallelism of the NAS loop at the cost of parallelism of the single trials) more clearly.\n\n> Improve the clarity of the paper.\nImprove results presentation.\nPresent paper contributions in a more convincing manner.\nShow that optimized allocation is transferable.\nShow large-scale experiments.\nPresent takeaway messages more clearly.\n\nWe believe to have addressed all the points raised in our previous answers and are happy to answer any follow-up questions. \n\n\n\n\n\n\n\n\n Global/general Response:\n```\nWe would like to thank all reviewers for their valuable feedback and suggestions to improve the paper. We have significantly extended our experimental evaluations, now including runs on two more types of GPUs (NVIDIA V100 and AMD MI50). Moreover, another NAS search space, where sampled models are trained on the complete imagenet dataset (see Subsections 4.4 and 4.5 and Figure 6), is included to the revised manuscript. These evaluations are inline with the previous conclusions and main takeaway, which has been that an increase of the NAS parallelism is favorable, even at the cost of decreasing the parallelism of the single trials. Due to space constraints, the Figures with lines including standard deviations as uncertainty measurements are now included in Appendix A1. As it took us some time and a large amount of computing resources to run the new experiments, four days were needed for updating the paper. All major additions to the paper are highlighted in red. \n\nDear reviewers,\n\nas the end of the author-reviewer discussion period is approaching, we are reaching out to resolve any additional questions you may have regarding our responses to your comments and the revised paper. We want to ensure that all your concerns are thoroughly addressed before the deadline.\n\nWe thank you for your time and consideration. \nDear Area Chair,\ndear Reviewers,\n\nwe would kindly remind you that the second phase of the discussion period ends this Monday. We are in your service if any additional clarification to our responses is needed.\n\nWe thank you for your time and consideration.\n```\n", "role": "user" }, { "content": "Title: Increased score\nThank you for the clarifications and for the additional content!", "role": "assistant" } ]
lmtNt--6dw	AutoML 2023 Conference	Reviewer_XDAv	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_XDAv, and the following are the paper you need to review:\n```<<lmtNt--6dw>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper studies the impact of using variable numbers of workers and GPUs on the speed of NAS within NATS-Benchmark. More specifically it studies two main research questions: 1) Efficiency of NAS as a function of number of GPUs used, and 2) Impact on the speed and accuracy of NAS when using various numbers of trials per GPU. From the results: 1) Efficiency has a decreasing trend when using more GPUs compared to using a single GPU. 2) NAS is generally faster when using more workers on a GPU, but if there are too many workers, then the NAS time can start increasing. There is no significant impact on accuracy. The authors study the question further by controlling for the amount of simultaneously processed data and find that the main reason hindering efficient resource usage are the switching and reloading processes between trials.\noverall_review: Positives\n* The paper studies the problem of resource allocation from multiple perspectives: 1) scaling across multiple GPUs, 2) scaling workers within 1 GPU\n* The paper is relatively easy to read\n* The topic of the paper may be interesting in case there are any surprising findings that would challenge how resources are allocated\n* Multiple common HPO algorithms (and hence of interest to researchers) are evaluated\n\nNegatives:\n* Any specific recommendations in the paper are likely to depend on the exact combination of benchmark and GPU used. Because only NATS-Benchmark and one type of GPU was evaluated it is hard to use any of the more specific suggestions such as using 32 GPUs in parallel and not more. To make the specific recommendations more general (and hence useful) more settings/combinations should be considered, but it can give mixed results.\n* A large part of the take-aways about efficiency in RQ1 depends on using threshold of 0.75. Various people may prefer different thresholds and the results can be quite different when considering different setups, so claims that we should use up to 32 can be unjustified. The results generally show that efficiency decreases as more GPUs are used.\n* While in some cases average across 3 runs is reported, no results show standard deviations. These would be especially useful in Table 1 to see if the points raised in the discussion actually hold or are simply due to noise.\n* The high-level findings of the paper are not particularly surprising as they essentially suggest that “it is advisable to increase the parallelism of the NAS loop to achieve a decreased total runtime”.\n\nreview_summary: Based on the paper it is hard to take specific and justified insights about how many workers or GPUs one should use when facing a new setup. The current paper is more of a case-study of the impact of using different numbers of workers on the specific setup considered.\n\nactions_required_to_increase_overall_recommendation: It would be useful to conduct evaluation using more settings (especially various types of GPU but also other benchmarks would be useful) so that the guidance is more likely to generalize to new settings.\nRating: 5 marginally below the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Reply to Reviewer XDAv\nWe thank the reviewer for the assessment of our work and the provided feedback. We address all concerns raised by the reviewer below. \n\n> The paper can be seen as a case study showing the impact of allocating resources in various ways on NATS-Benchmark when using a specific GPU, with the high-level findings likely to generalize. However, the general findings agree with earlier findings (existing small-scale studies mentioned in section 2.4) so in this sense the paper does not bring particularly surprising findings. As a result, I expect only limited impact.\n\nWe thank the reviewer for this comment. We are aware that small-scale studies on the scalability and resource allocations of NAS methods have been performed before, however, these are different from our evaluations in the sense that the type of hardware that is used (e.g. just CPUs), the number of workers (GPUs) are different and the models trained are much smaller. These types of evaluations do not transfer to experimental setups currently used by researchers, which is where our study provides guidance. We have added a sentence to Section 3.2 for clarification and hope that the reviewer now agrees with the impact of this work. \n\n> The evaluation considers a rather limited setup of NATS-Benchmark on a specific type of GPU so it is hard to draw specific conclusions that would tell us more about how to allocate resources. In this sense, the claims in the paper are not sufficiently sound as the level of empirical evaluation does not enable us to make the more general claims in the paper. More specifically one would expect that the identified numbers in the paper strongly depend on what GPU is used - most GPUs (e.g. 2080 Ti or 1080 Ti) would likely have those thresholds lower than A100). \n\nWe agree with the reviewer that performing experiments on different types of GPUs, in particular less powerful GPUs than A100s would bolster our findings and thank him for this suggestion. We have therefore decided to perform additional experiments on clusters featuring NVIDIA V100 and AMD MI50 GPUs. We chose those types of GPUs as they are more prevalent in current HPC and cluster systems than, e.g., commercial NVIDIA Ti GPUs. The results are presented in the new Subsection 4.5 and Figure 6 of the revised paper. The results of these experiments support our findings that increasing the parallelism of the NAS loop while decreasing the resources of the single trials leads to lower runtimes. \n\n> On the other hand the general findings are relatively obvious and may have been identified in earlier papers (although without rigorous studies - based on section 2.4 of this paper). \n\nWe would like to emphasize the main finding of our work that increasing the parallelism of the NAS loop (which at the same time means decreasing the parallelism of the single trials, as the total number of workers in a setting stays the same) reduces the total runtime of the NAS loop. \n\n> Some evaluations are averaged over 3 seeds, but e.g. Table 1 would benefit from including standard deviations to see if we can be confident about the behavior we see there.\n\nWe agree that showing standard deviations is important for quantifying the reliability of the results and thank the reviewer for pointing this out. As the lines in the plots are close together (see e.g. Fig. 3), we have decided for omitting them in the plots in the main paper. However, we do now include the untruncated figures with the standard deviations as uncertainty measurement in Appendix A1 of the revised manuscript, whereas Table 1 of the revision now includes standard deviations. \n\n> The paper is generally clear and it is simple to read. There are quite a few typos and simple grammar mistakes, so these should be fixed. Examples include:\nL119 training is is especially\nL151 nestorv momentum for\nL194 allocated to a single trials\nL211 these kinds of decision can\n\nWe have fixed these typos in the revised paper. \n\n> Any specific recommendations in the paper are likely to depend on the exact combination of benchmark and GPU used. Because only NATS-Benchmark and one type of GPU was evaluated it is hard to use any of the more specific suggestions such as using 32 GPUs in parallel and not more. To make the specific recommendations more general (and hence useful) more settings/combinations should be considered, but it can give mixed results.\n\nWe agree with the reviewer that to make more general recommendations it is important to perform experiments in different settings. We have therefore added two more types of experiments: one using older NVIDIA V100 GPUs and AMD MI 50 GPUs to assess the impact of the type of hardware and one using larger models from a ResNet50 search space. It should be noted that the results of the new evaluations support our previous conclusions (see Subsections 4.4 and 4.5 and Figure 6). \n\n\n\n\n\n\n> A large part of the take-aways about efficiency in RQ1 depends on using threshold of 0.75. Various people may prefer different thresholds and the results can be quite different when considering different setups, so claims that we should use up to 32 can be unjustified. The results generally show that efficiency decreases as more GPUs are used.\n\nAs the reviewer points out correctly, different thresholds might be suitable for different scenarios. We have therefore added a sentence to Section 5 to emphasize that users that are comfortable with a parallel efficiency of 0.6 can use up to 64 GPUs, but more computational resources will be spent on communication and scheduling instead of computation in that case. \n\n> While in some cases average across 3 runs is reported, no results show standard deviations. These would be especially useful in Table 1 to see if the points raised in the discussion actually hold or are simply due to noise.\n\nThe standard deviations are now reported in Table 1. As they are small, they do not influence the final results. \n\n> The high-level findings of the paper are not particularly surprising as they essentially suggest that “it is advisable to increase the parallelism of the NAS loop to achieve a decreased total runtime”.\n\nSee our previous answer to the first question. \n\n> It would be useful to conduct evaluation using more settings (especially various types of GPU but also other benchmarks would be useful) so that the guidance is more likely to generalize to new settings.\n\nIn addition to the new Subsection 4.5 with experiments on various types of GPUs, we have also added new experiments on a different NAS search space; a custom ResNet50 model trained on the complete imagenet-1k dataset(see Subsection 4.4 and Figure 6). The results of these new experiments confirm that our main findings do not depend on the type of hardware used.\n\nWe hope to have addressed all concerns raised by the reviewer and would be happy to answer any follow-up questions. \n\n\n\n\n Global/general Response:\n```\nWe would like to thank all reviewers for their valuable feedback and suggestions to improve the paper. We have significantly extended our experimental evaluations, now including runs on two more types of GPUs (NVIDIA V100 and AMD MI50). Moreover, another NAS search space, where sampled models are trained on the complete imagenet dataset (see Subsections 4.4 and 4.5 and Figure 6), is included to the revised manuscript. These evaluations are inline with the previous conclusions and main takeaway, which has been that an increase of the NAS parallelism is favorable, even at the cost of decreasing the parallelism of the single trials. Due to space constraints, the Figures with lines including standard deviations as uncertainty measurements are now included in Appendix A1. As it took us some time and a large amount of computing resources to run the new experiments, four days were needed for updating the paper. All major additions to the paper are highlighted in red. \n\nDear reviewers,\n\nas the end of the author-reviewer discussion period is approaching, we are reaching out to resolve any additional questions you may have regarding our responses to your comments and the revised paper. We want to ensure that all your concerns are thoroughly addressed before the deadline.\n\nWe thank you for your time and consideration. \nDear Area Chair,\ndear Reviewers,\n\nwe would kindly remind you that the second phase of the discussion period ends this Monday. We are in your service if any additional clarification to our responses is needed.\n\nWe thank you for your time and consideration.\n```\n", "role": "user" }, { "content": "Title: Response\nThanks for the additional explanations and for conducting the additional experiments using various GPUs. My main worry is that these have not been used for also answering RQ1 where specific guidance about efficiency is given (e.g. use up to 32 GPUs if you want efficiency >0.75). One could expect that the answer to RQ1 depends on what GPU you use, which could lead to challenges with what message the paper gives.\n\nThe additional clarifications have been helpful, especially highlighting that the overall number of workers remains the same in RQ2. I see some value in studying this as people may not be fully utilising GPUs (due to running one trial on one GPU). But on the other hand if a GPU is well-utilised with a single trial on it (e.g. by using the max batch size that fits there for speed - which can be very common to do), then it may not be useful to put multiple trials there. So a limitation of this study is that the findings likely only apply to problems that do not fully utilise GPU, making the findings problem-dependent. So overall the guidance about speedups can apply only to certain (limited) cases.\n\nOverall I increase my rating by one level only (borderline leaning reject), due to the limitations outlined here.", "role": "assistant" } ]
tQz8u2KU3zy	AutoML 2023 Conference	Reviewer_jnWR	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_jnWR, and the following are the paper you need to review:\n```<<tQz8u2KU3zy>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper studies how to perform HPO for domain adaptation algorithms where the labeled validation set cannot be accessed and provides an empirical study of HPO and model selection for domain adaptation problems.\noverall_review: 1. This paper provides an empirical study of HPO and model selection for domain adaptation problems and shows some useful findings for domain adaptation. But this paper does not have much potential impact on the AutoML field.\n2. It seems the HPO method used in this paper does not been introduced. Is it Eq. (4)? Why not use Bayesian HPO methods?\n3. A typo in Line 563.\n4. Different validators in Tab. 2 have different time complexity. It is better to consider the time complexity when performing HPO for domain adaptation problems.\nreview_summary: I think this paper provides some useful empirical findings for domain adaptation but limited impact on the AutoML field. I lean to accept this paper but I don't object to reject it either.\nactions_required_to_increase_overall_recommendation: If the authors can address my concerns in the \"Overall Review\" part, I am willing to raise my score.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Clarifications on HPO and time complexity of validators\nThank you for your feedback.\n### On HPO:\nWe use random search (Detailed in Sec C.2).\nBetter performance can potentially be reached by using e.g. BayesOpt like you mention. In this work, we focus on analysis and prefer the simpler random search to (1) enable computing a correlation score between the validation criteria and test performance (correlation computed over both high and low-quality checkpoints), as we report in our main tables. Also because (2) our comparisons involve comparing the “best” possible checkpoint with the one discovered by each validator, we did not want to risk aggressively optimising a bad validator and thus having no good checkpoints available for selection by the oracle.\n\nThis clarification has now been added to Sec C.2.\n\n### On validator runtime:\nWe now include figures in the appendix Sec D.2, measuring clock time of validators on two datasets of different sizes. This shows how the clustering-based validators are more compute-intensive but not prohibitively so compared to the cost of adaptation.\n\n### On relevance to AutoML:\nPlease see [general comment](https://openreview.net/forum?id=tQz8u2KU3zy&noteId=7ZMTnv1Y38).\n Global/general Response:\n```\nWe were surprised the reviewers thought the relevance was unclear. In order to automate anything (algorithm, hyperparameter, checkpoint selection), we need a metric to optimise. In the case of supervised learning, everyone agrees on this metric (validation performance), but it’s not straightforward for UDA/SFDA/TTA due to the lack of labels. The whole point of this paper is to clarify what the AutoML optimisation metric should be for domain adaptation. Once this is clarified, all the rest of AutoML tooling (BayesOpt, etc) can be directly applied. Thus this paper is about laying the foundations that allow bringing AutoML to DA.\n\nTowards this, we have identified some problem settings and metrics where the validators work sufficiently well such that AutoML might work already today (UDA); and some others where the validation metrics are insufficient, meaning AutoML can’t work until better validators are developed (TTA on big domain shift).\n\nWe are happy to update our text to better motivate this contribution.\n```\n", "role": "user" }, { "content": "I acknowledge the authors' response. The reply answers almost all of my questions. I reaffirm my positive endorsement for this paper.", "role": "assistant" } ]
tQz8u2KU3zy	AutoML 2023 Conference	Reviewer_UUAd	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_UUAd, and the following are the paper you need to review:\n```<<tQz8u2KU3zy>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper introduces a benchmark for domain adaptation techniques across unsupervised domain adaptation, source-free domain adaptation, and test-time adaptation that focuses on ensuring that the evaluation methods are aligned with practical application by avoiding using test data labels for tuning / model selection when in practice these labels would not be available. Within this context, the paper evaluates a variety of validators and adaptation algorithms across various datasets and source-model hyperparameters to identify the combinations that perform well in practice, showing that while strong results can potentially be obtained via UDA, that results for SFDA and TTA are mixed, with high volatility and many algorithms underperforming the no-adaptation baseline. The paper additionally highlights that using a separate target validation set for selection / tuning rather than purely the target train set results in better performance in the UDA setting.\noverall_review: The paper provides a useful empirical analysis of DA techniques and reports important negative results that should be taken into careful consideration when thinking about integrating DA into AutoML systems. The paper additionally highlights that using a separate target validation set for selection / tuning rather than purely the target train set results in better performance in the UDA setting.\n\nA lack of multiple random seeds does hinder the analysis, as well as there being little analysis on any \"auto\" component to assist AutoML in applying DA practically. In particular, I am still confused on the train/val target set details, which I hope is clarified during the rebuttal. Overall, I am borderline leaning accept, and will consider updating my score once my key questions to the authors are addressed.\nreview_summary: Overall, I am borderline leaning accept, and will consider updating my score once my key questions to the authors are addressed.\nactions_required_to_increase_overall_recommendation: 1. Most importantly I'd like to understand the train/val target set reasoning (Is it reasonable to assume these would exist separately from the test target set? What if they don't?). What are the downsides of splitting the target set into train/val?\n\n2. Additional metrics or indicators of stability of the approaches would be useful when evaluating if a DA technique is worth the risk of a worst-case outcome. I listed some example ideas earlier in my review, but I am curious what the authors think would be meaningful metrics in this regard, given the apparent volatility of DA algorithm performance.\n\n3. What, if anything, is \"auto\" in this work? Does AutoML have any application in this work beyond referencing the results tables and choosing a validator/algorithm pair to apply for DA after AutoML training?\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Reasoning behind val splits and measures of stability\nThank you for bringing up a few important points for us to clarify.\n\n### On train/val/test splits:\nYou bring up a good point. There are cases where we would not recommend using a val split.\n\nWe can consider two settings, a transductive case where we wish to evaluate our method on all target data given to us or an inductive case where we assume more target data is coming, which makes generalisation more important.\n\nIn the transductive case, you may prefer not to use a validation split. We would argue, however, that the inductive case is more realistic, as unlabelled data is often widely accessible in the modern age and organisations often receive continuous streams of data without any clear cut-off points. Due to the importance of generalisation in this setting, using a val split becomes more important.\n\nAdditionally, the usefulness of a val split depends on the amount of target data available. We agree that for a bounded-size benchmark, making a train/val split out of the actual train set costs some data, which will give a performance penalty if the benchmark is small. On the other hand, we show that the quality of the validator improves with a val split. So we end up with a trade-off between adaptation and validation quality. At a certain point, the gains from using more data for adaptation diminishes and become more useful for validation.\n\nOverall, we would recommend a val split if we are in the inductive setting and the dataset size is not small.\n\n### On the stability of algorithms and validators:\nWe now include tables reporting the \"percentage of results that are equal or better than no-adaptation source-only\" for both algorithms (independent of validator) and validators (independent of algorithms). These can be found in Tables 7 & 8 in the revised paper. These will help give an idea of the stability of an algorithm or validator.\n\n### On relevance to AutoML:\nPlease see [general comment](https://openreview.net/forum?id=tQz8u2KU3zy&noteId=7ZMTnv1Y38).\n Global/general Response:\n```\nWe were surprised the reviewers thought the relevance was unclear. In order to automate anything (algorithm, hyperparameter, checkpoint selection), we need a metric to optimise. In the case of supervised learning, everyone agrees on this metric (validation performance), but it’s not straightforward for UDA/SFDA/TTA due to the lack of labels. The whole point of this paper is to clarify what the AutoML optimisation metric should be for domain adaptation. Once this is clarified, all the rest of AutoML tooling (BayesOpt, etc) can be directly applied. Thus this paper is about laying the foundations that allow bringing AutoML to DA.\n\nTowards this, we have identified some problem settings and metrics where the validators work sufficiently well such that AutoML might work already today (UDA); and some others where the validation metrics are insufficient, meaning AutoML can’t work until better validators are developed (TTA on big domain shift).\n\nWe are happy to update our text to better motivate this contribution.\n```\n", "role": "user" }, { "content": "Title: Response to Authors\nThank you for your response and updates! Regarding adjustments to my score, the primary remaining point that gives me hesitance is the relevance to AutoML, which I reply with more detail below.\n\n## On train/val/test splits:\n\nThank you for your answer which generally covers my questions. Below I provide some additional thoughts:\n\nI would note that from the AutoML perspective, it is likely that users would often not initially try the systems in the inductive setting for specialized scenarios, as the inductive setting is more complex than the transductive setting and the user might not understand how to properly separate train and test data. This confusion can be avoided by implementing cross-validation in order to leverage all data as both train (+val) and test, which removes the need for the user to understand this aspect, and enables the technique to be applied in all scenarios regardless if there is additional data available beyond the test data.\n\nFuture work investigating the benefits of cross-validation techniques to more efficiently use the available data for both inductive and transductive settings would be very interesting and significantly raise the relevance and applicability to AutoML, but I do not expect this investigation to be part of this work, as it requires re-running all experiments.\n\n## On the stability of algorithms and validators:\n\nThank you for adding these tables! One comment I have is that while these numbers are useful, they are not representative of the true per-task volatility of the algorithms/validators. Instead, they are based on \"Averages over all 21 domain\ntransfers evaluated\". It would be more useful to have the percentages in the table be based on the success rate on 21*6 tasks, rather than 6 averaged tasks, as in practice, the user will care about the success rate on a single task and not on the average of many tasks.\n\nAs an example, if an algorithm has a 55% success rate per-task, when averaging over 21 tasks this would likely look like a 100% success rate. The 55% success rate value is more relevant than the 100% success rate value, and should be the value reported.\n\nAlso, is there a reason success rate is evaluated separately for algorithm and validator? Wouldn't it be reasonable to evaluate success rate on all algorithm+validator pairings? This should then result in a table of identical dimensions to Table 3, but with success rate across 21 tasks for each algorithm+validator pair instead of average accuracy across 21 tasks.\n\n## On relevance to AutoML:\n\nWhile I agree with the authors that the goal they describe is relevant to AutoML, there is still lacking a consistent messaging as to the recommendation to the AutoML community based on the results.\n\nMy personal take-away is to currently avoid DA in AutoML except for UDA. The choice of optimal general purpose algorithm + validator to use remains unclear, however the paper does a reasonable job of showcasing techniques that should be avoided (Ex: CORAL, SND, etc.). For UDA, the best practice would involve a deeper dive than what was shown in the paper to identify a strategy that minimizes worst-case scenario results while often generating strong results.\n\nWhile this is my personal take-away as someone heavily involved in the AutoML field who would be a direct user of such DA techniques from a system integration perspective, it is an implied take-away and not one that the authors explicitly discuss nor recommend in their conclusion. If the authors proposed some approach-acceptance mechanism to say that a given technique is superior to Source-only-accuracy with some level of statistical certainty, and then recommend a mechanism in the conclusion or analysis sections, this would increase the impact and relevance to AutoML.\n\nAs a first step, I would recommend the authors filter out obviously bad methods via some criterion to only focus on the performant solutions (for example, filter out methods that are statistically significantly worse than source-only-accuracy), and then conduct more detailed comparisons & statistical tests to come up with a single recommendation to the AutoML community that is actionable and implementable in modern AutoML systems as a best practice.", "role": "assistant" }, { "content": "Title: Making actionable recommendations for the community\nWe thank the reviewer for their helpful feedback from a perspective within the AutoML community.\n\nIn the new tables, the algorithm success rates are in fact reported in the more fine-grained style you recommend. But the validator success rates are indeed not as helpful as they could be. We will update this.\n\nWe especially appreciate the points about how to provide more direct recommendations for the community, such as performing statistical tests to single out an algorithm/validator pairing for realising AutoML for DA right now. An excellent suggestion we will dig into. This should reveal when such applications are beneficial (likely UDA) and when they are likely to fail (TTA).", "role": "user" }, { "content": "I am grateful for the opportunity to discuss these points with the author, which has significantly improved the quality of this work.", "role": "assistant" } ]
tQz8u2KU3zy	AutoML 2023 Conference	Reviewer_1Bfh	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_1Bfh, and the following are the paper you need to review:\n```<<tQz8u2KU3zy>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: In this paper, the authors analyze the general field of Domain Adaptation by benchmarking a suite of candidate validation criteria and then use these to compare popular adaptation algorithms for the purpose of hyper-parameter optimization and model selection. They consider different branches of domain adaption, Unsupervised Domain Adaptation (UDA), Source free domain adaption (SFDA), and Test time augmentation (TTA) because there seems to be a lack of consistency amongst the validation criteria used to compare different models. In the analysis of methods, covering 15 validators each with 10 hyper-parameter choices, they find that previously unexplored validators (they bring up RankMe and V-Measure) consistently perform well but generally, the best validator is setting/algorithm dependent. \noverall_review: Pros\n+ The author choose a very broad and thorough sweep of methods within UDA, SFDA, and TTA as well as a large sweep of validators.\n+ The questions both asked and the results answered in the analysis section do a good job of answering the high-level question in the paper regarding the evaluation of the UDA methods.\n+ The final investigation of TTA seems very strong in particular. From the analysis surrounding the statement \"we suggest that the strong success of TTA methods on synthetic benchmarks may not be representative of real-world adaptation problems\", this would be a major contribution that changes how future TTA methods are developed.\n\nCons\n- How to read the correlations is confusing. It seems to me like the correlation of even the final methods (RankMe and V-Measure) the authors propose doesn't seem to be that high. I was wondering where the authors drew the conclusions that these were good validators from.\n- Not a con necessarily but something that I felt might be incomplete was the analysis of tuning sensitivity. Tuning sensitivity was mentioned in the results section 4.1.2 but then was not addressed in 4.2.1 or 4.3.1. How does tuning sensitivity compare to the TTA or SFDA regimes?\nreview_summary: Based on my points above, I think that this is a very solid piece of work that merits acceptance to the conference. Their effort to make a common ground for evaluation and thorough investigation of different methods and validators is definitely accepted paper quality work. I believe as the authors point out that they could potentially sample more hyper-parameter settings for each algorithm because, with the minor improvements incurred, might choose better checkpoints.\nactions_required_to_increase_overall_recommendation: I am fairly confident that Accept is the right evaluation for this piece of work. I have a few questions that I listed in the cons but overall think that they would not cause me to raise my score unless the authors were able to perform an even larger hyper-parameter suite to make their results even stronger and maybe tighten up the writing to make the paper less dense to read.\nRating: 8 accept, good paper\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Correlations, tuning sensitivity and more\nThank you for your feedback.\n\n### On correlations:\n(1) Yes, it is true that most validators have low correlations. This is to some extent an open problem, which is one of our take home messages. We also consider correlation not to be a great measure of use as a metric to drive HPO, since most validators seem to exhibit non-linear relationships with test performance. We instead think it better to look at final selection performance, which is what we focus most discussion on. We report which the best validators are for each setting, but agree that often they are not strong enough to recommend for practical use (in e.g. TTA).\n\n### On tuning sensitivity:\nIn order to get a better picture of tuning sensitivity, we now include tables reporting the \"percentage of results that are equal or better than no-adaptation source-only\" for both algorithms (independent of validator) and validators (independent of algorithms). These can be found in Tables 7 & 8 in the revised paper.\n\n### On data split sizes:\nWe split all domains of all datasets into train (60%), val (20%) and test (20%) splits. This has been added to Sec C.1.\n\n### On rotating figures:\nWe will edit these figures to make sure they are large enough to read without rotating.\n Global/general Response:\n```\nWe were surprised the reviewers thought the relevance was unclear. In order to automate anything (algorithm, hyperparameter, checkpoint selection), we need a metric to optimise. In the case of supervised learning, everyone agrees on this metric (validation performance), but it’s not straightforward for UDA/SFDA/TTA due to the lack of labels. The whole point of this paper is to clarify what the AutoML optimisation metric should be for domain adaptation. Once this is clarified, all the rest of AutoML tooling (BayesOpt, etc) can be directly applied. Thus this paper is about laying the foundations that allow bringing AutoML to DA.\n\nTowards this, we have identified some problem settings and metrics where the validators work sufficiently well such that AutoML might work already today (UDA); and some others where the validation metrics are insufficient, meaning AutoML can’t work until better validators are developed (TTA on big domain shift).\n\nWe are happy to update our text to better motivate this contribution.\n```\n", "role": "user" }, { "content": "Title: Thank you for your response.\nThank you for your response. This satisfies the questions I had and I think these additions made the already good paper stronger.", "role": "assistant" } ]
tQz8u2KU3zy	AutoML 2023 Conference	Reviewer_Xd7a	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_Xd7a, and the following are the paper you need to review:\n```<<tQz8u2KU3zy>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper identifies the problem that many unsupervised domain adaptation (UDA) works use unrealistic validation methods especially for AutoML, so they propose methods for and perform benchmarks using only validation data that is available in each UDA scenario of increasing constraints. \noverall_review: Positive aspects:\n\n* The problem is identified and motivated very clearly.\n\n* A good breadth of datasets (including both simulated and natural domain shifts), data availability constraints, and backbones are studied.\n\nNegative aspects:\n\n* While some HPO is performed within each experiment, there is no analysis of how the training recipe (such as the number of source-only checkpoints, number of target hyperparameter sets, or the optimizer and it's fixed hyperparameters) affects the variance of results.\n\n* Could these studies be used to automatically select the best validators (and subsequently the best algorithms) for a given UDA scenario? This could just be proposed within this paper to help improve the impact.\n\n* The title is a bit too broad. Perhaps something like \"Better Practices for Validation of Unsupervised Domain Adaptation\" would be better.\n\n* Can you further analyze the \"+SO\" results? Why did you decide to include the source-only model, and when does it affect the results? How realistic is this setting? Like Table 6, can you add results for the algorithms without the source-only model?\n\n* Why are certain rows/columns mostly blank in Tables 4-6? \n\n* What happens outside of image classification? Are there other relevant tasks and data modalities that could be benchmarked to see if your claims extend beyond image classification? In Section 6, you mention a \"preliminary study of regression adaption\" but don't seem to actually include these results.\n\nMinor comments:\n\n* Within-paragraph numbering is inconsistent: sometimes numbers (1),(2),(3) are used, other times numerals (i), (ii), (iii) are used.\n\n* In line 24, you state \"a different conclusion\", but some of your conclusions are similar. \n\n* Explicitly mention how to find your code in the main paper, in case the reader does not see it in the checklist.\n\nreview_summary: This paper is a clear and impactful study on valid validation of UDA techniques. It could be further supported by additional analysis detailed above. I think it is worth accepting, especially if the recommended actions are taken.\nactions_required_to_increase_overall_recommendation: * Clarify and further analyze your tables of results\n* Explicitly propose how to practically apply the findings of this paper\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Training recipe and clarifications on tables\nThank you for your comments.\n\n### On training recipe:\nWith more checkpoints etc, we would expect that the oracle performance level and the best validators would improve, but the worse ones would not. So stretching of the results range. But we don’t expect any fundamental conclusion change.\nWe have prioritised a large number of datasets over increasing the number of hyperparameter sets etc.\n\nThank you for your comments.\n\n### On automatically selecting the best validator:\nNo, this would not be possible. We can’t automatically pick a validator, because this requires test performance. But our empirical results can be used to indicate a good validator overall (since we average over several datasets, the validator shown to be good is likely to generalise to new ones).\n\n### On the title:\nThis is a good suggestion, and we will consider updating the title.\n\n### On the +SO results:\nIf you exclude the source-only checkpoint, sometimes performance is much lower than the baseline. We decided to include it to test the hypothesis that the validator could pick the source-only model as a backup option. We included it for Tab 6 specifically, because this is the one where the performance is worse than baseline. It didn’t make much difference in the other tables. \n\n### On blank rows/columns in tables:\nRow: We only have a single source-only model available in the SFDA/TTA settings, so there it doesn’t make sense to apply validators to a pool. So you can consider the printed value to apply to the whole row.\nColumn: Source validation accuracy requires access to source data. We don’t have this for TTA/SFDA.\n\n### On other modalities:\nThe regression results we discuss are included in the appendix Sec. A. We could benchmark other modalities, but this is out of scope of our current study.\n Global/general Response:\n```\nWe were surprised the reviewers thought the relevance was unclear. In order to automate anything (algorithm, hyperparameter, checkpoint selection), we need a metric to optimise. In the case of supervised learning, everyone agrees on this metric (validation performance), but it’s not straightforward for UDA/SFDA/TTA due to the lack of labels. The whole point of this paper is to clarify what the AutoML optimisation metric should be for domain adaptation. Once this is clarified, all the rest of AutoML tooling (BayesOpt, etc) can be directly applied. Thus this paper is about laying the foundations that allow bringing AutoML to DA.\n\nTowards this, we have identified some problem settings and metrics where the validators work sufficiently well such that AutoML might work already today (UDA); and some others where the validation metrics are insufficient, meaning AutoML can’t work until better validators are developed (TTA on big domain shift).\n\nWe are happy to update our text to better motivate this contribution.\n```\n", "role": "user" }, { "content": "Title: Response to the authors\nI appreciate your responses to my concerns. However, few changes were affected to the article itself, so I will maintain my rating.", "role": "assistant" } ]
tQz8u2KU3zy	AutoML 2023 Conference	Reviewer_2u7o	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_2u7o, and the following are the paper you need to review:\n```<<tQz8u2KU3zy>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper examines the current state of domain adaptation using approapriate evaluation methods by testing various validation criteria and applying them to assess popular adaptation algorithms. It shows that all three branches of domain adaptation consistently face significant challenges. Although attainable performance are generally lower than expected, using appropriate validation splits and metrics is beneficial. \n\noverall_review: Pros:\n\nThe authors clearly identify bad practices for HPO in conjunction with DA and motivate the paper in an appropriate way accordingly.\nThe experimental setup covers a large spectrum for DA scenarios and considers a large variety of validators.\nThe outcomes of different setups show some general trend despite some inconsistencies which makes them valuable take-home messages. \n\nCons:\n\nThe paper barely fits into the scope of an AutoML conference since hyperparameters are fixed in their respective setups. The promise that the introduced data split combined with an appropriate validator could help reseach on HPO under DA conditions was not sufficiently supported (although it sounds plausible) \n\n\nreview_summary: In general, this is a good paper. The fact that it is a bit off topic and has little to do with AutoML makes it potentially not suitable for this venue. I am willing to upgrade my voting if the author manage to show a stronger connection to the conference topic.\nactions_required_to_increase_overall_recommendation: Given the empirical aspect of the paper, it is hard to substantially improve the quality of the paper.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Relevance to AutoML\nThank you for your comments.\nPlease see the [general comment](https://openreview.net/forum?id=tQz8u2KU3zy&noteId=7ZMTnv1Y38) and the updated introduction in the revised paper, both clarifying how our study relates to the AutoML community.\n Global/general Response:\n```\nWe were surprised the reviewers thought the relevance was unclear. In order to automate anything (algorithm, hyperparameter, checkpoint selection), we need a metric to optimise. In the case of supervised learning, everyone agrees on this metric (validation performance), but it’s not straightforward for UDA/SFDA/TTA due to the lack of labels. The whole point of this paper is to clarify what the AutoML optimisation metric should be for domain adaptation. Once this is clarified, all the rest of AutoML tooling (BayesOpt, etc) can be directly applied. Thus this paper is about laying the foundations that allow bringing AutoML to DA.\n\nTowards this, we have identified some problem settings and metrics where the validators work sufficiently well such that AutoML might work already today (UDA); and some others where the validation metrics are insufficient, meaning AutoML can’t work until better validators are developed (TTA on big domain shift).\n\nWe are happy to update our text to better motivate this contribution.\n```\n", "role": "user" }, { "content": "Thank you for the detailed response from the authors. The response covers nearly all of the issues I raised. I maintain my positive recommendation for this work.", "role": "assistant" } ]
y7rbX884LZrX	AutoML 2023 Conference	Reviewer_7dfr	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_7dfr, and the following are the paper you need to review:\n```<<y7rbX884LZrX>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The authors provide a deeper understanding of various zero-cost proxies by interpreting them based on a pruning perspective, and they propose a new combination of zero-cost proxies which achieve state-of-the-art performance. More specifically, they introduce a new gradient-based metric called SSNR and use layer-wise naswot as an activation-based metric to propose a combined proxy called T-CET. And the superiority of the proposed proxy is confirmed through various experiments.\noverall_review: The strengths of this paper lie in the analysis of various zero-cost proxies in the pruning literature, the proposal of a new proxy that combines existing proxies, and the demonstration of high performance. The theoretical and experimental analysis of the paper is expected to be of great help to future research in the zero-cost NAS field. However, the lack of detailed descriptions of the experimental environment and the brief analysis of the experimental results serve as drawbacks.\nreview_summary: This paper is technically correct and the proposed method consistently beat comparing methods on various search spaces and datasets. Also, the authors provide a novel view to analyze zero-cost proxies. Therefore, my recommendation is a strong accept.\nactions_required_to_increase_overall_recommendation: It would be better if more information about the experimental setup could be added. Additionally, although it seems to be due to the page limit, the size of Table 2 is too small to be easily read on the typical A4 size.\nRating: 9 strong accept, excellent work\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Thank you for your support of our work\nWe thank the reviewer for the positive feedback, and appreciate the time and effort that the reviewer has shown to help us improve our work. We have further updated our paper and hope to have incorporated all your suggestions satisfactorily below. \n\n> The lack of a detailed description of the experimental environment and a brief analysis of the experimental results. \n\nWe thank the reviewer for the constructive feedback on our paper. We will make the theoretical part more concise and provide further analysis of our results to justify our contribution for the camera-ready version.\n\n> … The font size of Table 2 is too small to read \n\nWe thank the reviewer for raising this concern and have accordingly increased the size of table 2 by around 20% for the camera-ready version.\n\n", "role": "user" }, { "content": "Thanks to the authors for their clarification. The reply answers almost all of my questions. I continue to recommend this work for acceptance.", "role": "assistant" } ]
y7rbX884LZrX	AutoML 2023 Conference	Reviewer_QQVT	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_QQVT, and the following are the paper you need to review:\n```<<y7rbX884LZrX>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The authors of this paper claim two main contributions, as follows:\n- revisiting and unifying the understanding of metrics used to predict architecture performance in neural architecture search (NAS)\n- proposing novel metrics that outperform existing ones from a theoretical and practical perspective\noverall_review: This paper revisits and unifies the understanding of metrics used to predict architecture performance in neural architecture search (NAS). In addition, authors also propose novel metrics that they claim outperform existing ones from a theoretical and practical perspective. The topic is relevant and, if contributions were achieved, the paper would have significant impact on the field, since it brings a better and broader understanding of existing techniques. In addition, given the relevant of metrics to predict architecture performance in NAS, improved metrics are very welcome and would be readily adopted by the community. Furthermore, the paper is generally well-written, though the assessment part of the paper is considerably brief. Since the theoretical analysis section was rather long and verbose, there is clearly room to better balance the different parts of the paper. More importantly, the experimental assessment is very limited for contributions to be validated. Nonetheless, since the theoretical part of the paper escapes my area of expertise, I evaluate it as a borderline paper.\nreview_summary: This paper addresses metrics used to predict architecture performance in neural architecture search (NAS). In detail, authors revisit, unify, and propose novel metrics. The topic is relevant, since it brings a better and broader understanding of existing techniques. Furthermore, the paper is generally well-written, though the assessment part of the paper is considerably brief. More importantly, the experimental assessment is very limited for contributions to be validated. Since the theoretical part of the paper escapes my area of expertise, I evaluate it as a borderline paper.\nactions_required_to_increase_overall_recommendation: - a thorough experimental assessment of the metrics considered and proposed\n- a revision of the paper, making the theoretical discussion more straightforward and substantiating the paper with experimental results\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper\n", "role": "assistant" }, { "content": "Title: Thank you for your feedback. \nWe thank the reviewer for the positive feedback on our work. We hope to have answered all of your questions satisfactorily below; Please do let us know if you see any further issues in the paper that need to be clarified or addressed.\n\n> The Theoretical analysis section is too long, while the experimental lack sufficient analysis.\n\nWe thank the reviewer for the constructive feedback on our paper. We will make the theoretical part more concise and provide further analysis of our results to justify our contribution in the camera-ready version. Please also refer to our other response for further experiments that clarify the contribution of our approaches. \n", "role": "user" }, { "content": "Thanks for the authors' response. The response covers nearly all of the issues I raised. I keep my recommendation for this paper to be accepted.", "role": "assistant" } ]
y7rbX884LZrX	AutoML 2023 Conference	Reviewer_6bMH	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_6bMH, and the following are the paper you need to review:\n```<<y7rbX884LZrX>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: In this paper, the authors identify some existing metrics and classify them into gradient property based and activation pattern based. They identify the factors that contribute to the success of ZiCo and reformulate NASWOT to perform better. They also produce a state of the art zero cost proxy.\noverall_review: The paper is well written, well motivated and easy to understand. However, the resulting scoring function (T-CET) does not deliver sufficiently improved KT/SPR to have a more profound impact on the community. Focusing on identifying more zero cost proxies and classifying them as done in Figure 1 and drawing insights to develop better zero cost proxies may help solidify the contributions of the paper, as well as more results (NDS search space can be easily initialized from their open source code for generating KT/SPR, as well as more search results demonstrating superior Top1-accuracy) would help.\nreview_summary: The paper is well written, and takes an interesting approach towards improving existing zero cost proxies. Unfortunately, the lack of a significant improvement over existing zero cost proxies, as well as lack of random search based results on more architecture spaces limits the impact of this work in practice. \nactions_required_to_increase_overall_recommendation: 1. More 'random search' results to demonstrate effectiveness of T-CET in practice.\n2. NDS search space correlation scores.\n3. More indepth classification of different zero cost proxies as done in Figure 1 would certainly be interesting.\nRating: 5 marginally below the acceptance threshold\nConfidence: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\n", "role": "assistant" }, { "content": "Title: Response to Reviewer Comments on T-CET and SSNR Metrics\nWe thank the reviewer for the valuable and constructive comments on our work. We hope to have answered all of your questions satisfactorily below. Please do let us know if you see any further issues in the paper that are unclear or need to be addressed.\n\n> The paper is interesting, but the improvement in rank correlation in most cases is ~0.03, which brings into question the effectiveness of T-CET and consequently its impact in practice.\n\nOur paper proposes two novel types of metric: SSNR and T-CET. Firstly, we consider an alternative description for the success of ZiCo, which not only provides an interpretation of ZiCo's success that is more consistent with empirical results but also highlights a method for improving existing metrics, SSNR (which we have modified in accordance with the comments of reviewer 2i84). We find a 0.21 average Kendall Tau correlation improvement of SNIP-SNR over SNIP on NATSBench-SSS and a 0.20 average improvement on NATSBench-TSS. This demonstrates the effectiveness of the first class of metric that we propose. However, the reviewer's specific concern was the effectiveness of T-CET. \n\nThis metric was motivated by our finding that gradient- and activation- based compression metrics were complementary in Section 3.1. In creating T-CET our aims were to produce a high-performance metric, robust to the gradient- or activation- based biases of various search spaces. We acknowledge that the majority of T-CET’s performance can be attributed to SSNR, however, in most of the search spaces (and datasets) that we consider T-CET outperforms SSNR, albeit if the improvement is often rather small. Moreover, it is worth highlighting to the reviewer that the aforementioned modification to SSNR (and thus T-CET) further improved T-CETs performance over existing metrics (discussed further in our response to reviewer 2i84) - later in our response we provide results showing that T-CET(SNIP) outperforms ZiCo in kendall tau correlation on average across NDS by 0.20. \n\nFurthermore, T-CET stems from a comprehensive framework encompassing a novel viewpoint on diverse high-performance metrics and a new (direct) approach to metric combination. We firmly believe that both aspects of this framework will offer valuable insights to the wider research community. A direct approach to metric combination is helpful for some search algorithms and may better discern the highest performing architectures, evidenced by our improvement over SOTA results on the top-1 ZenNAS search space. For the reasons outlined above, although we appreciate that the improvements of T-CET upon our other proposed metric class (SSNR) are modest, we believe that together SSNR and T-CET provide a significant contribution - and are the only metrics we could find, along with ZiCo, that outperform params on average over NATSBench- TSS and SSS. \n\n> The papers comparision of existing metrics (from Figure 1) is interesting, and does indeed support their distinction of expressivity and trainability. However, contextualizing this comparision with more existing metrics would make the point stronger of which proxies fall in which of the introduced classification. (A larger graph with several of the existing proxies would help: synflow, jacob_cov, grad_norm, SNIP, Angle-NAS, TE-NAS, EZNAS, NASWOT, ZenNAS, ZiCo). Since the paper studies compositionality of zero cost proxies, being more thorough with such an evaluation would be valuable.\n\nWe agree that a more formal investigation into the metrics most ripe for combination is an important task. Such research is ongoing and has arguably been prompted by the work of Krishnakumar et al. referred to in our paper. This, however, was not the main focus of this paper. Section 3.1 acted to motivate our interests of understanding, generalising and combining gradient- and activation- based compressibility metrics, thus we restrict our use of metric comparison techniques to specific metric classes relevant to these aims in order to prevent overwhelming the reader - but we do direct them to similar work conducted by Krishnakumar et al. for a more comprehensive analysis. \n\n> The proposed approach of developing T-CET is sound, but its resulting effectiveness is not impressive. To further support the effectiveness of their proxy, I would highly recommend testing on the NDS-CIFAR-10 search space, which is especially difficult for synflow as shown in NASWOT. It is not necessary that a higher SPR/KT would equate to better performance in search algorithms, thus, evaluating top-1 accuracies on a few more search spaces aside from the one in Table 4 may highlight the usefulness of T-CET.\n\nWe agree that the top-1 accuracies demonstrate some of the more significant contributions of this paper. As the reviewer requested, we further performed searching on ZenNAS search space with CIFAR-100 datasets and also compared our approach with more recent SOTA Zico in both CIFAR-10 and CIFAR-100 Setting:\n\n\n\n\| Method \| CIFAR-10 \|CIFAR-100\|\n\|----------------\|------------------------------\| -----------\|\n\| Random \| 93.5+-0.7 \| 71.1+-3.1 \|\n\| FLOPs \| 93.1 \| 64.7 \|\n\| Grad-norm \| 92.8 \| 65.4 \|\n\| Synflow \| 95.1 \| 75.9 \|\n\| NASWOT \| 96.0 \| 77.5 \|\n\| TE-Score \| 96.1 \| 77.2 \|\n\| Zen-Score \| 96.2 \| 80.1 \|\n\| Zico \| 97.0 \| 80.2 \|\n\| T-CET(Synflow) \| 96.6 \| 80.0 \|\n\| T-CET(SNIP) \| 97.2 \| 80.43 \|\n\nT-CET(SNIP) is consistently finding better models than recent SOTA metrics Zen-Score and Zico in the same search settings. We hope this should help the reviewer to acknowledge the advantages of our approaches. \n\nUnder your recommendation, we have expanded our analysis of different search spaces in NDS datasets, the results (Kendall tau correlations) are listed in the following:\n\n\| \| Darts \| Darts-fix-w-d \| NASNet \| ENAS \| Amoeba \| PNAS \| ResNet \| Resnext-b \|\n\|----------------\|-------\|---------------\|--------\|------\|--------\|------\|--------\|-----------\|\n\| Synflow \| 0.30 \| -0.09 \| 0.02 \| 0.13 \| -0.06 \| 0.18 \| 0.14 \| 0.43 \|\n\| SNIP \| 0.27 \| -0.13 \| -0.05 \| 0.09 \| -0.09 \| 0.15 \| 0.26 \| 0.45 \|\n\| ZiCo \| 0.34 \| 0.09 \| 0.09 \| 0.20 \| -0.02 \| 0.19 \| 0.14 \| 0.47 \|\n\| Synflow-SNR \| 0.53 \| 0.10 \| 0.29 \| 0.41 \| 0.25 \| 0.39 \| 0.59 \| 0.48 \|\n\| SNIP-SNR \| 0.54 \| 0.16 \| 0.29 \| 0.41 \| 0.24 \| 0.39 \| 0.60 \| 0.48 \|\n\| T-CET(Synflow) \| 0.54 \| 0.15 \| 0.34 \| 0.44 \| 0.28 \| 0.41 \| 0.32 \| 0.54 \|\n\| T-CET(SNIP) \| 0.56 \| 0.20 \| 0.34 \| 0.45 \| 0.27 \| 0.42 \| 0.34 \| 0.54 \|\n\nWe observe that our approaches significantly improved correlation in different datasets, where vanilla Synflow and SNIP correlated poorly with ground truth accuracy, which further highlights our contribution that compressibility is important for estimating model performance in zero-cost NAS.\n", "role": "user" }, { "content": "Title: Response To Authors\nThank you for your response.\n\nI appreciate that the authors have expanded their analysis of different search spaces in NDS data-set, but still believe that some key related works in the above table are missing. For instance, EZNAS significantly outperforms T-CET on several NDS design spaces. Please note that in some of the NDS search spaces, FLOPs and Params seem to be better or similar metrics for accuracy (DARTS, ENAS).\n\nI still feel that the paper is well written and well motivated and agree with the insights the paper can provide to the community, but I believe more empirical evidence is required for the results to have a more notable impact in the NAS community. Specifically, it may be very helpful to have more NAS search results to further strengthen the paper given that the KT/SPR correlation results are a little weak. The NAS search results can help us understand how these zero cost proxies are able to differentiate between the top-x% networks in a given design space, a problem highlighted in [1].\n\nSpecifically, ZenNAS CIFAR-10 and CIFAR-100 correlation may be strong enough that the new experiment reported above does not lend us any insight (please correct me if I am wrong about the architecture correlation across data-sets, generally this has been true for several design spaces.)\n\n[1] Abdelfattah, M. S., Mehrotra, A., Dudziak, Ł., & Lane, N. D. (2021). Zero-Cost Proxies for Lightweight NAS.\n", "role": "assistant" }, { "content": "Title: Further response\nWe thank the reviewer for pointing to EZNAS as a reference paper, which significantly contributes to the field of zero-cost NAS. The main contribution of EZNAS is the generation of interpretable and generalisable zero-cost metrics by evolving more fundamental pre-existing operations with expression trees based on the statistics from existing benchmark search spaces. However, we argue that EZNAS is orthogonal to our work, which means the success of EZNAS does not conflict with our contributions, and we will list our reasons below.\n\nFirstly, in our work, we focus on exploring different types of compressibility and propose a unified framework to combine different compressibilities that significantly improve existing zero-cost metrics. Our methods provide unique operation types that EZNAS did not consider: a row-wise signal-to-noise ratio (SSNR). More specifically, EZNAS is limited to creating expression trees; it does not contain a stand-alone standard deviation operation and does not possess an operation for restricting a pre-existing operation to an axis of a tensor, i.e. row/column-wise. For these reasons, we believe there is, at most, a negligible chance of EZNAS finding ZiCo, SSNR, or T-CET. \nIn our further analysis, we find that T-CET outperforms EZNAS on all tasks in NATSBench-TSS/SSS (by up to 0.11 Spearman - shown in the table below), further suggesting the significance of the operations we propose that are unaccounted for by EZNAS.\n\nMoreover, we don’t believe the comparison on NDS is fair, as the EZNAS form is generated based on a fitness objective derived from NDS. Furthermore, their use of approximately 24 hours of CPU time to search for their method, as opposed to an analytic framework, arguably gives them an unfair advantage over our method. We maintain that the NAS community can bolster existing methods by considering the notion of compressibility presented in our paper; therefore, we believe that our work is orthogonal and not in competition with methods like EZNAS. We will further include those analyses in our paper.\n\n\| Search Space (Dataset-Correlation) \| TSS (C10-KT) \| TSS (C100-KT) \| TSS (IN120-KT) \| SSS (C10-SR) \| SSS (C100-SR) \| SSS (IN120-SR) \| DARTS (C10-KT) \| Amoeba (C10-KT) \| ENAS (C10-KT) \| PNAS (C10-KT) \| NASNet (C10-KT) \|\n\|-------------------------------------\|---------------\|------------------\|----------------\|---------------\|---------------\|------------------\|----------------\|---------------\|---------------\|------------------\|----------------\|\n\| FLOPs \t\| 0.56 \t\| 0.54 \t\| 0.50 \t\| 0.61 \t\| 0.28 \t\| 0.58 \t\| 0.51 \t\| 0.26 \t\| 0.47 \t\| 0.34 \t\| 0.20 \t\|\n\| #Params \t\| 0.56 \t\| 0.54 \t\| 0.50 \t\| 0.87 \t\| 0.72 \t\| 0.68 \t\| 0.50 \t\| 0.26 \t\| 0.47 \t\| 0.32 \t\| 0.21 \t\|\t\n\| NASWOT \t\| 0.57 \t\| 0.61 \t\| 0.55 \t\| 0.45 \t\| 0.18 \t\| 0.41 \t\| 0.47 \t\| 0.22 \t\| 0.37 \t\| 0.38 \t\| 0.30 \t\|\n\| EZNAS \t\| 0.65 \t\| 0.65 \t\| 0.61 \t\| 0.89 \t\| 0.74 \t\| 0.81 \t\| 0.56 \t\| 0.45 \t\| 0.52 \t\| 0.51 \t\| 0.44 \t\|\n\| SNIP-SNR \t\| 0.68 \t\| 0.65 \t\| 0.63 \t\| 0.92 \t\| 0.76 \t\| 0.92 \t\| 0.54 \t\| 0.24 \t\| 0.41 \t\| 0.39 \t\| 0.29 \t\|\n\| T-CET(SNIP) \t\| 0.69 \t\| 0.65 \t\| 0.62 \t\| 0.92 \t\| 0.63 \t\| 0.85 \t\| 0.56 \t\| 0.27 \t\| 0.45 \t\| 0.42 \t\| 0.34 \t\|\n\n\n\nWith respect to FLOPs and #Params, although T-CET and SSNR didn’t significantly outperform these baseline metrics’ Kendall tau correlations on Amoeba, DARTS and ENAS, they provide significant gains in NASNet, PNAS and all of NATSBench. As requested, we have further considered the accuracies of the top-x models found by each metric. When considering the top-1 model accuracy on NDS, we found that one of either SNIP-SNR or T-CET(SNIP) at least matched both FLOPs and #Params on all search spaces and strictly outperformed them in five out of eight occasions. For the top-5%, T-CET(SNIP) strictly outperformed FLOPs and #Params in every space, which should lead to NAS algorithms with a more stable performance. \n\nMoreover, we found further evidence that a compressibility perspective bolsters existing metrics: in all NDS spaces that we considered, SNIP-SNR strictly outperformed SNIP at top-5% accuracy (and in all but one space for top-1 accuracy). NASWOT proved to be the strongest pre-existing metric we considered in the top-x search; however, we still strictly outperformed it on all but one occasion for top-5% accuracy and were competitive with it for top-1 accuracy. \n\n\| Search Space \\ Top-1 Acc \| FLOPs \| #Params \| NASWOT \| Synflow \| SNIP \| Synflow-SNR \| T-CET(Synflow) \| SNIP-SNR \| T-CET(SNIP) \|\n\|-----------------\|-----------\|-----------\|-----------\|-----------\|-------\|-------------\|----------------\|-----------\|-------------\|\n\| Darts \| 92.67 \| 92.67 \| 93.53 \| 93.71 \| 87.48 \| 92.67 \| 92.67 \| 92.67 \| 93.53 \|\n\| Darts-fix \| 92.17 \| 92.17 \| 92.90 \| 84.99 \| 2.31 \| 92.17 \| 92.90 \| 92.17 \| 92.90 \|\n\| NASNet \| 92.63 \| 92.63 \| 94.25 \| 92.93 \| 86.97 \| 92.63 \| 92.63 \| 92.63 \| 92.63 \|\n\| Resnext-b \| 94.19 \| 94.05 \| 94.05 \| 94.09 \| 94.18 \| 94.09 \| 94.48 \| 94.09 \| 94.48 \|\n\| ENAS \| 93.02 \| 93.02 \| 93.02 \| 93.35 \| 83.31 \| 93.02 \| 93.02 \| 93.02 \| 93.02 \|\n\| Amoeba \| 93.64 \| 93.64 \| 93.30 \| 91.22 \| 86.04 \| 93.64 \| 93.64 \| 93.64 \| 93.64 \|\n\| PNAS \| 93.38 \| 93.38 \| 93.94 \| 92.86 \| 92.86 \| 93.83 \| 93.69 \| 93.83 \| 93.83 \|\n\| ResNet \| 93.59 \| 93.59 \| 92.92 \| 93.26 \| 92.85 \| 93.59 \| 92.99 \| 93.70 \| 92.99 \|\n\| #Best-Acc \| 1 \| 1 \| 3 \| 2 \| 0 \| 1 \| 3 \| 2 \| 3 \|\n\n\| Search Space \\ Top-5% Acc \| FLOPs \| #Params \| NASWOT \| Synflow \| SNIP \| Synflow-SNR \| T-CET(Synflow) \| SNIP-SNR \| T-CET(SNIP) \|\n\|-----------------\|-----------\|-----------\|-----------\|-----------\|-------\|-------------\|----------------\|-----------\|-------------\|\n\| Darts \| 93.44 \| 93.44 \| 93.64 \| 92.52 \| 92.07 \| 93.75 \| 93.82 \| 93.84 \| 93.90 \|\n\| Darts-fix \| 91.89 \| 91.91 \| 92.00 \| 90.87 \| 89.17 \| 92.09 \| 92.13 \| 92.23 \| 92.18 \|\n\| NASNet \| 92.67 \| 92.64 \| 92.63 \| 90.38 \| 88.26 \| 92.66 \| 92.80 \| 92.67 \| 92.82 \|\n\| Resnext-b \| 93.26 \| 93.58 \| 93.49 \| 93.52 \| 93.54 \| 93.59 \| 93.88 \| 93.59 \| 93.88 \|\n\| ENAS \| 93.14 \| 93.14 \| 92.92 \| 90.94 \| 90.55 \| 92.97 \| 93.13 \| 93.19 \| 93.25 \|\n\| Amoeba \| 91.24 \| 91.26 \| 91.14 \| 88.63 \| 87.32 \| 91.75 \| 91.95 \| 91.62 \| 91.80 \|\n\| PNAS \| 93.50 \| 93.51 \| 93.83 \| 93.20 \| 92.85 \| 93.56 \| 93.59 \| 93.51 \| 93.58 \|\n\| ResNet \| 93.19 \| 93.82 \| 93.13 \| 93.31 \| 93.24 \| 93.92 \| 94.01 \| 93.83 \| 94.00 \|\n\| #Best-Acc \| 0 \| 0 \| 1 \| 0 \| 0 \| 0 \| 3 \| 1 \| 4 \|\n\nOur analysis highlights the advantages of SSNR and T-CET, illustrating that compressibility provides an important perspective that is largely overlooked by EZNAS’ operation vocabulary and hints at ways in which methods like EZNAS may be further improved. ", "role": "user" }, { "content": "Title: Response to authors\nThank you for your thorough response.\n\nThanks for remarking that a comparision on NDS specifically may be unfair, I agree. I also comprehend that the paper presents the idea of compressibility, which appears to strengthen existing zero-cost proxies. This could be beneficial for the community, but I'm still concerned about the extent of its impact.\n\nThank you for sharing the mentioned results; over how many trials were they conducted? Generally, there can be considerable variability in NAS search results.\n\nI have revised my rating and will also review other discussions once they have finished to determine if I need to adjust my score further.\n\n\n\n\n", "role": "assistant" }, { "content": "Title: Thank you for your comments\nThank you for your comments. We scored all networks in the search space and reported the accuracy of the top-1 and the average of the top-5% models. Thus we only run experiments with the same initialization seeds one time. We understand that zero-cost scores can vary slightly because of random initialization and random batches, but the overall correlation should be relatively stable across different seeds. We thank for the reviewer raising this concern on the robustness of randomness; we will add an experiment for different random seeds and report accuracy in a mean +- std manner. \n\n", "role": "user" }, { "content": "I truly enjoyed the above exchange with the author, which has significantly improved the quality of this work.", "role": "assistant" } ]
y7rbX884LZrX	AutoML 2023 Conference	Reviewer_2i84	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_2i84, and the following are the paper you need to review:\n```<<y7rbX884LZrX>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The article discusses Neural Architecture Search (NAS) and how it has been successful in discovering high-performance neural network architectures. However, current NAS approaches require extensive evaluation or training, so zero-cost proxies have been proposed to efficiently predict architecture performance. The article argues that existing proxies should be revisited and analyzed through the lens of network compressibility, trainability, and expressivity. By doing so, a better understanding of the high-level relationship between different proxies can be built and some can be refined. The article proposes a novel saliency and metric aggregation method informed by compressibility, orthogonality, and network topology that yields state-of-the-art results in popular NAS benchmarks.\noverall_review: In general, this paper will be interesting to people working in NAS and AutoML.\nI have voiced my concerns above.\nreview_summary: This paper provides a concrete analysis of the new proxy it proposes (which is very interesting), and conducts comprehensive experiments.\nI vote weak accept. I hope the author could address my concerns above.\nactions_required_to_increase_overall_recommendation: I am open to further discussions, and willing to adjust my score if the author could respond my questions above.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Thank you for your support of our work and Further Comments\nWe thank the reviewer for the positive feedback, and appreciate the time and effort that the reviewer has shown to help us improve our work. We have further updated our paper and hope to have incorporated all your suggestions satisfactorily below. \n\n> … Where is K^l defined(line223)\n\nThis definition was indeed lacking and we have clarified this for the camera-ready version. \n\n> In addition to the explanation on page 7, are there any ablation studies on different ways to combine S and K in Eq. 8? E.g. with or without logarithm.\n\nWe are indebted to this recommendation. In T-CET there are two log’s: one on the determinant (NASWOT) and one on the signal to noise ratio (SSNR). The logdet on the NASWOT kernel is required to prevent data-overflow, thus is a non-negotiable. The log on the signal to noise ratio was motivated by 2 considerations: to make SSNR’s formulation consistent with ZiCo and to stabilise the computation such that no layer dominates the score. \n\nUpon performing an ablation study we find that the logarithm does improve performance on some search spaces: NasBenchMacro, NDS-Resnext-b, NATSBench-SSS, ZenNAS - albeit not on others. We explain the disparity in the effect of the logarithm on search space results by the variance in layer-size. We believe that search spaces containing networks exhibiting significant variation in layer width, ‘unstable search spaces’, require the log to stabilise the results. By its very nature, the logarithm effectively replaces layer-wise addition of scores with layer-wise multiplication. This prevents the layers with the largest values from completely dominating the final score, even if some layers’ values are exponentially larger than others. Conversely, in search spaces with more consistent layer-widths, ‘stable search spaces’, we contend the logarithm is unnecessary. Rather than enhancing performance, the logarithm further contracts the already small differences between layer scores, making the system more susceptible to noise and consequently degrading its performance. In summary, the inclusion or exclusion of the log does not change the fundamental tenets of SSNR but merely helps to stabilise performance in some search spaces. \n\nAlthough these arguments for explaining the observed behaviour are not particularly rigorous, and more theory would be needed to more confidently justify our changes we can at the very least provide an empirically-justified recommendation for when to use the logarithm with SSNR. Accordingly, in the camera-ready version our table results will only stabilise SSNR with the log when the stability of the search spaces requires it.\n\nIn more detail, our ablation study assessed the average Kendall tau correlation of log-less SSNR (the signal to noise ratio with no logarithm acting on it) and SSNR (as proposed in our paper) over a number of search spaces and tasks. We ranked log-less Synflow-SNR, Syflow-SNR, log-less SNIP-SNR and SNIP-SNR on a number of search spaces and datasets. On the less stable search spaces (NasBenchMacro, NDS-Resnet-b, NATSBench-SSS) the logarithm proved useful, outperforming its log-less counterpart.\n\n\| \| Resnext-b \| Macro - SS1 \| SSS (C10) \| SSS (C100) \| SSS (IN120) \| Rank Average \|\n\|----------------------\|-----------\|-------------\|-----------\|--------------\|-------------\|--------------\|\n\| Log-less Synflow-SNR \| 0.45 \| 0.22 \| 0.69 \| 0.53 \| 0.67 \| 3.8 \|\n\| Synflow-SNR \| 0.48 \| 0.52 \| 0.75 \| 0.56 \| 0.76 \| 1.8 \|\n\| Log-less SNIP-SNR \| 0.49 \| 0.23 \| 0.69 \| 0.53 \| 0.68 \| 2.8 \|\n\| SNIP-SNR \| 0.48 \| 0.52 \| 0.75 \| 0.56 \| 0.76 \| 1.6 \|\n\n\n\n\n\nOn the more stable search spaces (NATSBench-TSS, NASBench1shot1, BLOX and the rest of NDS) the Kendall tau correlations were higher for log-less SNR as opposed to SNR.\n\n\| \| Darts \| Darts-fix-w-d \| NASNet \| ENAS \| Amoeba \| PNAS \| ResNet \| BLOX \| 1shot1 (SS1) \| 1shot1 (SS2) \| 1shot1 (SS3) \| TSS (C10) \| TSS (C100) \| TSS (IN120) \| RankAverage \|\n\|----------------------\|-------\|---------------\|--------\|------\|-------\|------\|--------\|-------\|--------------\|--------------\|--------------\|------\|------------\|-------------\|-------------\|\n\| Log-less Synflow-SNR \| 0.53 \| 0.10 \| 0.29 \| 0.41 \| 0.25 \| 0.39 \| 0.59 \| 0.39 \| 0.52 \| 0.60 \| 0.49 \| 0.63 \| 0.60 \| 0.59 \| 2.29 \|\n\| Synflow-SNR \| 0.35 \| 0.08 \| 0.09 \| 0.20 \| -0.01 \| 0.20 \| 0.15 \| -0.34 \| 0.50 \| 0.57 \| 0.48 \| 0.60 \| 0.60 \| 0.59 \| 3.86 \|\n\| Log-less SNIP-SNR \| 0.54 \| 0.16 \| 0.29 \| 0.41 \| 0.24 \| 0.39 \| 0.60 \| 0.40 \| 0.58 \| 0.62 \| 0.52 \| 0.68 \| 0.65 \| 0.63 \| 1.14 \|\n\| SNIP-SNR \| 0.35 \| 0.09 \| 0.09 \| 0.21 \| -0.01 \| 0.20 \| 0.16 \| -0.23 \| 0.52 \| 0.58 \| 0.50 \| 0.63 \| 0.63 \| 0.62 \| 2.71 \|\n\nWe found similar, albeit slightly more modest, improvements for the log vs log-less variants of T-CET. We will update our tables in our paper with improved results for the camera-ready version. \n\n> The novelty is slightly limited, as both Gradient (SynFlow over variance) and Action-space (layer-wise Naswot) Compressibility are proposed by existing works or modified versions.\n\nWe disagree with this characterization of the literature. We do agree that NASWOT is, at least implicitly, described as a measure of compressibility; however, this is not the case for gradient-based compressibility. Arguments around compressibility are not used to explain the effectiveness of ZiCo and we need to be aware of any previous literature that considers the row-wise compressibility of the gradient i.e., saliency over variance. Infact, a key part of the novelty of this work derives from the fact that it generalises the tenets of some of the highest-performance metrics and from these insights generates a novel class of row-compressed metrics, SSNR. \n", "role": "user" }, { "content": "Thank you for the detailed response from the authors. It adequately addresses the points I raised. I will continue to recommend accepting this work.", "role": "assistant" } ]
y7rbX884LZrX	AutoML 2023 Conference	Reviewer_Qh4i	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_Qh4i, and the following are the paper you need to review:\n```<<y7rbX884LZrX>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper discusses Neural Architecture Search (NAS) to discover high-performance neural network architectures. However, current NAS approaches can be costly, requiring extensive evaluations or training of large super networks. Zero-cost proxies have been proposed to predict performance efficiently to reduce search costs, but little attention has been given to their mutual effects. The paper analyzes existing proxies through network compressibility, trainability, and expressivity and designs a new saliency and metric aggregation method informed by these factors, leading to state-of-the-art results across popular NAS benchmarks.\noverall_review: The proposed method in the paper appears to be a unique approach to combining different zero-cost NAS (ZC-NAS) methods for evaluating neural architectures. However, the results achieved are only marginally better or sometimes worse than those of single ZC-NAS methods, which raises some concern. Additionally, the paper mentions TE-NAS in the related work section but does not provide further details about it in other parts of the paper, which could be confusing.\n\nIt is also worth noting that the paper \"Generic Neural Architecture Search via Regression\" presents a near zero-cost method with significantly better results. This suggests that while the approach in the current paper is interesting, there may be other more effective methods for zero-cost NAS that the authors could consider exploring.\n\nIn general, I believe this is a well-written paper. However, if the authors could improve the significance of their results, I would be inclined to give it a higher rating.\n\nreview_summary: Please refer to the above section.\nactions_required_to_increase_overall_recommendation: Please refer to the above section.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Thank you for your support of our work\nWe thank the reviewer for the positive feedback, and appreciate the time and effort that the reviewer has shown to help us improve our work. We have further updated our paper and hope to have incorporated all your suggestions satisfactorily below. \n\n\n> Kindly ask the author to provide the topology in Appendix for searched structure in the ZenNAS search space.\n\nWe will add that searched architecture in the appendix.\n\nT-cet(SNIP): `SuperConvK3BNRELU(3,64,1,1)SuperResK1K3K1(64,152,1,40,5)SuperResK1K3K1(152,128,1,48,5)SuperResK1K5K1(128,256,2,24,4)SuperResK1K5K1(256,40,2,16,1)SuperResK1K5K1(40,80,2,32,1)SuperConvK1BNRELU(80,64,1,1)`\n\nT-cet(Synflow):\n`SuperConvK3BNRELU(3,176,1,1)SuperResK1K3K1(176,152,1,48,4)SuperResK1K3K1(152,152,1,32,5)SuperResK1K3K1(152,240,1,32,4)SuperResK1K3K1(240,112,2,48,1)SuperResK1K3K1(112,120,2,16,1)SuperResK1K3K1(120,112,2,40,1)SuperConvK1BNRELU(112,96,1,1)`\n\nZico:\n`SuperConvK3BNRELU(3,96,1,1)SuperResK1K3K1(96,80,1,48,1)SuperResK1K3K1(80,192,2,48,5)SuperResK1K3K1(192,64,2,40,1)SuperResK1K3K1(64,192,1,32,5)SuperResK1K3K1(192,96,2,40,4)SuperConvK1BNRELU(96,80,1,1)`\n\n> Please consider making all the terms consistent with the original paper, such as NASWOT (instead of naswot)\n\nWe thank the reviewer for highlighting this inconsistency and have amended it for the camera-ready version. We are also amending inconsistent notation for FLOPS and GraSP in the tables. \n\n> However, the results achieved are only marginally better or sometimes worse than those of single ZC-NAS methods, which raises some concern. \n\nAs explained further in our response to reviewer 6bMH, our paper provides two distinct contributions. Our first proposed class of metrics is SSNR, which is a single metric that often significantly outperforms SOTA metrics. Although T-CET only produces a modest improvement upon SSNR, it significantly outperforms pre-existing SOTA and provides a proof-of-concept for direct metric combination, an as yet unexplored area of research. \n\n> Additionally, the paper mentions TE-NAS in the related work section but does not provide further details about it in other parts of the paper, which could be confusing.\n\nWe respectfully disagree with this account. TE-NAS is introduced in the related work section for two reasons. Firstly, it attempts to combine metrics for expressivity and trainability. Secondly, it is an example of an indirect means of metric combination from which our method is proposed in response to. With regards to the first point, towards the end of Section 3.1 we reconsider ‘trainability’ and ‘expressivity’ directly citing TE-NAS. With regards to the second point, when introducing the indirect combination method RankAve in Section 3.3 we directly cite TE-NAS. \n\n> It is also worth noting that the paper \"Generic Neural Architecture Search via Regression\" presents a near zero-cost method with significantly better results. This suggests that while the approach in the current paper is interesting, there may be other more effective methods for zero-cost NAS that the authors could consider exploring.\n\nWe credit GenNAS as being an insightful and important contribution to the field of zero-cost NAS, however its focus is somewhat different to the motivations behind our paper, and zero-cost metrics more generally. GenNAS predicts a network's ability, on a range of tasks, by considering the ability of extracted tensors from its intermediate layers to regress towards signal bases. It provides a strong performance predictor of the network’s ability to perform a variety of difficult tasks via considering its performance on a generic low-cost proxy task. This however does not provide insight into why a given network performs well, beyond its ability to perform well on the generic tasks considered by GenNAS. \n\nIn contrast, zero-cost proxies extract more fundamental properties of the network, be that synaptic flow, gradient variance, etc. These likely can’t compete on performance with methods like GenNAS, but this sacrifice allows for interpretability. Zero-cost metrics implicitly provide an explanation for why a given network is performing well. The narrative of this paper is not that SSNR and T-CET outperform all (near) zero-cost methods but that they outperform other common zero-cost metrics and provide an interpretable overarching explanation for network success in terms of (layer-wise) compressibility be that gradient- or activation- based. \n\n", "role": "user" }, { "content": "Thanks to the authors for their clarification. The response covers nearly all of the issues I raised. I still support my favorable recommendation for this manuscript.", "role": "assistant" } ]
ydqqogDD5RW	AutoML 2023 Conference	Reviewer_z9vf	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_z9vf, and the following are the paper you need to review:\n```<<ydqqogDD5RW>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper focuses on the research problem of self-adjusting hyper-parameters of Bayesian Optimization (BO) to balance the exploration and exploitation. \nThe contibutions of this paper are: \n1. Clearly formulation of the problem, i.e., when and how should we adjust the hyper-parameters. \n\n2. The proposed method can self-adjust the important hyper-parameter of acquisition function.\noverall_review: Pros\n\n1. This paper is well-motivated and well-written.\n\nCons\n1. The advantages of the current method is not significant enough. For example, the final rank of the proposed method is over 4 (of 11) on two types of test-beds.\n\n2. More issues please refer to \"Actions Required To Increase Overall Recommendation\".\nreview_summary: This paper is well-motivated and well-written. The research problem is important. However, the proposed method is not novel and the experimental results are weak.\nactions_required_to_increase_overall_recommendation: 1) More experiments on some non-synthetic tasks would further improve this work. \n\n2) More compared baselines, such as Multi-armed bandits.\n\n3) I notice that the authors have leave the comparison with DAC (RL-based method) as the future work. However, RL is a popular and promosing method in recent work in this research area, and shows the better performance than the heuristic-based methods, especially in some black-box optimization methods [1, 2]. I believe just compare one simple method based on RL (e.g., DQN with action that increase, decrease, or keep $\\alpha$) will significantlly strengthen the results.\n\n[1] Learning Step-Size Adaptation in CMA-ES. PPSN, 2020.\n\n[2] Multi-agent Dynamic Algorithm Configuration. NeurIPS, 2022.\n\n4) Some theoretical analysis will also significantly strenghthen the results.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Initial Response\nThank you for your review. We would like to address your comments in the following paragraphs.\n\n## Significance\n\n> The advantages of the current method is not significant enough. For example, the final rank of the proposed method is over 4 (of 11) on two types of test-beds.\n\nThis is because the performance depends on the specific landscape and it is quite unlikely to have an optimal search attitude for each problem/instance - and this is not even what we aim for. We show that even with an average rank of 4 on BBOB, our method works out of the box and is on par with other strong baselines. So, for practitioners with no knowledge about the black-box landscape (as the name black-box suggests) do not know anything about our problem structure, SAWEI is a very robust and adaptive choice, not only w.r.t. final performance but also anytime performance.\nThis is especially apparent in the results on HPOBench (Figure 4)\n\n## Experiments\n\n> More experiments on some non-synthetic tasks would further improve this work.\n\nWe emphasize that we evaluated SAWEI on 5 different models with different configuration spaces on 8 datasets. So, overall we show results on 40 HPO tasks. We believe that this is much more than most papers on BO provide as evidence and is considered a very thorough analysis on non-synthetic benchmarks.\n\n## More Baselines / Multi-Armed Bandits\n\nWe note that we included a multi-armed bandit-based method which is Portfolio Allocation [Hoffman et al., 2011]. There, they have a portfolio of nine acquisition functions (three parametrizations of each EI, PI and LCB). They optimize their portfolio with their algorithm GP-Hedge, adjusting the selection probabilities (which acquisition function is used to determine the next configuration) by the mean of the selected configuration. We note that although we consider only the adaptation of a single acquisition function (WEI), we clearly outperform GP-Hedge.\nTo add even more baselines, we also evaluated the well-known LCB that is also outperformed by SAWEI.\n\n## Comparison with RL-based method\n\nWhile RL promises more tailored solutions, it is not an easy-to-integrate baseline. First, RL needs to be trained offline and the number of required data points is huge compared to the offline-free, on-the-fly adaptation of SAWEI. From our own experiences, we expect that RL approaches require thousands or even millions of samples to perform well and robustly. In view of that each sample in BO refers to a training of machine learning algorithm, this is often not feasible in practice. In addition, it is not clear from the beginning what are good state features to describe the state of BO. Although we see promise in applying RL-based methods, comparing against DAC-BO is out of scope of this paper.\n\nThe strong point of our method is that a simple idea, working adaptively on the fly with no pretraining necessary, yields strong performance.\n\n Global/general Response:\n```\nWe thank all reviewers for their thoughtful and valuable feedback. We address individual comments in direct replies to the reviews.\nAs a result of the feedback received we made several updates to the paper:\n\n- Added LCB as another baseline\n- Improved markers and colors for Figures 2, 3 (b,d), 4\n- Updated Figure 6 (Hyperparameter Sensitivity) (same scale, better font sizes)\n- Reworded ablation study\n- Refined description of search attitude\n- Moved Appendix section “How to Detect Adjustment Point” to “SAWEI in a Nutshell”, better describes usage of hyperparameter $\\epsilon$\n- Improved names for baselines, i.e. the modulated PI (PI*) is now denoted as WEI($\\alpha=1$) to avoid confusion\n- Add short appendix section to show that UBR changes after changing optimizer settings (e.g. switching acquisition function)\n- Add Figure 5b and description to directly compare the different behavior of SAWEI on BBOB and HPOBench in terms of $\\alpha$\n- Rephrased beginning of “Related Work”\n\nWe marked the updated text with green color.\n```\n", "role": "user" }, { "content": "Title: Response to authors\nThanks for your replies. I will increase my score. ", "role": "assistant" } ]
ydqqogDD5RW	AutoML 2023 Conference	Reviewer_7ELw	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_7ELw, and the following are the paper you need to review:\n```<<ydqqogDD5RW>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This work proposes an adaptive strategy to adjust the hyper-parameter in weighted Expected Improvement (WEI). The strategy is simple yet efficient.\noverall_review: Strength:\n\n++The proposed SAWEI is simple yet efficient.\n\nWeakness:\n\n-- I have one concern about generalizability. This method aims to adjust the hyper-parameter of the weighted EI acquisition function adaptively, and it can degrade to eighter pure exploration, EI, or PI. I am afraid that this method cannot integrate other acquisition functions, such as UCB, TS, Knowledge gradient, etc.\n\n-- I have concerns about the settings of the hyper-parameter. \n\n1. Why did the authors fix the $\\Delta \\alpha$? How about adjusting it according to the UBR?\n\n2. According to Fig. 6(a), it seems that $\\epsilon=0.05$ has the best performance, then why the authors set $\\epsilon=0.1$ by default (in Line 166)?\n\n\nreview_summary: I am afraid that this method should be further refined so as to integrate more rich acquisition functions. Moreover, the settings of $\\epsilon$ and $\\Delta \\alpha$ might be kind of groundless.\nactions_required_to_increase_overall_recommendation: Please see the weakness.\nRating: 4 possibly reject, but has redeeming facets\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Initial Response\nThank you for your review. In the following we address your concerns about SAWEI’s hyperparameters and generalizability.\n\n## Setting of SAWEI’s Hyperparameters\n\nRegarding the hyperparameter settings:\n\n[ $\\Delta\\alpha$] We agree that setting $\\Delta\\alpha$ as a function of the UBR proposes an elegant solution.\nHowever, with a fixed $\\Delta\\alpha$ we are already able to determine well-performing $\\alpha$ schedules per problem at hand. Therefore, we leave this as a future work to further improve SAWEI.\n\n[$\\epsilon$] This is a good spot. However we can only view those results in the context of the ablation benchmark which was BBOB. The chosen $\\epsilon$ performs well on both inherently different benchmarks, BBOB and HPOBench. Since these are our results on new test functions (not used for development), we cannot tune SAWEI’s hyperparameters on that without risking overfitting and thus providing over-optimistic empirical results.\n\n## Generalizability\n\nThere are two core ideas behind this work. First, we want to adjust the exploration-exploitation trade-off online. Second, we want to adjust it whenever we converge with the current settings. The idea of adjusting the weight of a linear combinations not exclusive to WEI: We can think of linearly combining any two acquisition functions, e.g. UCB or TS. We chose WEI because its summands can be attributed to explorative or exploitative behavior. With the first goal in mind, to adjust the trade-off, other acquisition functions need to be chosen with care.\n\nIn addition, SAWEI exhibits two different attitudes on the two suites that we consider (please see the comparison of $\\alpha$ schedules in the newly added Figure 5b and paragraph in 4.1 ). This (1) motivates the need for a dynamic and self-adjusting AF, (2) is interpretable given the very different nature of the two summands (more exploitative/more explorative) and (3) renders it likely that SAWEI is applicable on a multitude of domains.\n\n Global/general Response:\n```\nWe thank all reviewers for their thoughtful and valuable feedback. We address individual comments in direct replies to the reviews.\nAs a result of the feedback received we made several updates to the paper:\n\n- Added LCB as another baseline\n- Improved markers and colors for Figures 2, 3 (b,d), 4\n- Updated Figure 6 (Hyperparameter Sensitivity) (same scale, better font sizes)\n- Reworded ablation study\n- Refined description of search attitude\n- Moved Appendix section “How to Detect Adjustment Point” to “SAWEI in a Nutshell”, better describes usage of hyperparameter $\\epsilon$\n- Improved names for baselines, i.e. the modulated PI (PI*) is now denoted as WEI($\\alpha=1$) to avoid confusion\n- Add short appendix section to show that UBR changes after changing optimizer settings (e.g. switching acquisition function)\n- Add Figure 5b and description to directly compare the different behavior of SAWEI on BBOB and HPOBench in terms of $\\alpha$\n- Rephrased beginning of “Related Work”\n\nWe marked the updated text with green color.\n```\n", "role": "user" }, { "content": "Thanks for the authors' clarification. They've addressed nearly all of my concerns, and I am considering increasing the score.", "role": "assistant" } ]
ydqqogDD5RW	AutoML 2023 Conference	Reviewer_naa8	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_naa8, and the following are the paper you need to review:\n```<<ydqqogDD5RW>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This work suggests a new acquisition function, self-adjusting weighted expected improvement, in Bayesian optimization. To control an ability to explore and exploit, it adds a new parameter to balance both exploration and exploitation. In particular, such a parameter is adaptively adjusted. The authors show some experimental results of diverse acquisition functions.\noverall_review: I think that it proposes a novel acquisition function, which is very interesting. However, the proposed acquisition function is somewhat heuristic. I think that more detailed justification is required.\n\n* I think that the authors need to do more exhaustive analysis on the proposed method.\n\n* Baselines are unclear. Please elaborate them. In particular, please add a description of EI -> PI* (Linear) and PI* -> EI (linear).\n\n* According to the previous work, the upper bound regret is used to determine when we terminate a Bayesian optimization round. Why can it be employed to adjust exploration and exploitation?\n\n* Could you explain how you implement Line 10 of Algorithm1?\n\n* I think that Figures 3(a) and 3(c) are redundant. If a space for contents is tight, it is okay to move them to the appendix.\n\n* It can be a preference, but the y-axes of Figures 2, 4, and 5 should be flipped, i.e., higher ranks are upper.\nreview_summary: Please see the text boxes above.\nactions_required_to_increase_overall_recommendation: I am on a positive side now. I would like to hear how you will revise the paper and justify the proposed method.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Initial Response\nThank you for reviewing our work. In the following we will address your comments.\n\n## Clarity and Presentation \n\nWe updated our figures for better accessibility (new colors, bigger font sizes, no overlaps).\nWe keep the rank orientation as is to fit the notions of \"lower is better\" for the log regret and UBR. We hope you are fine with this.\nIn addition we improved the naming of our baselines to avoid confusion.\n\n## UBR as a signal to adjust\n\nThe upper bound regret (UBR) can be used to stop Bayesian optimization (BO) [Makarova et al., 2022]. This means the UBR signals whether it is worth continuing optimization. We add our intuition that this holds for the current optimizer settings. This is empirically supported by observing the UBR for the switching policies (EI to PI) where we see sharp bends in the UBR after switching.\nIn our case, current setting implicitly describes the search attitude whether it is exploring or exploiting. Our intiuition is that if we converged with a lot more exploration, more exploitation should help next, and vice versa. Therefore we can use the UBR to signal when we should change our search attitude. Just to clarify: We do not use the UBR itself to adjust the exploration-exploitation trade-off.\nWe added a short section with this in the Appendix and refer to it in \"When To Adjust\".\n\n## Smooth with IQM (Line 10 in Algorithm 1)\n\nThe UBR exhibits a rough signal therefore we smooth it as follows: We apply a moving interquartile mean (IQM) (25%-75%, to remove the spikes) with a window size of 7. This is basically the moving average but with IQM. We updated the description of the algorithm accordingly in \"SAWEI in a Nutshell\".\n\n Global/general Response:\n```\nWe thank all reviewers for their thoughtful and valuable feedback. We address individual comments in direct replies to the reviews.\nAs a result of the feedback received we made several updates to the paper:\n\n- Added LCB as another baseline\n- Improved markers and colors for Figures 2, 3 (b,d), 4\n- Updated Figure 6 (Hyperparameter Sensitivity) (same scale, better font sizes)\n- Reworded ablation study\n- Refined description of search attitude\n- Moved Appendix section “How to Detect Adjustment Point” to “SAWEI in a Nutshell”, better describes usage of hyperparameter $\\epsilon$\n- Improved names for baselines, i.e. the modulated PI (PI*) is now denoted as WEI($\\alpha=1$) to avoid confusion\n- Add short appendix section to show that UBR changes after changing optimizer settings (e.g. switching acquisition function)\n- Add Figure 5b and description to directly compare the different behavior of SAWEI on BBOB and HPOBench in terms of $\\alpha$\n- Rephrased beginning of “Related Work”\n\nWe marked the updated text with green color.\n```\n", "role": "user" }, { "content": "Title: Thank you for your response\nThank you for your response.\n\nI acknowledge that I have read your rebuttal. I will keep my score.", "role": "assistant" } ]
ydqqogDD5RW	AutoML 2023 Conference	Reviewer_2V2m	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_2V2m, and the following are the paper you need to review:\n```<<ydqqogDD5RW>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper proposes a way to set a trade-off parameter $\\alpha$ in a Bayesian optimization (BayesOpt) policy proposed in previous work.\nSpecifically, this parameter $\\alpha$ controls the behavior of the resulting policy on the exploration/exploitation spectrum of Expected Improvement (EI) -- Probability of Improvement (PI), and the paper proposes to adjust the value of $\\alpha$ based on optimization history in the following way.\nFirst, we keep track of a metric called the Upper Bound Regret (UBR), which is the difference between a pessimistic estimate of the incumbent value and a lower bound on the lowest function value.\nIf this metric grows smaller as optimization progresses, it is an indication that optimization with the current setting is converging, and we might benefit by adjusting the value of $\\alpha$.\n\nTo update the value of this parameter, we inspect the trade-off between exploration and exploitation in the most recent queries.\nIf recent queries have favored exploration, we adjust the parameter to bias exploitation and vice versa.\nThis strategy aims to counteract getting stuck at a local optimum from how we are currently balancing exploration and exploitation.\n\nThe authors then present comprehensive experiment results comparing the proposed algorithm against a wide range of benchmarks.\noverall_review: The paper is well-structured and has a nice flow.\nThe proposed algorithm is natural and easy to understand: it involves adjusting the trade-off parameter in a direction opposite to recent behavior to avoid getting stuck at local optima.\n\nThe paper extends a previously proposed BayesOpt policy by designing a way to automatically set its trade-off parameter.\nAs such, I worry that the contributions are somewhat incremental.\n\nAs noted in the Technical Quality And Correctness portion, some of the details about the algorithm are not clear or missing from the manuscript.\nThis makes appreciating the proposed algorithm more difficult.\nIt's also somewhat disappointing that the paper doesn't mention theoretical guarantees for the proposed policy.\nGiven its connection to PI and EI, both of which have their own guarantees, I was hoping some results could be derived for SAWEI.\n\nFinally, I believe the experiment section should include other commonly used BayesOpt policies to ensure that there's value in adopting the current approach.\nreview_summary: I recommend weak rejection.\nWhile I think this paper proposes an interesting approach to the problem of controlling the behavior of WEI, I fear the contributions are limited and some of the crucial details are missing to make this work a strong submission.\nactions_required_to_increase_overall_recommendation: I'd like to understand the exploration term in the acquisition function of SAWEI better, so some discussion on this will be appreciated.\nCrucially, more details about how to compare recent exploration vs. exploitation and how the UBR is computed and tuned (by setting $\\beta_t$) is needed.\n\nEstablishing some theoretical guarantee would help justifying the algorithmic design.\nIncluding stronger baselines such as entropy search policies will make the experiment section more well-rounded.\nRating: 4 possibly reject, but has redeeming facets\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Initial Response\nThank you for reviewing our paper!\n\n## Figures\n\nWe updated our figures for better accessibility.\n\n## UBR and $\\beta_t$\n\nWe set $\\beta_t = 2 \\log (d t^2 / \\beta), \\beta = 1$ for UCB/LCB as done in SMAC3 [Lindauer et al, 2022] following the original UCB [Srinivas et al., 2010].\nWe added this information in the experimental setup.\nWe agree that an analysis on the behavior of UBR depending on $\\beta_t$ is interesting work to refine the stopping criterion [Makarova et al., 2022] and our work.\n\n## Search Attitude\n\nEmpirically, we see that both versions, PI and the modified PI (the exploitation-term of WEI), perform almost on par, with PI a little better. We added this as a comment in “How to Adjust” and a small appendix section. The scales of both terms are different due to the weighting of $(\\mu(x) - f_{min})$.\nWe would also like to stress that we do not denote $\\alpha=0$ as pure exploitation, merely as the exploring term in WEI.\n\n## True Exploration\n\nWe agree that WEI does not perform true exploration (even with $\\alpha = 0$). Nevertheless, as argued in the original paper it provided a stronger exploration attitude. In principle our method could be extended to linear combinations of other acquisition functions or even to adjust the $\\beta$ in UCB to trade-off exploration-exploitation. In the most extreme case, we could also use our approach to adjust the probability of sampling a point based on the acquisition function or from random sampling – we note that for example, SMAC3 uses a fixed probability for that.\n\n## Inclusion of More Baselines\n\nWe included LCB as an additional baseline which does not change our general conclusions.\n\n## Rank as a Metric\n\nWe choose ranks as a metric because comparing the log regret across functions and instances is difficult: Often we see a wide spread in performance and aggregating obscures what we are originally interested in: What method works best on this specific instance?\nFor completeness, we have boxplots of the log regrets for each function / task in the appendix.\n\n[Lindauer et al., 2022] Marius Lindauer, Katharina Eggensperger, Matthias Feurer, André Biedenkapp, Difan Deng, Carolin Benjamins, Tim Ruhkopf, René Sass, Frank Hutter:\nSMAC3: A Versatile Bayesian Optimization Package for Hyperparameter Optimization. J. Mach. Learn. Res. 23: 54:1-54:9 (2022)\n\n[Makarova et al., 2022] Makarova, A., Shen, H., Perrone, V., Klein, A., Faddoul, J., Krause, A., Seeger, M., and Archambeau, C. (2022). Automatic termination for hyperparameter optimization. In Guyon, I., Lindauer, M., van der Schaar, M., Hutter, F., and Garnett, R., editors, International Conference on Automated Machine Learning, AutoML 2022, 25-27 July 2022, Johns Hopkins University, Baltimore, MD, USA, volume 188 of Proceedings of Machine Learning Research, pages 7/1–21. PMLR.\n\n[Srinivas et al., 2010] Srinivas, N., Krause, A., Kakade, S., and Seeger, M. (2010). Gaussian process optimization in the bandit setting: No regret and experimental design. In Proc. of ICML’10, pages 1015–1022.\n\n Global/general Response:\n```\nWe thank all reviewers for their thoughtful and valuable feedback. We address individual comments in direct replies to the reviews.\nAs a result of the feedback received we made several updates to the paper:\n\n- Added LCB as another baseline\n- Improved markers and colors for Figures 2, 3 (b,d), 4\n- Updated Figure 6 (Hyperparameter Sensitivity) (same scale, better font sizes)\n- Reworded ablation study\n- Refined description of search attitude\n- Moved Appendix section “How to Detect Adjustment Point” to “SAWEI in a Nutshell”, better describes usage of hyperparameter $\\epsilon$\n- Improved names for baselines, i.e. the modulated PI (PI*) is now denoted as WEI($\\alpha=1$) to avoid confusion\n- Add short appendix section to show that UBR changes after changing optimizer settings (e.g. switching acquisition function)\n- Add Figure 5b and description to directly compare the different behavior of SAWEI on BBOB and HPOBench in terms of $\\alpha$\n- Rephrased beginning of “Related Work”\n\nWe marked the updated text with green color.\n```\n", "role": "user" }, { "content": "Title: remaining questions\nI thank the authors for their response.\n- Regarding the two \"versions\" of the PI term, my original question was along the lines of, how are we able to balance between the exploration term and the exploitation term in WEI if the two terms aren't in the same units? I agree that the two versions of PI should perform on par since it seems to me they have the same maximizer, but my concern is: (1) is it possible for one term to always be greater than the other, making it impossible to control for the tradeoff?, and (2) how can the user know that the balance is being facilitated effectively?\n- Thanks for the clarification on pure exploration; that clears up my confusion. Perhaps the text could be updated to indicate that pure exploration isn't being pursued with $\\alpha = 0$.\n- Thanks for adding in LCB. Do you have any thoughts other policies such as entropy search policies and knowledge gradient? I understand your time was limited, but feel free to share any comments.\n- Any thoughts on theoretical guarantees?", "role": "assistant" }, { "content": "Title: Answer to Remaining Questions\n## PI Term\n\nSorry, we misunderstood what you meant. One clarification before we discuss the scale of the terms: We use the normal WEI (Eq. 1) for proposing the next configuration to evaluate.\nWe ran our method SAWEI with the attitude determined on the original WEI terms, dubbed SAWEI (modPI), and observed that for BBOB it performs on par and on HPOBench the performance is not convincing (see [here](https://imgur.com/a/RZyQJwE)). This is because `a_exploit` (the modulated PI term) is always smaller than a_explore (the EI term) steering $\\alpha$ towards exploitation on HPOBench. Using the pure PI as `a_exploit` however rightfully determines the attitudes and steers $\\alpha$ towards exploration.\nTo address your questions, (1) in our results $\\alpha$ is adapted in different directions in our experiments, which can be seen in the Appendix plots. (2) Whether the balance is set effectively is difficult to determine online. Until the convergence of BO (with WEI and fixed $\\alpha$) SAWEI cannot be worse. Afterwards we can only improve further by now using a different $\\alpha$.\n\n\n## Pure Exploration Wording\n\nWe made it explicit in the text that $\\alpha=0$ does not equal pure exploration.\n\n## Other Acquisition Functions\n\nIn the experiments we saw that static acquisition functions (e.g., EI, LCB, PI) are strong on different types of problems. We hypothesize that entropy search and knowledge gradient acquisition functions show a similar behavior, i.e., that they work differently well on different kinds of problems but this needs to be empirically validated.\n\n## Theoretical Guarantees\nWe believe that with careful work theoretical guarantees can be derived. For WEI itself there are no guarantees yet, neither for the static nor the dynamic case.\nOne could think of using theoretical guarantees for EI and PI and superposing them. Factoring in the dynamicity of WEI is not straightforward, one might start with providing guarantees for $\\alpha=0$ and $\\alpha=1$.\nHowever, SAWEI also depends on the Upper Bound Regret (UBR), to be more precise, on the gradients of the UBR. Makarova et al. (2022) do provide a bound for the target objective in terms of statistical error and the simple regret. Both quantities are not known and estimated. For the statistical error they use a cross-validation estimator which is not applicable in our case as we do not necessarily optimize ML models on a dataset. \nA straightforward but unelegant method yielding theoretical convergence guarantees would be interleaving with random samples once in a while, as in Hutter et al. (2011). In the limit, we are then guaranteed to converge to the optimum. A more direct analysis of SAWEI could possibly be done along the lines of Doerr, Doerr & Yang (2016), which provides performance guarantees for a parameter control scheme using an $\\epsilon$-greedy multi-armed bandit approach. \n\n[Hutter et al., 2011] Hutter, F., Hoos, H. H., & Leyton-Brown, K. (2011). Sequential model-based optimization for general algorithm configuration. In Learning and Intelligent Optimization: 5th International Conference, LION 5, Rome, Italy, January 17-21, 2011. Selected Papers 5 (pp. 507-523). Springer Berlin Heidelberg.\n\n[Makarova et al., 2022] Makarova, A., Shen, H., Perrone, V., Klein, A., Faddoul, J. B., Krause, A., ... & Archambeau, C. (2022, September). Automatic termination for hyperparameter optimization. In International Conference on Automated Machine Learning (pp. 7-1). PMLR.\n\n[Doerr, Doerr, Yang, 2016] Doerr, B., Doerr, C., & Yang, J. (2016). k-Bit mutation with self-adjusting k outperforms standard bit mutation. In Parallel Problem Solving from Nature–PPSN XIV: 14th International Conference, Edinburgh, UK, September 17-21, 2016, Proceedings 14 (pp. 824-834). Springer International Publishing.\n", "role": "user" }, { "content": "I am grateful for the opportunity to discuss these points with the author, which has certainly strengthened the paper.", "role": "assistant" } ]
ydqqogDD5RW	AutoML 2023 Conference	Reviewer_D7zQ	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_D7zQ, and the following are the paper you need to review:\n```<<ydqqogDD5RW>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: The paper proposed a self-adjusting acquisition function for BO, based on EI and PI. It compares to a collection of other ways to combine EI and PI on both synthetic and HPO problems.\noverall_review: The strengths of the paper are that it proposes a self-adjusting acquisition function, which they compare to a collection of related benchmarks. It is mostly clearly written, tested on a benchmark relevant to the AUTOML community, and engaged with the interesting question of what acquisition function to use. \n\nThat being said, there are some important elements that are unclear, and which require fixing. It uses two different versions of WEI, which seems to make the baselines less comparable, and lacks details in how the attitude is tracked. The ablation study and the related work start out with claims of what they do that don’t match up with the rest of the paragraph.\n\nOverall, the paper seems to be almost ready: the plots need some work to be more accessible and the writing needs to be clarified in parts. I expect the authors to be able to remedy most, if not all of these concerns in the rebuttal. And in that case the review will of course increase.\n\nreview_summary: It is an interesting paper, relevant to the AUTOML community, but which in its current form is unclear, even misleading in parts, and some details are missing. These parts can be remedied by clarifying the writing and plotting, and I expect to increase the score once this has been done.\nactions_required_to_increase_overall_recommendation: These are listed in the above sections. In particular, clearing up the difference between the different types of WEI and addressing the concerns in clarity and accessibility.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Initial Response\nWe would like to thank you for your in-depth review, we appreciate it! We will address your comments in the next paragraphs. \nWe addressed your low and medium concerns as much as possible in the updated paper version.\nWe would like to address your major concerns in our response in more detail.\n\n## Clarity and Accessibility\n\nWe updated the figures for better accessibility and refined the terms of the baselines for clarity. In addition we clear misunderstandings about how we use WEI.\n\n## PI vs. PI\n\nThank you for pointing out the possible confusions between the PI=WEI($\\alpha=1$) and the true PI. The baselines switching from EI to PI after a certain time are not based off WEI. In contrast, the linear step function from $\\alpha=0.5$ to $\\alpha=1$ and vice versa parametrize WEI, therefore used the notation PI. Table 1 and the figures both are correct. We updated our notation in the text and the figures.\n\n## When SAWEI is better\n\nYou are right with your suggestions: At times e.g. EI is better and sometimes it is not. For the BBOB functions where we know the function landscape we could link the landscape to the performance of the individual methods. On HPOBench, this is not possible.\nThe important thing is that on average SAWEI performs better or similarly with the best baseline. What baseline is best really depends on the problem and this robustness is the strength of our method, i.e. we do not need prior knowledge to obtain solid performance.\n\n## Alpha Schedules are Different for Different Benchmarks\n\n> L276-279. You don’t show this in the paper. Worth referring to the appendix so people know where to find it.\n\nThank you for pointing this out. You are right and we added the direct comparison of the $\\alpha$ schedules on BBOB and HPOBench to directly compare the different behavior of SAWEI in Figure 5b.\n\n## Related Work\n\nThank you for pointing this out. You are right that we missed elaborations on Qin et al (enhancing EI) and Balandat et al. (BoTorch) and Volpp et al. (neural AFs) and we adjusted the text accordingly to include a short summary for both.\n\n## Different WEI Schedules and Tracking the Search Attitude\n\nWe understand that our former presentation of the algorithm can be misleading. We use standard WEI to propose the next configuration to evaluate. However, to determine the search attitude we modify one summand. This modification followed our intuition that the second summand is related to PI, and PI is more exploitative than (W)EI. Empirically, both methods perform on par (we added a figure in the appendix).\nWe updated the text accordingly.\nRegarding tracking the search attitude, we record the summands of WEI for the configuration proposed by WEI to be evaluated next.\nWe updated the section \"How to Adjust\" and \"SAWEI in a Nutshell\" to be more explicit on how the search attitude is calculated. Thank you for providing this feedback to improve our paper.\n\n## Effect of Switching in UBR\n\nThe effect of switching from EI to PI is expressed two folds in the UBR: We can see sharp downward bends in the log regret and sharp (often upward) bends in the UBR. This is nicely visible in Appendix Figure 20 (BBOB F13) or Figure 25 (BBOB F18). Same holds for HPOBench, but here log regret often stagnates after the switch. The magnitude of this effect depends on the problem at hand.\nWe updated the text to point to the Figure 3(d) and also added a small section in the Appendix showing this on more functions.\n\n## Ablation Study\n\nWe reworded the ablation section to make it easier to follow, and especially reworded L232-233 to reflect our motivation for the sensitivity analysis.\nIn particular, we can track the attitude up to the point where the incumbent changed the last time, i.e. until the point where we found a new best performing configuration. Finding a new incumbent is a desirable event in BO and thus lead to this setting.\n\n Global/general Response:\n```\nWe thank all reviewers for their thoughtful and valuable feedback. We address individual comments in direct replies to the reviews.\nAs a result of the feedback received we made several updates to the paper:\n\n- Added LCB as another baseline\n- Improved markers and colors for Figures 2, 3 (b,d), 4\n- Updated Figure 6 (Hyperparameter Sensitivity) (same scale, better font sizes)\n- Reworded ablation study\n- Refined description of search attitude\n- Moved Appendix section “How to Detect Adjustment Point” to “SAWEI in a Nutshell”, better describes usage of hyperparameter $\\epsilon$\n- Improved names for baselines, i.e. the modulated PI (PI) is now denoted as WEI($\\alpha=1$) to avoid confusion\n- Add short appendix section to show that UBR changes after changing optimizer settings (e.g. switching acquisition function)\n- Add Figure 5b and description to directly compare the different behavior of SAWEI on BBOB and HPOBench in terms of $\\alpha$\n- Rephrased beginning of “Related Work”\n\nWe marked the updated text with green color.\n```\n", "role": "user" }, { "content": "Title: Thank you for your response\nThank you for your response and the updates to the paper. Especially for the presentation of the search attitude which is now much clearer, and the related work which is much more informative. Figure 6 is also much better now. I still think some of the figure captions are too small, hopefully you can increase them in a camera-ready version. I still think the more clearly presented expression for alpha exploit (eq 4) should be the one you use.\n\nTypo: l148-149: You say alpha explore twice, one of them should be exploit.\n\nI have improved my score accordingly.", "role": "assistant" } ]
zaBJZmaSOf7	AutoML 2023 Conference	Reviewer_rDoP	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_rDoP, and the following are the paper you need to review:\n```<<zaBJZmaSOf7>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: In this work, the authors have introduced a new setting called \"ordered transfer HPO\", which is basically using or taking into consideration the \"order\" of tuning tasks during the transfer HPO process. The results obtained by the new approach OTHPO show to be promising and can outperform the standard transfer HPO methods. However, this new approach was not tested on a wide range of ML models and also needs to be tested on other datasets in the future.\noverall_review: The contribution in this paper gives a new and promising view on the tuning of hyperparameters optimization problems. The results show a competitive/ outperforming performance for the proposed approach, however, it still needs to be tested on more datasets and more ML models to study better its performance for AutoML in a general view.\nreview_summary: I recommend this paper to be accepted for AutoMLconference because it proposes an interesting and high-quality work but also brings an interesting approach to the field of HPO.\nactions_required_to_increase_overall_recommendation: Please address the recommendation mentioned previously.\nRating: 8 accept, good paper\nConfidence: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\n", "role": "assistant" }, { "content": "Title: Response to rDoP and Comment relating revised paper to review by rDoP\nThank you for the positive reviews and the suggestions. We will try to make our contribution clearer in the abstract and the variants of the methods better explained in the new revision. Do you have any concrete suggestions for improving Section 5.4, or what is unclear? The compute budget can be found in the appendix Section H. \nFor Section 5.2 we are adding introductory summaries for each of the methods.\n\n Further to our comment yesterday we have now uploaded a new paper with updates to the abstract, Section 5.2 and Section 5.4. Please let us know if anything is still unclear.\n\n Global/general Response:\n```\nWe’ve uploaded a revision of the paper. We’ve marked the changes in blue, and summarise them below:\n- Update to abstract making contributions clearer.\n- Update to Related work with additional references.\n- Update to Section 5.2, adding introductory sentences to each method and further improvements to the writing.\n- Update to Section 5.4 to make it clearer. \n- Addition of Practical recommendation paragraph at the end of the Conclusion. \n- Update to Submission checklist 4d, about the other assets listed in Appendix A.\n- Addition of experiments combining the SimpleOrdered approach with BO using density-ratio estimation for the surrogate model. These are presented in the new appendix G.7.\n- A small number of grammatical fixes.\n\n```\n", "role": "user" }, { "content": "Thanks to the authors for their clarification. It satisfactorily addresses the major issues I had. I continue to recommend this work positively.", "role": "assistant" } ]
zaBJZmaSOf7	AutoML 2023 Conference	Reviewer_9SXD	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_9SXD, and the following are the paper you need to review:\n```<<zaBJZmaSOf7>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper introduces Ordered Transfer Hyperparameter Optimization (OTHPO), a version of transfer learning for hyperparameter optimization (HPO) where tasks follow a sequential order. This paper proposes a formal definition, outlines the differences to related problems, and presents a basic OTHPO method that outperforms state-of-the-art transfer HPO. The authors investigate gradually accumulating data and empirically show the importance of taking order into account in ten benchmarks spanning XGBoost, random forest, approximate k-nearest neighbor, elastic net, support vector machines, and a separate real-world motivated optimization problem. The benchmarks are open-sourced to foster future research on ordered transfer HPO. This paper provides guidance for HPO practitioners in a deployed system.\noverall_review: Positive aspects of the paper:\n\nNovelty and practical relevance: The paper introduces the concept of ordered transfer hyperparameter optimization (OTHPO) that considers the sequential nature of hyperparameter tuning in practical scenarios where data is collected over time. This is a novel idea that has practical relevance in real-world machine learning systems.\n\nEmpirical evaluation: The authors provide empirical evidence to support their proposed OTHPO method by evaluating it on ten benchmarks spanning various machine learning methods such as XGBoost, support vector machines, random forest, approximate k-nearest neighbor, elastic net, and a real-world motivated optimization problem. They show that OTHPO outperforms state-of-the-art transfer HPO methods and highlight the importance of considering the order of tasks in hyperparameter optimization.\n\nNegative aspects of the paper:\n\nLimited Impacts: While the intuition is straightforward and reasonable, it doesn't provide many useful conclusions for practical usage. \n\nLack of theoretical analysis: Although the paper provides a formal definition of OTHPO and highlights its differences from related problems, such as standard transfer HPO and continual learning, it does not provide any theoretical analysis of the method. A theoretical analysis could help provide a deeper understanding of why and how the proposed method works, and under what conditions it is expected to perform well or poorly.\n\n\nreview_summary: The empirical evaluation validates the methods. A deeper analysis from the theoretical aspect and exploration of the conclusion would further improve the paper.\nactions_required_to_increase_overall_recommendation: See Overall Review\nRating: 5 marginally below the acceptance threshold\nConfidence: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper\n", "role": "assistant" }, { "content": "Title: Response to 9SXD and Comment relating revised paper to review by 9SXD\nWe would like to thank the reviewer for raising these points. Our learning from this work is that when the budget is low, running SimpleOrdered can already give good performance and we would suggest practitioners to start with this simple method before trying more complicated ones. This will be included in the updated draft which will be uploaded soon. It is aligned with the intuition from many practitioners but this has not been studied and reported to the best of our knowledge.\n\nWe agree with your point regarding the lack of theoretical analysis. That being said, we believe providing theoretical guarantees for transfer learning methods is a research problem on its own and will probably be out of scope for a single paper, in particular given that the first guarantees in this field are recent [1]. \n\n[1] On the Optimality Gap of Warm-Started Hyperparameter Optimization. Parikshit Ram Proceedings of the First International Conference on Automated Machine Learning, PMLR 188:12/1-14, 2022.\n\nIn the new paper version we have added our practical recommendation explicitly at the end of the Conclusion. \n Global/general Response:\n```\nWe’ve uploaded a revision of the paper. We’ve marked the changes in blue, and summarise them below:\n- Update to abstract making contributions clearer.\n- Update to Related work with additional references.\n- Update to Section 5.2, adding introductory sentences to each method and further improvements to the writing.\n- Update to Section 5.4 to make it clearer. \n- Addition of Practical recommendation paragraph at the end of the Conclusion. \n- Update to Submission checklist 4d, about the other assets listed in Appendix A.\n- Addition of experiments combining the SimpleOrdered approach with BO using density-ratio estimation for the surrogate model. These are presented in the new appendix G.7.\n- A small number of grammatical fixes.\n\n```\n", "role": "user" }, { "content": "Thanks for the authors' response. They've addressed nearly all of my concerns, and I may raise the score accordingly.", "role": "assistant" } ]
zaBJZmaSOf7	AutoML 2023 Conference	Reviewer_4oaY	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_4oaY, and the following are the paper you need to review:\n```<<zaBJZmaSOf7>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: In this paper, the authors tackle the issue of ordered hyperparameter optimization, which arises when there are changes in the underlying dataset or environmental variables, specifically data accumulation. \n\nTo address this issue, the authors propose a straightforward solution that involves using the top-performing hyperparameters from previous tasks, where the dataset was smaller than the current one, to either meta-initialize Bayesian Optimization (BO) on the new task or to use these hyperparameters as a starting point for BO. The effectiveness of this approach is evaluated on three benchmarks with a total of 10 datasets and several model classes.\n\nThe experimental results demonstrate that selecting the best-performing configurations as the data accumulates is an efficient method to obtain good results. However, the authors acknowledge that there is no universal method that is preferred after five evaluations, as it depends on the set of tasks being considered.\noverall_review: The paper has several notable strengths:\n\n- The authors tackle a practical use case and propose a straightforward solution that is effective and efficient. The proposed solution addresses the problem of hyperparameter optimization of a pre-defined model when the dataset changes over time. \n\n- The paper is well-written and presented in an easily understandable manner.\n\n- The results are well-visualized and straightforward to interpret. \n\nHowever, there are some negative aspects to the paper:\n\n- The proposed idea is very simplistic and does not yield conclusive results. The authors merely document an empirical reference that aligns with practitioners' initial intuitions of reusing top-performing configurations from previous tasks.\n\n- It is unclear how the hyperparameters of the first task are selected. If they are chosen randomly, then it is not clear why the performance of SimpleOrdered and SimpleOrderedShuffled are not the same at task 2 in Figure 5, especially since it reports the performance at the first iteration.\n\nAdditionally, there are some concerns:\n\n- Based on the revised problem definition, it is concerning that TransferBO, which extends the domain of the surrogate model along an additional sequence dimension, does not show reliable results against SimpleOrdered and its variants.\n\n- The authors mention that they reran the experiments for 50 seeds, which usually implies selecting new random configurations to initialize BO methods. However, the authors provide no further information about the source of randomness. A better way to validate the approach might be to introduce another source of randomness that affects the instances of the dataset at each task.\n\n- The evaluation protocol is limited to only 25 hyperparameters, which is lower than the conventional value. \n\n- Additionally, the baselines used in the paper are not very recent, and there are more up-to-date approaches available in the literature, such as those presented in references [1-2], that have demonstrated good transfer learning results.\n\n- Simpler baselines are also not evaluated, although addressed and referenced [3]. \n\nReferences: \n\n1. Wistuba, Martin, and Josif Grabocka. \"Few-Shot Bayesian Optimization with Deep Kernel Surrogates.\" International Conference on Learning Representations. 2021.\n\n2. Jomaa, Hadi Samer, et al. \"Transfer learning for bayesian hpo with end-to-end landmark meta-features.\" Fifth Workshop on Meta-Learning at the Conference on Neural Information Processing Systems. 2021.\n\n3. Feurer, M., Springenberg, J., and Hutter, F. (2015). Initializing Bayesian hyperparameter optimization 401\nvia meta-learning. In Proceedings of the AAAI Conference on Artifcial Intelligence, volume 29.\n\nreview_summary: The proposed problem setting is intriguing and casts transfer learning in a new light. However, the proposed approach is rather simplistic, and the authors do not provide conclusive recommendations. The idea of reusing the best hyperparameters, either from all previous tasks or only the immediate previous one, to initialize BO is aligned with the intuition of many practitioners.\nactions_required_to_increase_overall_recommendation: The proposed solution is compared to transfer learning solutions designed for another context, which is interesting but puts the approach at a disadvantage in terms of empirical findings. It would be valuable to evaluate the recommended approach, SimpleOrdered, against other meta-initialization techniques, such as those that leverage meta-feature similarity, and simply random initialization for GP. Additionally, SimpleOrdered can be coupled with several other surrogate models aside from GP to concretely demonstrate its superior performance.\nRating: 2 reject, significant issues present\nConfidence: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\n", "role": "assistant" }, { "content": "Title: Response to 4oaY\nFirst, we would like to thank the reviewer for their time. We have structured the response by quoting the review\n> like this\n\nand adding our responses like this below.\n\nResponse:\n\n>However, there are some negative aspects to the paper:\n>\n> - The proposed idea is very simplistic and does not yield conclusive results. The authors merely document an empirical reference that aligns with practitioners' initial intuitions of reusing top-performing configurations from previous tasks.\n\nIt would be helpful for us to understand which parts of the results seem inconclusive, could you elaborate? The main message of this paper is to introduce ordered transfer learning and to show that conceptually simpler methods, such as SimpleOrdered, actually perform competitively to more sophisticated approaches designed for standard transfer learning. \n\n> - It is unclear how the hyperparameters of the first task are selected. If they are chosen randomly, then it is not clear why the performance of SimpleOrdered and SimpleOrderedShuffled are not the same at task 2 in Figure 5, especially since it reports the performance at the first iteration.\n\nThe hyperparameter evaluations for the first task are optimised using Bayesian Optimisation (the baseline we call BO). Despite specifying the seeds there is some additional randomness in Syne Tune. We evaluate four hyperparameter configurations in parallel, and the optimisation trajectory is impacted by which hyperparameter evaluation finishes first. While the performance is not exactly the same for SimpleOrdered and SimpleOrderedShuffled in Figure 5 for task 2, they are close and within the 2 standard errors shown, and so are as we would expect them to be.\n\n>Additionally, there are some concerns:\n> - Based on the revised problem definition, it is concerning that TransferBO, which extends the domain of the surrogate model along an additional sequence dimension, does not show reliable results against SimpleOrdered and its variants.\n\nWe would like to thank the reviewer for raising this point. We agree that TransferBO intuitively makes sense (that is the main reason we take it as a baseline) and it indeed performs better than the non-transfer learning methods in our results. The interesting insight from our results is that the simple methods beat TransferBO. We don’t see why it has to be better than SimpleOrdered, especially in our setting where the sequence of the tasks are ordered: SimpleOrdered is getting only very relevant information, while TransferBO needs to separate that out from the less relevant evaluations of early tasks.\n\nIn addition, we want to restate that our goal is to show that considering the order is important in the setting and even simple methods such as SimpleOrdered could already lead to better results than other transfer learning baselines, which aligns with “practitioners' initial intuitions”, as mentioned already by the reviewer.\n\n\n> - The authors mention that they reran the experiments for 50 seeds, which usually implies selecting new random configurations to initialize BO methods. However, the authors provide no further information about the source of randomness. A better way to validate the approach might be to introduce another source of randomness that affects the instances of the dataset at each task.\n\nWe kindly ask the reviewer to clarify this suggestion. Does the reviewer mean to have a different train/validation split for every seed? Why is that a better approach to validate a method? Fixing train/test split is a common setting in many HPO benchmarks. For YAHPO, we rely on already defined train/test splits, so this is not possible to change. It accounts for eight of our ten benchmarks, and we want to keep our evaluation consistent between benchmarks. Therefore we use the same split even for the other benchmarks \n\nIn our experiments, the randomness includes the seed specified to Syne Tune, which we set to 50 different values. There is some extra randomness that we don’t control within Syne Tune. We evaluate 4 hyperparameter evaluations in parallel. When one finishes we start another one before the others finish. Which evaluation finishes first therefore impacts the choice of where to sample next and introduces randomness. \n\n> - The evaluation protocol is limited to only 25 hyperparameters, which is lower than the conventional value.\n\nRunning 25 hyperparameters may seem short in general. In our experiments however, many of them actually converge already after 25 iterations, please see Figure 16 in the appendix for some examples.\n\n> - Additionally, the baselines used in the paper are not very recent, and there are more up-to-date approaches available in the literature, such as those presented in references [1-2], that have demonstrated good transfer learning results.\n> - Simpler baselines are also not evaluated, although addressed and referenced [3].\n> \n> References:\n> 1. Wistuba, Martin, and Josif Grabocka. \"Few-Shot Bayesian Optimization with Deep Kernel Surrogates.\" International Conference on Learning Representations. 2021.\n>\n> 2. Jomaa, Hadi Samer, et al. \"Transfer learning for bayesian hpo with end-to-end landmark meta-features.\" Fifth Workshop on Meta-Learning at the Conference on Neural Information Processing Systems. 2021.\n>\n> 3. Feurer, M., Springenberg, J., and Hutter, F. (2015). Initializing Bayesian hyperparameter optimization 401 via meta-learning. In Proceedings of the AAAI Conference on Artifcial Intelligence, volume 29.\n\nWe would like to thank the reviewer for the additional references. We will include them in the related work section. Notice that in [3], the method is to find the most similar datasets and take the best HP configurations from those datasets to initialize BO for a new dataset, which is almost identical to our approach SimpleOrdered if we take the dataset sizes as meta-feature. In addition, we believe the selected baselines in the paper are still relatively recent from 2020 and 2019. We will add a new baseline where we combine SimpleOrdered with BORE [4], as required by the reviewer who asked for SimpleOrdered to be coupled with other surrogate models. \n\n[4] Tiao, Louis C., et al. \"BORE: Bayesian optimization by density-ratio estimation.\" International Conference on Machine Learning. PMLR, 2021.\n\n> Actions Required To Increase Overall Recommendation:\nThe proposed solution is compared to transfer learning solutions designed for another context, which is interesting but puts the approach at a disadvantage in terms of empirical findings. It would be valuable to evaluate the recommended approach, SimpleOrdered, against other meta-initialization techniques, such as those that leverage meta-feature similarity, and simply random initialization for GP. Additionally, SimpleOrdered can be coupled with several other surrogate models aside from GP to concretely demonstrate its superior performance.\n\nWe already compare to random initialisation for GP; this is the baseline called BO. Notice that in [3], the method is to find the most similar datasets and take the best HP configurations from those datasets to initialize BO for a new dataset, which is almost identical to our approach SimpleOrdered if we take the dataset sizes as meta-feature. We are running a new baseline where we combine SimpleOrdered with BORE [4] and will include these results in the next revision of the paper.\nWe have added the extra references mentioned into the Related work section in the new paper version. We have also added appendix G.7, which has results for combining the SimpleOrdered approach with a new surrogate model: density-ratio estimation.\n\n Global/general Response:\n```\nWe’ve uploaded a revision of the paper. We’ve marked the changes in blue, and summarise them below:\n- Update to abstract making contributions clearer.\n- Update to Related work with additional references.\n- Update to Section 5.2, adding introductory sentences to each method and further improvements to the writing.\n- Update to Section 5.4 to make it clearer. \n- Addition of Practical recommendation paragraph at the end of the Conclusion. \n- Update to Submission checklist 4d, about the other assets listed in Appendix A.\n- Addition of experiments combining the SimpleOrdered approach with BO using density-ratio estimation for the surrogate model. These are presented in the new appendix G.7.\n- A small number of grammatical fixes.\n\n```\n", "role": "user" }, { "content": "Thank you for the authors' reply. They've addressed nearly all of my concerns, and I'm open to increasing the score.", "role": "assistant" } ]
zvV7hemQmtLl	AutoML 2023 Conference	Reviewer_JBRh	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_JBRh, and the following are the paper you need to review:\n```<<zvV7hemQmtLl>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper introduces two new population-based ensemble selection methods in post hoc optimization(QO-ES and QDO-ES) for AutoML, a version of ensemble selection where the search of optimism is not based on greedy search but based on population optimization. The authors present a background of ensemble learning and the motivations to pursue this work on using population-based methods instead of greedy methods. The new approach was explained in section 4 and provides guidance for each step of the proposed method.\noverall_review: The paper has several notable strengths:\n The authors tackle a practical use case and propose a straightforward solution that is effective and efficient. The proposed solution addresses the problem of post hoc ensemble selection methods in AutoML.\n The paper is well-written and in an easily understandable manner.\n The results are well-visualized and straightforward to interpret.\n\nHowever, there are some negative aspects to the paper:\n The proposed approach is not well presented\n It is not very clear how the method could influence on the decision making part in AutoML\n The results representation /explanation is a bit complicated to follow and need to be improved.\n The approach need to be tested on more benchmarks to make intuitive conclusions.\nreview_summary: The suggested problem setting is intriguing and offers a fresh perspective on post hoc ensemble selection. However, the suggested strategy is rather straightforward, and the authors offer no firm recommendations. replacing a greedy search with a population-based method to predict higher performance but still not conclusive to make better decisions for AutoML.\nactions_required_to_increase_overall_recommendation: See overall review and adress comments.\nThe paper needs to be reorganized to show better the potential of the proposed approach.\nRating: 5 marginally below the acceptance threshold\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer JBRh and Changes Addressing Your Concerns\nDear Reviewer JBRh,\n\nThank you for your feedback. \n\nFollowing your remarks, we are working on ways to improve the clarity of Section 4. However, to allow for discussions, we decided to answer the rest of your comments sooner rather than later\n\nYou mentioned that our approach should be tested on more benchmarks.\nWe compared our method on 71 datasets (10-fold cross-validation) across two metrics following the procedure of the AutoML benchmark, which is currently setting the standard for best practices in AutoML. To the best of our knowledge, no related work for ensemble selection or post hoc ensembling has ever done such an extensive comparison. Note that the estimated wall clock time of our current experiments is approximately 3.9 years (we provide details on this number in the Appendix of the updated manuscript). We think that adding more experiments would hinder the reproducibility of our results without adding substantial value to the paper. \n\nYou state:\n> “It is not very clear how the method could influence on the decision making part in AutoML”. \n\nSadly, it is slightly unclear to us what exactly you are referring to here. Our methods do not directly influence the decision making part of AutoML (i.e., CASH) but are a post-processing step, i.e., post-hoc ensemble selection. Could you please elaborate on this such that we can properly reply? \n\nMoreover, you state:\n> “However, the potential impact of the solution is low but still offers meaningful insights.”\n\nIn the introduction, we listed many AutoML systems using post hoc ensembling, specifically greedy ensemble selection (GES). Because our method improves over GES, it can potentially impact most AutoML systems currently in use.\n\nDear Reviewer JBRh,\n\nOur updated manuscript contains several changes to improve clarity and notational quality. \nAdditionally, the Appendix now includes several formalisations of components of our approach, which aim to resolve the unclear aspects of Section 4. Thank you for pointing this out to us. \n\nMoreover, we formulate an ablation study to provide conclusive data to make better decisions for AutoML. \n\nWe believe these changes address your concerns w.r.t. clarity and conclusiveness and hope you consider increasing your rating based on the improved manuscript.\n\nFor a detailed list of all improvements, see our comment “The Updated Manuscript is Available” above.\n\n Global/general Response:\n```\nDear Reviewers,\n\nWe updated the manuscript taking into account your valuable feedback. All changes to the manuscript are coloured in green.\n\nIn detail, we changed the following major points:\n\n* We estimated the wall-clock time of our experiments, stated this in the result section, and provided details on the estimation in Appendix F.1. \n* In the result section, we expanded our discussion on ensemble efficiency and provided an overview in Table 1. \n* We highlight that performance differences are only statistically significant on validation data, although Q(D)O-ES often outrank GES on both validation and test data. We further include validation results in Figure 3, showing that our approaches strongly outperform GES on validation data. The claim that our approaches perform well is further backed up by the performance overviews per dataset in Appendix H.\n* We mention the algorithmic diversity created by SiloTopN explicitly in the result section and point to Table 2 in the Appendix showing the algorithmic diversity per dataset. \n* We provide formalisations of the diversity measures in Appendix D.1 and formalisations of the initialisation approaches in Appendix D.7. \n* We touch up ablation studies w.r.t. components of Q(D)O-ES and the configurations of these components in the result section. These ablation studies are detailed in Appendix F.3 and F.4. \n* EDIT-1: We addressed the additional comments of reviewer 7tmR regarding the new formal definitions and Figure 1. \n\nAdditionally, we changed other minor things to improve clarity and notational quality. Moreover, we add all new code related to these changes to Anonymous GitHub. \n\nWe believe these changes address the concerns raised by you. Therefore, we politely ask you to verify if your concerns have been resolved and reconsider your rating.\n\nThank you for your time and valuable feedback. We are happy to participate in further discussions and willing to resolve additional concerns.\n\n```\n", "role": "user" }, { "content": "I'm grateful for the authors' response. They have adequately addressed the majority of my concerns, and I am considering increasing the score.", "role": "assistant" } ]
zvV7hemQmtLl	AutoML 2023 Conference	Reviewer_e8Qx	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_e8Qx, and the following are the paper you need to review:\n```<<zvV7hemQmtLl>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper studies post hoc ensemble election based on a population-based quality and diversity optimization for AutoML. The experimental results compared against greedy ensemble selection. The study found that diversity is one of the key factors to achieve a better performance for ensemble models. \noverall_review: It is an interesting paper for the AutoML community, but further discussion and explanation of the evaluation and the results are needed to provide a comprehensive understanding.\nreview_summary: The paper is well-written and easy to follow but lacks detail on experiment setup and evaluations that may impact comprehensiveness of the manuscript's findings.\nactions_required_to_increase_overall_recommendation: Answering the above questions and revising the paper accordingly would be beneficial for the AutoML community.\nRating: 8 accept, good paper\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer e8Qx and Changes Addressing Your Concerns\nDear Reviewer e8Qx,\n\nThank you for your feedback. As we want to start a potential discussion early, given the short timeframe of the rebuttal, we are sharing our comments now and are stating below where we plan to adjust the manuscript. The updated manuscript will follow by the end of the week.\n\nIn your review, you state:\n> “Although the proposed approach and recommendations are not a novel idea (see Nam et al. [1]),”.\n\nIt is true that the idea of using diversity in ensembling is not novel. However, to the best of our knowledge, we are the first to propose and evaluate population-based algorithms for greedy ensemble selection and AutoML, as discussed in the related work section. \n\nRegarding the paper of Nam et al. [1], we believe that it is not directly related to our work. In contrast, the papers on diversity mentioned in our introduction (Lines 45-57) are more closely related papers. Nam et al. [1] discuss a model distillation approach motivated by ensemble diversity literature. Model distillation aggregates model parameters or behavior to build one model. In contrast, ensembles aggregate the predictions of multiple models. In our work, we talk about a post hoc ensembling method motivated by ensemble diversity. \n\n> “In Line 227, the authors stated [...] but it raises several important questions which are not answered in this study, such as: what is the cost of this improvement? How much additional time is required per dataset to achieve this performance gain? (There is only a brief description in Line 321). Adding ensemble models are expensive compared to a single model How does the proposed approach handle large datasets?.” \n\nWe assume you are referring to Line 277. We agree that this raises the questions you have mentioned, but most prominent AutoML systems use post hoc ensembling, and it is not the primary goal of this study to evaluate how much cost is induced by it. However, we show that our method does not incur significantly higher costs than the state-of-the-art method GES.\nMoreover, as our approaches function similarly efficiently to GES (Line 321) and handles large datasets similarly well to GES, answering questions related to the general efficiency of post hoc ensembling was not a goal of our paper. We note that post hoc ensembling methods are in general efficient even for larger datasets, as AutoML systems store the validation predictions [2]. Multiple large datasets are part of the AutoML benchmark datasets, see Table 1 in the Appendix. Regarding this point and your later point “Furthermore, it would be beneficial to have fine-grained evaluation results per dataset for a more comprehensive analysis.” – our code provides the results per dataset, but these were not in the manuscript. We will provide an overview of performance per dataset in the updated manuscript. \n\n> “My understanding of the current experimental setup is that model selection applies to only 10% of data for each dataset in each iteration, which makes it feasible to have diverse ensemble algorithms.”\n\nIn post hoc ensemble selection, CASH is not performed. Instead, post hoc ensemble selection is concerned with constructing an ensemble given a set of models. In our experiments, Auto-Sklearn’s model selection is performed on all data (using a hold-out split for validation). Our ensemble methods are then trained on the validation data from the hold-out split. By default and in our experiments, this validation data corresponds to 33% of the training data. We noticed that we only stated this in Line 295. For clarity, we will state this earlier in Section 5 in the updated manuscript.\n\n> “In addition, the authors emphasize that their report is based on a 10-fold cross-validation, which is not clear whether the performance is based on average of selected ensemble model in each iteration or if it is based on the performance of selected ensemble on the whole dataset.”\n\nOur evaluation protocol follows the AutoML benchmark. Consequently, the used datasets were split into 10 folds for cross-validation. For each split, we compute the score of each ensemble method and then average these fold scores. A method's fold score is the final ensemble's performance for the whole dataset. We did not evaluate the performance over time or iterations. \n\n \n\nBesides our comments above, we will update the explanations of the experimental setup and evaluations to increase clarity in the updated manuscript. Thank you again for your valuable feedback.\n\nReferences\n\n[1] Nam, Giung, et al. \"Diversity matters when learning from ensembles.\" Advances in neural information processing systems 34 (2021): 8367-8377.\n\n[2] Feurer, M., Klein, A., Eggensperger, K., Springenberg, J., Blum, M., & Hutter, F. (2015). Efficient and Robust Automated Machine Learning. Advances in Neural Information Processing Systems, 28. \n\nDear Reviewer e8Qx,\n\nWe have added several improvements to the manuscript to address your concerns. For a detailed list of all improvements, see our comment “The Updated Manuscript is Available” above.\n\nWe now provide more details on the efficiency of all compared approaches in the result section. Specifically, stating the additional time added on average on top by running the methods for $50$ iterations. Moreover, we clarified what data is used during model selection in Section 5. Furthermore, we provided fine-grained evaluation results per dataset in Appendix H and made our usage of UK English consistent. Thank you for raising these concerns in your review.\n\nWe believe these changes address your concerns and hope you consider increasing your rating based on the improved manuscript.\n\n Global/general Response:\n```\nDear Reviewers,\n\nWe updated the manuscript taking into account your valuable feedback. All changes to the manuscript are coloured in green.\n\nIn detail, we changed the following major points:\n\n* We estimated the wall-clock time of our experiments, stated this in the result section, and provided details on the estimation in Appendix F.1. \n* In the result section, we expanded our discussion on ensemble efficiency and provided an overview in Table 1. \n* We highlight that performance differences are only statistically significant on validation data, although Q(D)O-ES often outrank GES on both validation and test data. We further include validation results in Figure 3, showing that our approaches strongly outperform GES on validation data. The claim that our approaches perform well is further backed up by the performance overviews per dataset in Appendix H.\n* We mention the algorithmic diversity created by SiloTopN explicitly in the result section and point to Table 2 in the Appendix showing the algorithmic diversity per dataset. \n* We provide formalisations of the diversity measures in Appendix D.1 and formalisations of the initialisation approaches in Appendix D.7. \n* We touch up ablation studies w.r.t. components of Q(D)O-ES and the configurations of these components in the result section. These ablation studies are detailed in Appendix F.3 and F.4. \n* EDIT-1: We addressed the additional comments of reviewer 7tmR regarding the new formal definitions and Figure 1. \n\nAdditionally, we changed other minor things to improve clarity and notational quality. Moreover, we add all new code related to these changes to Anonymous GitHub. \n\nWe believe these changes address the concerns raised by you. Therefore, we politely ask you to verify if your concerns have been resolved and reconsider your rating.\n\nThank you for your time and valuable feedback. We are happy to participate in further discussions and willing to resolve additional concerns.\n\n```\n", "role": "user" }, { "content": "Title: Revised Version\nI thank the authors for revising the manuscript to address my concerns. Although not all major issues have been resolved, in my opinion, the paper's presentation and experiments are more convincing. I have adjusted my rating accordingly.", "role": "assistant" } ]
zvV7hemQmtLl	AutoML 2023 Conference	Reviewer_wmVj	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_wmVj, and the following are the paper you need to review:\n```<<zvV7hemQmtLl>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: 1. Using Quality-Diversity (QD) algorithms for Ensemble Selection (ES) problems. \n\n2. Experimental results show the advantages of QDO-ES and QO-ES. \noverall_review: Pros\n1. The application of QD algorithms into ensemble selection is interesting and promising.\n\nCons\n\n1. The experiment part is not solid enough. Please refer to \"Actions Required To Increase Overall Recommendation\".\nreview_summary: This paper apply QD algorithms into ensemble selection problems. However, the experiment part should be further improved.\nactions_required_to_increase_overall_recommendation: 1. Abalation studies, including the operators and the hyper-parameters.\n\n2. More types of experiments are needed.\n\n3. The verification (or visualization) of the diversity of obtained solutions should be discussed.\n\n4. More related methods should be compared, such as [1] \n\n[1] DivBO: Diversity-aware CASH for Ensemble Learning.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper\n", "role": "assistant" }, { "content": "Title: Response to Reviewer wmVj\nDear Reviewer wmVj,\n\nThank you for your feedback. As we want to start a potential discussion early, given the short timeframe of the rebuttal, we are sharing our comments now and are stating below where we plan to adjust the manuscript. The updated manuscript will follow by the end of the week.\n\nIn your review, you state: \n> “This paper apply current QD algorithms into the ensemble selection problems to obtain a set of high-performing and diverse solutions..”\n\nWe would like to provide additional clarification regarding this statement to avoid further confusion. While we are applying a QD-like algorithm with QDO-ES, QO-ES is not a QD algorithm. Our work in general focuses on population-based search, which is not only restricted to QD algorithms. In this sense, we constructed another competitor without making use of the concepts of QD, which is given by the QO-ES method.\n\n> “However, the lacking of compared methods (except for greedy and single best), and the lacking of ablation studies weaken the quality” \n\nWe compared our approach to ensemble selection to the only other available and widely used method in AutoML, greedy ensemble selection (GES). We do not compare our approach to ensemble weighting methods (like stacking) since we are interested in efficient, i.e., sparse, ensembles, see background (Lines 98-100). Additionally, the contribution of this paper is an improvement of GES by extending it to a population-based search. Therefore, we believe comparing our methods only to GES is appropriate. Nevertheless, we agree that the current literature lacks a general comparison of post hoc ensembling methods; we believe this to be a promising avenue for future work.\n\nWe would be grateful if you could provide us with references to other post hoc ensembling methods for AutoML that you would like us to compare. The referenced method that you provide in your review, DivBO [1], is not directly applicable to post hoc ensembling (which is concerned with constructing an ensemble itself given a set of models) but instead tackles the problem of how to efficiently find models in a CASH fashion to then construct an ensemble.\n\nIn detail, DivBO [1] describes an approach to introduce diversity during CASH. However, our methods look at diversity during post hoc ensembling. The method discussed in [1] is comparable to the papers we have mentioned in the related work (Lines 116-120), which is very different from our approach.\n\n \n\nYou mention in your review that the paper is well organised but rated it as “2: The work is not presented in a clear way in most parts.”. We would be willing to address any more concerns you have about the clarity.\n\nIn “Actions Required”, you stated:\n> “More types of experiments are needed.” \n\nWe compared our method on 71 datasets (10-fold cross-validation) across two metrics following the procedure of the AutoML benchmark, which is currently setting the standard for best practices in AutoML. To the best of our knowledge, no related work for ensemble selection or post hoc ensembling has ever done such an extensive comparison. Note that the estimated wall clock time of our current experiments is 3.9 years (we provide details on this in the Appendix of the updated manuscript). We think that adding more experiments would hinder the reproducibility of our results without adding substantial value to the paper. \n\n \n\nRegarding your comment, “Abalation studies, including the operators and the hyper-parameters.”. Thank you for the feedback, we note that we have the performance values for all tested configurations. We only used this for analyzing hyperparameter importance in the paper, but will include additional ablation studies in the updated manuscript. \n\n> “The verification (or visualization) of the diversity of obtained solutions should be discussed.” \n\nWe agree that such an analysis would be interesting in principle. However, the goal of this work was not to foster diversity in the final solution but to achieve peak performance with the final solutions. In other words, our aim was to obtain high-performing solutions and not solutions that are diverse, as stated in the introduction. We admit our application of QDO is atypical; nevertheless, having access to a diverse population during optimisation can improve performance, even if one is only interested in a single best solution, as this was our main motivation.\nWe provide a glimpse of the algorithmic “diversity” of base models in relation to the pruning methods TopN and SiloTopN in Appendix E.1, Table 1. Specifically, the columns regarding the mean number of distinct algorithms show that SiloTopN produces a much more algorithmically diverse set of base models than TopN. However, we did not reference this explicitly in Lines 306-308, which will be corrected in the updated manuscript\n\n \n\nThank you again for your time and valuable feedback!\n\n \n\nReferences\n\n[1] DivBO: Diversity-aware CASH for Ensemble Learning.\n\nDear Reviewer wmVj,\n\nWe have added several improvements to the manuscript to address your concerns. For a detailed list of all improvements, see our comment “The Updated Manuscript is Available” above. Thank you for your feedback allowing us to improve our manuscript.\n\nWe provided details on ablation studies w.r.t. the hyperparameters and their values (e.g., crossover operators) in Appendix F.3 and F.4; showing the importance of different components of our approach. Moreover, we now elaborate on the algorithmic diversity created by SiloTopN and point to values showing this algorithmic diversity per dataset in the appendix. \n\nWe believe these changes address your concerns and hope you consider increasing your rating based on the improved manuscript.\n\n Global/general Response:\n```\nDear Reviewers,\n\nWe updated the manuscript taking into account your valuable feedback. All changes to the manuscript are coloured in green.\n\nIn detail, we changed the following major points:\n\n* We estimated the wall-clock time of our experiments, stated this in the result section, and provided details on the estimation in Appendix F.1. \n* In the result section, we expanded our discussion on ensemble efficiency and provided an overview in Table 1. \n* We highlight that performance differences are only statistically significant on validation data, although Q(D)O-ES often outrank GES on both validation and test data. We further include validation results in Figure 3, showing that our approaches strongly outperform GES on validation data. The claim that our approaches perform well is further backed up by the performance overviews per dataset in Appendix H.\n* We mention the algorithmic diversity created by SiloTopN explicitly in the result section and point to Table 2 in the Appendix showing the algorithmic diversity per dataset. \n* We provide formalisations of the diversity measures in Appendix D.1 and formalisations of the initialisation approaches in Appendix D.7. \n* We touch up ablation studies w.r.t. components of Q(D)O-ES and the configurations of these components in the result section. These ablation studies are detailed in Appendix F.3 and F.4. \n* EDIT-1: We addressed the additional comments of reviewer 7tmR regarding the new formal definitions and Figure 1. \n\nAdditionally, we changed other minor things to improve clarity and notational quality. Moreover, we add all new code related to these changes to Anonymous GitHub. \n\nWe believe these changes address the concerns raised by you. Therefore, we politely ask you to verify if your concerns have been resolved and reconsider your rating.\n\nThank you for your time and valuable feedback. We are happy to participate in further discussions and willing to resolve additional concerns.\n\n```\n", "role": "user" }, { "content": "Title: Responses to authors\nThanks for your replies. I think the revised manuscript is better now and I will increase my clarity and overall score. \n", "role": "assistant" } ]
zvV7hemQmtLl	AutoML 2023 Conference	Reviewer_7tmR	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_7tmR, and the following are the paper you need to review:\n```<<zvV7hemQmtLl>>```", "role": "user" }, { "content": "Review Content: summary_of_contributions: This paper proposes two population based methods for post hoc ensemble selection in AutoML: QO-ES (quality optimization) and QDO-ES (quality diversity optimization). The general idea is to investigate whether the behavioural diversity of a population of ensembles may be beneficial for the performance. The methods are compared to greedy ensemble selection (GES) as well as the single best model. The authors claim that QO-ES and QDO-ES often outperform GES, as well as that behavioural diversity can be beneficial for post hoc ensembling. \n\noverall_review: ## Positive\n\n- Structure of the paper\n- Most descriptions of methods and evaluation protocols\n- Interesting idea of behavioural diversity\n\n## Negative (see comments in other fields)\n\n- Results and conclusions thereof are oversold\n- Mathematical presentation could be improved\n\nreview_summary: The paper proposes an interesting alternative to GES, although it is oversold in my opinion. It is well structured and described in most parts of the paper. Nevertheless, especially the results and discussions thereof could be improved and clarified. Additionally, I was wondering why only a single AutoML tool has been used for evaluation. Furthermore, in the beginning of the paper, I assumed that diversity would play a major role in this paper and would be further analyzed, but it turned out diversity was not so present.\nactions_required_to_increase_overall_recommendation: All in all, the negative aspects described under correctness and clarity need to be addressed to convince me to raise my score.\nRating: 8 accept, good paper\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer 7tmR\n\nDear Reviewer 7tmR,\n\nThank you for your feedback. As we want to start a potential discussion early, given the short timeframe of the rebuttal, we are sharing our comments now and are stating below where we plan to adjust the manuscript. The updated manuscript will follow by the end of the week.\n\nIn your review, you stated:\n> “The authors have bold statements about the performance of their methods compared to GES, but none of the three ensembling methods investigated is significantly better than another on test data. Only on validation data, a significant performance improvement of Q(D)O-ES can be detected over GES, which is a sign of overfitting. “\n\nWe do agree that some of our statements, especially in the abstract, may be considered exaggerated, and we will fix this in the updated manuscript. Thank you very much for bringing this to our attention. However, in the result section, we believe that we provide a fair summary and comparison of the performance of the different methods. We also want to note that comparing only four algorithms via Nemenyi Post hoc tests results in requiring very strong performance differences (and therefore very stable mean ranks) to result in significant differences. Looking at the box plots, we can observe that Q(D)O-ES usually results in better relative performance in 3 out of the 4 scenarios (especially QO-ES showing strong relative performance)\n\nWe also agree that the significant performance improvement on validation data is a sign of overfitting, as discussed in the result section. However, it is also a sign that Q(D)O-ES can better solve the problem based on the given data (i.e., on validation data) than GES, despite the solution’s performance not generalising to the test data. As the poor generalisation likely follows from Auto-sklearn’s validation method (33% hold-out split), we do believe the significant performance differences on validation data are important. \n\n> “Therefore, the performance of Q(D)O-ES is oversold. [...] Still, the authors claim that Q(D)O-ES often outperforms GES, which is not true.”\n\nThe statement about ‘outperforms’ was related to mean rank and in relation to better relative performance, not in relation to significance. Since the mean rank of our methods is often better, it ‘outperforms’ GES, albeit not significantly. Especially considering the difference in relative performance (Figure 2). Our statement holds true but seems to have been misleading; we will add the point about non-significance in the abstract and clarify our statement in the introduction to avoid such confusion.\n\n> “The three initialization approaches (lines 204 - 207) should all be defined mathematically.“\n\nWe will add formalized versions of the initialization method to the Appendix, likely in terms of pseudo-code, mirroring the current implementation. \n\n> “Furthermore, the definition of the first ensemble consisting only of a base model (line 206) is wrong, as it would result in an empty set.”\n\nPlease note that the formula states that it is the set of multisets such that there exists one and only one element of $r$ that is $1$. As such, it can not result in an empty set but will always correspond to a set of one base model. Here, $\\exists!$ means \"exists exactly one\" (e.g., see https://en.wikipedia.org/wiki/Uniqueness_quantification or Chapter 1 in [1]). \n\n> “Within Algorithm 2, line 7, the remark does not make sense to me as it is the same as on the left hand side. “\n\nThe operation “<-” means assignment; while the mathematical operator “=” means equal. In other words, the right-hand side states that the output of `mutate()` is equal to the input $r_{sol}$. While the left-hand side only states that the output of `mutate()` is assigned as the new value of $r_{sol}$. \n\n> “Furthermore, an ablation study about which parts and parameters most affect the performance would be interesting. The mentioned method for assessing hyperparameter importance (Appendix F2) should be elaborated. Moreover, it is not distinguished, e.g., which initialization method has been used, and how important it is.”\n\nWe will elaborate on the hyperparameter importance in the updated manuscript and provide an additional ablation study w.r.t. the values of hyperparameters. Thank you for pointing out the limitation of our current approach. \n\n \n\nReferences:\n\n[1] Howson, A. G. (1972). A handbook of terms used in algebra and analysis. Cambridge University Press.\n> “Lastly, the overall used CPU hours to perform the final evaluation of this paper would be of interest.“\n\nWe did not state any time-related number w.r.t. our experiments in the manuscript, while it would indeed be beneficial. We will fix this in the updated manuscript. We estimated the wall clock time to be approximately 3.9 years (summed over all datasets, folds, metrics, and configurations; ignoring the parallelization of the ensemble method; we will provide details on this number in the Appendix of the updated manuscript). We did not capture the exact CPU time and cannot calculate this exactly in retrospect due to parallelization. \n\n\n> “In general, I was wondering why the experiments have not been repeated for different seeds.”\n\nWe have used multiple folds (10 per dataset) following the AutoML benchmark. As the AutoML benchmark did not use seeds, we also did not look at different random seeds per fold. Moreover, considering the number of datasets and randomness induced by resampling via cross-validation, we believe our experiments already cover a good amount of randomisation.\n\nFurthermore, we initially deemed it too expensive to include seeds due to the large number of configurations. In retrospect, considering the estimated wall clock time of approximately 3.9 years, we believe this to be the case indeed.\n\n> “Within the evaluation (around lines 235 and 250) it is not clear to what extent AutoSklearn und AutoGluon have been used. Overall, captions should contain a short description of the experiment it belongs to.” \n\nWe used Auto-Sklearn for the base models and only a pruning method from AutoGluon. This is a hyperparameter, as stated in Line 251. None of our experiments are specific to a pruning method. Like all other hyperparameters, the pruning method is selected by LOO CV and is not specific to certain experiments. Hence, we can not provide such descriptions in the captions. We will highlight this as part of the ablation study in the updated manuscript to further clarify this. \n\n \n\nWe will try to resolve the remaining points you raised w.r.t. clarity in the updated manuscript. Thank you for the valuable and detailed feedback! \nDear Reviewer 7tmR,\n\nWe have added several improvements to the manuscript to address your concerns. For a detailed list of all improvements, see our comment “The Updated Manuscript is Available” above.\n\nBased on your valuable and comprehensive feedback:\n\n* We adapted our somewhat exaggerated statements in the abstract and introduction that were overselling our contributions. \n* We improved the mathematical presentation for several components of our approach. Specifically, we formalised the diversity measures in Appendix D.1, and the initialisation approaches in Appendix D.7. Moreover, we were able to improve the overall mathematical notation for several aspects thanks to your feedback. \n* Likewise, we make sure to shortly explain something before pointing to the appendix for more details and moved the results for validation data from the appendix to the main paper. \n* Finally, we provided an estimate of the wall-clock time. Our calculations of this lower bound of the CPU time are explained in Appendix F.1. \n\nWe believe these changes address your concerns and hope you consider increasing your rating based on the improved manuscript. Thank you again for your feedback. \n\n Global/general Response:\n```\nDear Reviewers,\n\nWe updated the manuscript taking into account your valuable feedback. All changes to the manuscript are coloured in green.\n\nIn detail, we changed the following major points:\n\n* We estimated the wall-clock time of our experiments, stated this in the result section, and provided details on the estimation in Appendix F.1. \n* In the result section, we expanded our discussion on ensemble efficiency and provided an overview in Table 1. \n* We highlight that performance differences are only statistically significant on validation data, although Q(D)O-ES often outrank GES on both validation and test data. We further include validation results in Figure 3, showing that our approaches strongly outperform GES on validation data. The claim that our approaches perform well is further backed up by the performance overviews per dataset in Appendix H.\n* We mention the algorithmic diversity created by SiloTopN explicitly in the result section and point to Table 2 in the Appendix showing the algorithmic diversity per dataset. \n* We provide formalisations of the diversity measures in Appendix D.1 and formalisations of the initialisation approaches in Appendix D.7. \n* We touch up ablation studies w.r.t. components of Q(D)O-ES and the configurations of these components in the result section. These ablation studies are detailed in Appendix F.3 and F.4. \n* EDIT-1: We addressed the additional comments of reviewer 7tmR regarding the new formal definitions and Figure 1. \n\nAdditionally, we changed other minor things to improve clarity and notational quality. Moreover, we add all new code related to these changes to Anonymous GitHub. \n\nWe believe these changes address the concerns raised by you. Therefore, we politely ask you to verify if your concerns have been resolved and reconsider your rating.\n\nThank you for your time and valuable feedback. We are happy to participate in further discussions and willing to resolve additional concerns.\n\n```\n", "role": "user" }, { "content": "Title: Reviewer Response to Author Response\nDear Authors, \n\nthanks for the detailed response to my review and the adapted paper with highlighted changes. Most of my comments have been sufficiently addressed, except the following:\n\nFigure 1: Critical Difference Plots\n\nWithin the caption the used dataset is missing (as it has been added to Figure 2)\n\nAppendix D.7 Initialisation Approaches\n\n* I am still quite sure that the formal definition of $S_1$ is not correct, let me elaborate: You only check $r(i)=1$, but you do not check that the others are $r(i')=0$. My suggestions are the following, alternatively you could check the sum of $r$: \n * $$ S_1 = \\\\{ (P,r) \| ∃!𝑖 ∈ [m] : 𝑟(𝑖) = 1 \\wedge \\forall j \\in [m], j \\neq i: r(i)=0 \\\\} $$\n * $S_1 = \\\\{ (P,r) \| \\sum\\limits_{i =1}^{m} r(i) =1\\\\}$. \n* Similarly, I suggest the following for $S_2$: \n * $S_2 = \\\\{(𝑃, 𝑟) \| 𝑟(𝑗) = 1 ∧ ∃!𝑖 ∈[m], i \\neq j: 𝑟(𝑖) = 1 \\\\}$\n * $S_2 = \\\\{(𝑃, 𝑟) \| 𝑟(𝑗) = 1 ∧ \\sum\\limits_{i =1}^{m} r(i) =2 \\\\}$\n* Furthermore, $S_3$ can also be defined by set notation: First define the set of all possible ensembles with 2 elements:\n * $\\hat{S_3} = \\\\{ (P, r) \| ∃!𝑖,j ∈ [m], i\\neq j : 𝑟(𝑖) = 1 = r(j) \\\\}$\n * $S_3 \\subset \\hat{S_3}$ with $\|S_3\| = m$\n\nI will update my review rating (Weak Accept -> Accept), as I assume that these minor corrections will be done either before rebuttal ends, or latest for the camera ready version if the paper is accepted.", "role": "assistant" }, { "content": "Title: Thanks\nThanks a lot for you swift and constrcutive reply. We will definitly take a look at D.7 to make sure it is correct.\nThanks for putting in the the time to even check the appendix in such a detailed manner!\n", "role": "user" }, { "content": "Title: Forgot one Point\nI actually forgot one additional minor point: \n* $\\mathbb{Z}^+$ does not include $0$, therefore you will have a problem with $r(i)=0$, therefore you probably want to use $\\mathbb{Z}^+_0$", "role": "assistant" } ]
-Qr0uIfNbUN	CLeaR 2022 Conference	Reviewer_Prcm	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_Prcm, and the following are the paper you need to review:\n```<<-Qr0uIfNbUN>>```", "role": "user" }, { "content": "Review Content: Summary: This work is the first to work on relational causal models. It proposes \\sigma-separation for such causal models and show it is sound and complete under relatively strong assumptions.\nMain review: ## major comments\n1. I wonder how meaningful it is to study causal models with cycles. Can we always decompose the variables s.t. the cycles can be avoided? For example, we can decompose performance into P1 and P2, then we let P1 be the cause of quality and P2 be the effect of quality.\n2. It seems assumption 3 is likely to be violated in many cases. For example, in a standard probabilistic SCM, every observed variable would have an exogenous noise variable pointing to it.\n\n## details\n1. I think in definition 1 it should be d-separation instead of \\sigma-separation.\n2. There are many terminologies and notations introduced, it is better to have a table to summarize them.\n3. It is better to intuitively introduce the difference between \\sigma separation and d-separation with an example after introducing their definitions.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper\n", "role": "assistant" }, { "content": "Title: Response to reviewer Prcm\nWe thank the reviewer for the feedback and provide responses to each point.\n\n> I wonder how meaningful it is to study causal models with cycles. Can we always decompose the variables s.t. the cycles can be avoided? For example, we can decompose performance into P1 and P2, then we let P1 be the cause of quality and P2 be the effect of quality.\n\nIt might be possible to avoid a cycle by decomposing variables. However, the motivation behind the study of cyclic causal models often comes from a practical point of view where we have limited scope of observation. As a general example, in an economic system, prices of products can be a function of supplied quantities. On the other hand, supplied quantities can also be a function of price. The underlying dynamic process can be described through an acyclic causal graph over time. But if we don’t have enough observations over time and want to approximate such phenomena it eventually introduces a cycle that represents the steady state. Similarly, in our problem setup, we are given a relational schema that consists of the attributes that we can observe. So, even though decomposing variables is an option to avoid cycles theoretically, it cannot necessarily serve the practical purpose in general. We refer the reviewer to [Bongers et. al., 2021](https://arxiv.org/pdf/1611.06221.pdf) for further discussion. We can clarify this further in the paper.\n\n> It seems assumption 3 is likely to be violated in many cases. For example, in a standard probabilistic SCM, every observed variable would have an exogenous noise variable pointing to it.\n\nAssumption 3 is on the topology of the relational structure, i.e., the network which defines which entities are connected to each other. For example, if entities are social network users, this would exclude users of degree 0 or degree 1. However, there is no assumption on the dependence structure over the random variables. Under our assumptions it is entirely valid to have each random variable in the ground graph to have an exogenous noise variable pointing to it. We will clarify this further in the paper.\n\n> I think in definition 1 it should be d-separation instead of $\\sigma$-separation\n\nIndeed it should be “d-separated” instead of “$\\sigma$-blocked” and “d-blocked”. Thank you for catching this unintentional mistake. \n\n> There are many terminologies and notations introduced, it is better to have a table to summarize them. \n> It is better to intuitively introduce the difference between $\\sigma$-separation and d-separation with an example after introducing their definitions.\n\nWe appreciate the suggestions on improving the paper. We will include a terminology/notation table and an example showing the difference between $\\sigma$-separation and d-separation.\n", "role": "user" }, { "content": "Thank you for the authors' feedback. The response covers nearly all of the issues I raised. I will continue to recommend accepting this work.", "role": "assistant" } ]
-Qr0uIfNbUN	CLeaR 2022 Conference	Reviewer_BnHP	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_BnHP, and the following are the paper you need to review:\n```<<-Qr0uIfNbUN>>```", "role": "user" }, { "content": "Review Content: Summary: \nMain review: The focus of the work is on relational causal models, which generalize over previous causal models by capturing interactions between non-iid instances. A causal representation named abstract ground graph (AGG) was previously introduced for reasoning in relational models. But this representation does not allow for cycles. The contribution of this work is to generalize that line of work by extending the framework to allow for cycles. The authors introduce a presentation called sigma-AGG. Also introduce relational sigma-separation for reasoning about relational queries and show that it is sound and complete.\n\nA clear motivation for using relational models is not presented and the advantage of using relational models over, say, chain graphs is not clarified by the authors. This paper is of interest only to readers sufficiently interested in the original papers on relational models.\n\nThe authors show that under Assumption 3, relational sigma-separation is complete. The authors have not clarified what is the interpretation of this assumption and what real-life models are excluded due to this assumption.\n\nDefinition 3 is not clear. What does it mean that dependencies construct cycles. At this point of the paper, nothing is said yet about the connection between the distribution and a cyclic graph. \n\nIn general, the notion of cycle in relational models is not clear. In SCMs, we start with structural equations, and we obtain the edges based on those equations. Then those equation can result in cycles, which could be interpreted as steady state of the equations. What is the meaning of cycles in a relational model?\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to reviewer BnHP\nWe thank the reviewer for the feedback and provide responses to each point.\n\n> A clear motivation for using relational models is not presented and the advantage of using relational models over, say, chain graphs is not clarified by the authors. \n\nThe main motivation for our work is to enable the study of social phenomena through rigorous reasoning and identification of effects of interest (e.g., contagion). Some of the literature on interference and SCM considers only templates with pairwise relationships (E.g., Ogburn & VanderWeele 2014, Shalizi & Thomas 2011). This is unrealistic because social networks are not composed of pairs of nodes that are independent of other pairs of nodes. To be realistic, these templates would need to have dependencies with themselves. RCMs allow us to represent such more complex dependencies. What RCMs offer in this context is two-fold: 1) they offer a general expressive language to reason about causal dependencies in relational data where a single template can represent infinite instances. 2) They allow a rich set of tools (i.e. relational d-separation, AGG) which enable representation and learning of relational causal models from observational samples. \n\nChain graphs represent feedback as explicitly non-causal edges whereas the framework of $\\sigma$-separation also provides a rich representation for reasoning over cyclic relationships. It allows a cycle between two nodes to be represented as an explicit causal relationship, and there can be directed cycles of arbitrary length (chain graphs require a weakened form of acyclicity). \n\n> This paper is of interest only to readers sufficiently interested in the original papers on relational models.\n\nEven though the community where this work directly fits is small, it can be beneficiary to a wide range of audiences. We hope our work can serve a similar purpose to identification in non-relational settings. Researchers from network science and computational social science often want to measure certain influence/peer effect quantities or events and are unsure whether it's possible or not. They can use our work as a tool to identify what can be measured. \n\n> The authors show that under Assumption 3, relational $\\sigma$-separation is complete. The authors have not clarified what is the interpretation of this assumption and what real-life models are excluded due to this assumption.\n\nAssumption 3 is on the topology of the relational structure, i.e., the network which defines which entities are connected to each other. The interpretation is that entities that are not connected to any other entity or are connected to only one other entity are excluded. For example, if the entities are users in a social network, the framework would only consider users who have degree at least two, i.e., are connected to at least two other users. Assumption 3 prohibits one-to-one or one-to-many relationships and only allow many-to-many relationships which covers a wide range of real-world phenomena. For example, friendship ties in social networks, employers developing products, social network users and their platforms of choice are some common examples which fall within the scope of our work.\n\n> What does it mean that dependencies construct cycles. At this point of the paper, nothing is said yet about the connection between the distribution and a cyclic graph. \n\nWe use dependency to denote that the distribution of one random variable is causally influenced by the values of another. The connection between the structure of the graph and the underlying distribution is made via $\\sigma$-faithfulness (Def. 3). We will update the text to make this more clear.\n\n> What is the meaning of cycles in a relational model?\n\nThe meaning of cycles in relational models is very similar to the meaning of cycles in SCMs except that in relational models, the cycles can be between relational variables and we need to account for the fact that independencies can break when unrolling the relational model at the instance level. This is why relational d-separation was introduced in the first place. A person being influenced (e.g., in their buying habits, political views) by their peers and vice-versa is an example of a feedback loop between relational variables. One way to try to represent these feedback loops would be with SCM templates with pairwise relationships (e.g., the ones used in Ogburn & VanderWeele 2014, Shalizi & Thomas 2011) and assume SUTVA at the node pair level. However, this is not a realistic representation because social networks are not composed of pairs of nodes that are independent of other pairs of nodes. To be realistic, these templates would need to have dependencies with themselves. This is where RCMs help. The fundamental challenge in representation and learning of RCMs is to be able to give guarantees about causal effect identification based on the unfolding of relational templates which can have dependencies with themselves. We will clarify this further.\n", "role": "user" }, { "content": "I appreciate the authors' efforts in addressing my concerns. The response addresses most of my concerns. I stand by my favorable recommendation for this paper.", "role": "assistant" } ]
-Qr0uIfNbUN	CLeaR 2022 Conference	Reviewer_Wvcq	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_Wvcq, and the following are the paper you need to review:\n```<<-Qr0uIfNbUN>>```", "role": "user" }, { "content": "Review Content: Summary: This work generalizes the relational causal models to account for cycles by introducing \"relational sigma-separation\" based on the work of Forré and Mooij (2017). \nMain review: The most important problem with this manuscript is its lack of clarity. The concepts on which it builds, i.e. relational causal models and cyclic causal models, are non-standard and thus should be defined clearly to help the non-expert reader (like myself) undertand the contributions. I often had to refer to either Forré and Mooij (2017) and Maier et al. (2013b) to understand some notions specifically related to cyclic causal models and relational causal models.. This is not necessarily a bad thing given some notions are fairly complex and cannot be fully explained in the background section of a conference paper, but, once I understood the notion, it became clear that it could have been explained more clearly in the manuscript without eating up too much space. This lack of clarity prevents me from recommending acceptance at the moment.\n\n### Originality:\nTo the best of my knowledge, cyclic relational causal models are novel and were not investigated before. That being said, most theoretical results in this paper seem to be rather direct applications of the results of Forré and Mooij (2017) and Maier et al. (2013b), which leads me to think this work has medium originality.\n\n### Significance: \nThe motivation for relational causal models was almost completely missing from the manuscript. In addition, the cycle between \"employee.performance\" and \"product.quality\" in the motivating example is never explained. As a result, the overall contribution of this work lacked motivation which led me to believe its significance is somewhere between low and medium. \n\n### Clarity:\n- Related work is hard to follow for someone unfamiliar with relational causal models and cyclic models. Consider integrating it with the Preliminaries section.\n- Definition 1: should be d-separation instead of sigma-separation.\n- Section 3.4: This section was very hard to understand. Why are they more than one relational skeleton? The notion of \"relational path\" and \"perspective\" feels largely unmotivated. What is the difference between [Alice].salary and [Laptop, develops, Alice].salary? \n- In Figure 2, it would be useful to have a relational skeleton represented between the relational causal model and the ground graph.\n- The explanation of an \"abstract ground graph\" (AGG) in Section 3.5 was very fast given its central importance. After reading about it in Maier et al (2013b), I noticed that the definition in the present manuscript does not mention that the AGG is actually dependent on the perspective and the \"hop threshold\" h. This seems like an important point completely missed. This omission makes the relational counterpart for graphical d-separation in section 3.6 very confusing. Also, the notion of \"intersection variables\", which is important for later results, is very poorly explained. What does the reader should understand from \"Intersection variables are constructed from pairs of relational variables with non-empty intersections.\"? The intersection of relational variables was never defined. \n- Definition 5 what is s(B)? Couldn't find the definition.\n- Definition 6: the hop threshold \"h\" appears for the first time next to a AGG, without any prior explanation. Also, points 1.b and 2.b are not clearly explained. The former refers \"intersection variables\" which were ambiguously defined and the latter uses \"inherit\" in a very vague sense.\n- Shouldn't Definition 7 be a theorem or a proposition?\n\t\n\n### Technical quality:\n- I am not familiar with cyclic causal models, but the paragraph starting with \"Spirtes (1995); Neal (2000)\" just after definition 1 was confusing. It says that d-separation are not sufficient to entail all the corresponding conditional independence relations in a direct cyclic graph. However, this statement is meaningless without formalizing a connection between the causal graph and the distribution via a Markov property or a structural causal model. For instance, in an acyclic causal model, we can assume that the distribution is entailed by an SCM with some graph G. Then, it is meaningful to talk about the relation between the graph G and the conditional independences in the distribution. \n- On a related note, it is mentioned nowhere how one should translate the cyclic graph into conditional independence statements. I suspect this is what Definition 2 is for: \\sigma-separation implies corresponding conditional independence? It should be explicitly stated.\n- The lack of clarity made reading the main theoretical contributions of this work very difficult. As a consequence, I could not assess their quality.\n\n# Response to rebuttal \nI thank the authors for providing clarifications. I maintain my initial assessment: the lack clarity and motivation are the two main problems with this manuscript that prevent me from recommending acceptance at the moment. I encourage the authors to improve on these aspects in a future iteration of their work.\nRating: 4 possibly reject, but has redeeming facets\nConfidence: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper\n", "role": "assistant" }, { "content": "Title: Response to reviewer Wvcq\nWe thank the reviewer for the feedback and provide responses to each point.\n\n> Most theoretical results in this paper seem to be rather direct applications of the results of Forré and Mooij (2017) and Maier et al. (2013b), which leads me to think this work has medium originality.\n\nWhile we agree that our work builds upon the works of Forré and Mooij (2017) and Maier et al. (2013b), neither one considers the implications of introducing cycles in relational models. The contributions of our work include the following: \n1) To the best of our knowledge, this is the only work that aims to reason about cyclic relational causal models. While much progress has been made on representation and learning of RCMs, none of them address feedback loops and do not allow for reasoning about identification in the presence of feedback loops. We study cyclic RCMs since they offer the necessary representation to reason about a lot of real-world causal problems where SUTVA does not hold. \n2) We introduce a new abstract representation ($\\sigma$-AGG) for cyclic RCMs and a means (relational $\\sigma$-separation) to reason about relational queries on such models. We provide theoretical justification for the proposed representation. \n3) We provide sufficient conditions for the completeness of AGG which resolves the counterexample presented in (Lee et. al. UAI 2015). This non-trivial contribution greatly improves the applicability of AGG and has implications beyond the scope of cyclic relational models. \n\n> The motivation for relational causal models was almost completely missing from the manuscript. In addition, the cycle between \"employee.performance\" and \"product.quality\" in the motivating example is never explained.\n\nThe main motivation for our work is to enable the study of social phenomena through rigorous reasoning and identification of effects of interest (e.g., contagion). Some of the literature on interference and SCM considers only templates with pairwise relationships (E.g., Ogburn & VanderWeele 2014, Shalizi & Thomas 2011). This is unrealistic because social networks are not composed of pairs of nodes that are independent of other pairs of nodes. To be realistic, these templates would need to have dependencies with themselves. RCMs allow us to represent such more complex dependencies. What RCMs offer in this context is two-fold: 1) they offer a general expressive language to reason about causal dependencies in relational data where a single template can represent infinite instances. 2) They allow a rich set of tools (i.e. relational d-separation, AGG) which enable representation and learning of relational causal models from observational samples. \n\nIn the given hypothetical example, an employee's performance can directly influence the quality of the product. We can imagine that a good product influences the self esteem of the employee as well which eventually impacts their performance. This can go on in a loop such that we can imagine a steady state between these two observed variables which is represented by a feedback loop in the relational model. We will add this explanation.\n\n> Related work is hard to follow for someone unfamiliar with relational causal models and cyclic models. Consider integrating it with the Preliminaries section.\n\nThank you for that feedback.\n\n> Definition 1: should be d-separation instead of $\\sigma$-separation.\n\nIndeed it should be d-separation instead of $\\sigma$-blocked. Thank you for catching this unintentional mistake.\n\n> Why are there more than one relational skeleton?\n\nA relational schema can result in multiple instances. For example, we can assume a simplistic schema with an entity Person and a relationship Friendship which represents friendship ties among instances of entity Person. In one possible instance a person Alice can be friends with only one other person Bob, however in a different instance Alice can be friends with two other persons, Bob and Chris. Both these instances are valid and come from a single schema. Each of these instances can be represented as a separate skeleton. \n\n> What is the difference between [Alice].salary and [Laptop, develops, Alice].salary?\n\n[Alice].salary describes the salary attribute of Employee instance Alice from the perspective of employee class Employee. On the other hand, [Laptop, develops, Alice].salary describes the same instance-attribute pair but from the perspective of entity class Product. While they refer to the same thing in this example, the concept of perspective allows an expressive semantic for referring to relational variables. For example, we can refer to the set of employees who worked on a specific product from the perspective of Product but not from the perspective of Employee. We will include a further discussion about “perspectives” of relational models which should make this clearer and relates to the next point. \n> In Figure 2, it would be useful to have a relational skeleton represented between the relational causal model and the ground graph.\n\nThank you for this suggestion, it makes sense and we can include it.\n\n> Present manuscript does not mention that the AGG is actually dependent on the perspective and the \"hop threshold\" h. This seems like an important point completely missed.\n\nWe thank the reviewer for pointing this out. We mention perspective at the end of section 3.4 and hop threshold at point 1a of definition 6 but if the AGG’s dependence on them was not clear, then we will definitely make this more explicit. \n\n> What should the reader understand from \"Intersection variables are constructed from pairs of relational variables with non-empty intersections.\"?\n\nIntersection variables refer to the result of set intersection on multiple relational variables. For example, [Laptop, Develop, Employee] refers to the set of employees who develop the product laptop. On the other hand, [Keyboard, Develop, Employee] refers to the set of employees who develop the product Keyboard. If we consider the intersection of these two sets we get the employees who develop both laptop and keyboard. Such intersection variables are a key part of AGG. Since we directly use the notion from Maier et. al. we didn’t include a lot of detail on them in the paper due to space constraints. We plan to include a section in the appendix with a detailed description of intersection variables.\n\n> Definition 5 what is s(B)? Couldn't find the definition.\n\ns(B) refers to the entity instances of entity type B in the given skeleton s. We will clarify this notation.\n\n> Shouldn't Definition 7 be a theorem or a proposition?\n\nYes, it should be a theorem. That was an unintentional mistake that we will correct.\n\n> The paragraph starting with \"Spirtes (1995); Neal (2000)\" just after definition 1 was confusing. It says that d-separation are not sufficient to entail all the corresponding conditional independence relations in a direct cyclic graph. However, this statement is meaningless without formalizing a connection between the causal graph and the distribution via a Markov property or a structural causal model.\n\nWe will add a sentence in the beginning of that paragraph to formalize that connection. The statement itself is an established result from the works of Spirtes (1995); Neal (2000). \n\n> It is mentioned nowhere how one should translate the cyclic graph into conditional independence statements. I suspect this is what Definition 2 is for: $\\sigma$-separation implies corresponding conditional independence? It should be explicitly stated.\n\nThe conditional independencies in a cyclic graph are similar to the acyclic case except that one has to account for the fact that a node can serve both as a collider and non-collider between two other nodes in the presence of cycles. For example, not observing B in the graph B->A->B<-C makes A and C dependent because of the chain C->B->A despite B being a collider in A->B<-C. \\$sigma$-separation formalizes how these independences can be found. We will clarify this further. ", "role": "user" }, { "content": "Thanks for the authors' response. It addresses most of my concerns.", "role": "assistant" } ]
3QLFAlGeD3u	CLeaR 2022 Conference	Reviewer_Hjdm	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_Hjdm, and the following are the paper you need to review:\n```<<3QLFAlGeD3u>>```", "role": "user" }, { "content": "Review Content: Summary: \nMain review: An estimated CATE function could potentially reveal individual covariates, treatment status, or potential outcomes. The authors propose a general approach for multi-stage learning under differential privacy guarantees and apply it to CATE estimation. They show empirically the trade-off between privacy and estimation accuracy.\n\nThe paper is well-written, however familiarity with the differential privacy is assumed by the authors.\n\nThe only concern about this manuscript is that it does not have much of theoretical contribution, but on the other hand, it is has great practical value and should be of interest to applied data scientists. \n\nThe authors perform a thorough evaluation of the proposed approach. The results show the surprising different behaviors of DP-EBM-S-learner compared to DP-EBM-DR-learner and DP-EBM-R-learner. This shows that choosing a flexible CATE estimator versus a simple estimator is not a trivial decision.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Thanks ++\nThanks for the review. We're glad you like the paper. We'll try to add more introduction and discussion about DP to the paper for readers who are not familiar with it. Another reviewer also had questions about the theoretical contribution. We'll modify the paper to make it clearer that the multiple splitting procedure we adopt is the key to being able to prove tight DP bounds for our algorithm -- without it the theoretical guarantees would be much weaker and the impact of DP noise much worse. Thanks again for the review.", "role": "user" }, { "content": "I acknowledge the authors' response. It resolves the majority of my concerns. I continue to recommend this work positively.", "role": "assistant" } ]
3QLFAlGeD3u	CLeaR 2022 Conference	Reviewer_GapS	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_GapS, and the following are the paper you need to review:\n```<<3QLFAlGeD3u>>```", "role": "user" }, { "content": "Review Content: Summary: Review of Paper#67\nMain review: \nThis paper studies the problem of estimating conditional average treatment effects (CATE) with differential privacy guarantees. It proposes a general meta-algorithm that can work with simple, single-stage CATE estimators such as S-learner and more complex multi-stage estimators like DR and R-learner. The authors perform a privacy analysis utilizing the parallel composition property of differential privacy. Finally, experiments demonstrate how differential privacy noise affects the estimated CATE function.\n\nOverall, I think the author considers an interesting problem of addressing privacy concerns in estimating causal effects. Such a problem could be prevalent due to the sensitivity of healthcare data. I appreciate the authors for starting this conversation. However, I do have some questions and concerns, summarized as below.\n\n- This paper focuses on estimating conditional average treatment effect (CATE) while ensuring differential privacy (DP) constraints. This is a bit funny and seems ill-posed. For instance, suppose one wants to compare the CATE of a drug on male and female patients, and the actual CATEs of the male and female population diverge significantly. In this case, ensuring the privacy constraints on gender requires one to add DP noise on the CATE function until one cannot differentiate the CATE on the male and female population. There exists no ideal CATE function in this setting. It makes me wonder if DP formalism is the best tool to address the privacy challenges in causal inference tasks.\n\n- I would suggest writing Sec 2.1 using the formal language of causal inference, e.g., Pearl's structural causal model and the backdoor criterion.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper\n", "role": "assistant" }, { "content": "Title: Reply to Reviewer GapS\n1. \"Is DP the best tool to address the privacy challenges in causal inference tasks?\"\n\nDifferential Privacy has become the standard for protecting privacy in the machine learning and database communities. Differential Privacy focuses on protecting the privacy of individuals in a dataset, not protecting the privacy of groups such as male or female.\nMeeting the DP definition does not require equalizing predictions across distinct groups like male or female.\nAs long as the CATE function for gender does not reveal too much information about any one individual, then the algorithm will be able to learn different heterogeneous treatment effects for male and female while satisfying differential privacy.\n\n2. \"I would suggest writing Sec 2.1 using the formal language of causal inference, e.g., Pearl's structural causal model and the backdoor criterion.\"\n\nThe paper presents the setup in the potential outcome framework following previous work in the literature (Nie and Wager (2020), Kennedy (2020)). \nWe can potentially also present the same setup in a Structural Causal Model (SCM) framework. \nThe two are ultimately equivalent to each other. \nMore precisely, the DGP can be represented using the following structural causal model (Pearl (1995, 2009). Each observed variable $(X,T,Y)$ is a function of parent variables and unobserved background factors $(U_X,U_T, U_Y)$:\n$$X = f_X(U_X)$$\n$$T = f_T(X, U_T)$$\n$$ Y = f_Y(X, T, U_Y)$$\nOur conditional ignorability assumption, $(Y(0), Y(1))\\perp T\|X$, implies that $U_T\\perp U_Y \| X$. By intervening in the SCM we can generate the counterfactual outcome $Y(t)=f_Y(X,t,U_Y)$. The CATE is then\n$\\tau(x)=E[Y\|do(T=1), X=x] - E[Y\|do(T=0), X=x] = E[f_Y(X, 1, U_Y) - f_Y(X, 0, U_Y)\|X=x]$.", "role": "user" }, { "content": "I appreciate the authors' feedback. They have addressed most of the issues I raised.", "role": "assistant" } ]
3QLFAlGeD3u	CLeaR 2022 Conference	Reviewer_Qg9u	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_Qg9u, and the following are the paper you need to review:\n```<<3QLFAlGeD3u>>```", "role": "user" }, { "content": "Review Content: Summary: This paper studied the indices of DP for the entire procedure of estimating CATE when each component of the procedure is made DP and suggested some interesting directions based on results of numerical experiments.\nMain review: Strengths:\n\n1. Conditional, or heterogeneous, average treatment effect estimation (CATE) is a natural problem that differential privacy is of great concern, given the essential role of CATE in individualized decision making. Therefore, this paper is indeed trying to make progress on an important problem. The topic is very fresh -- there are indeed only a few papers that consider differentially private causal inference.\n\n2. The section before numerical experiments is relatively well-written, not dense with technical jargon, and accessible to people working on either causal inference or privacy data analysis.\n\nWeaknesses:\n\n1. The insight offered by the only theorem proved (Theorem 6) is quite limited. It is good to calculate the \"tight\" DP indices for the entire estimation procedure, but the only concrete example in the paper is about the trivial scenario: if all sub-routines are $(\\epsilon, \\delta)$-DP, then the entire procedure is $(\\epsilon, \\delta)$-DP. The motivation for proving this theorem is not strong, in my opinion.\n\n2. The numerical experiment section lacks sufficient detail. The authors did not provide any details on the DP-nuisance parameter estimation algorithms. It is difficult to judge the value of the experiments without knowing these details.\n\n3. The potential research questions suggested in this paper are only based on one simulation study. Since this paper is not about proving important theoretical results, more comprehensive simulation experiments should be performed. A pure \"intuition + simulation studies\" based paper, even without any theoretical guarantees, can still be very valuable to the community. But the experiments must be very comprehensive and aim at pushing the boundaries of the data generating mechanisms as hard as one could.\n\n4. The insights offered by the experiments are not very pedagogical. When the sample size is large, complicated methods can perform better than simpler methods; when the sample size is low, simpler methods might be more stable so one might be willing to sacrifice some bias. Isn't it true for almost any statistical practice? How much more could we learn by following these directions, except that we add the DP component?\n\n5. The type of differential privacy that the authors tackled is black-box. For instance, probably a most natural DP-related setting is what if some of the key confounders have to be taken to be differentially private while modeling the nuisance parameters and the final CATE function. A more concrete example will be more useful to illustrate the significance of either the theoretical or the empirical contribution of this paper.\n\nOverall, the paper is trying to tackle an interesting problem. Unfortunately, the current \"signal\" in the paper is quite low: the only theorem is not a core contribution to the field; the simulation experiments are not solid or comprehensive. As a result, I would not recommend acceptance of this paper.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Reply to Reviewer Qg9u\n1. \"The insight offered by Theorem 6 is quite limited...\"\n\nWe believe the motivation for Theorem 6 is strong and the insight it offers is nontrivial. The paper is about CATE estimation under differential privacy guarantees, which requires that any algorithm be accompanied by a formal proof of privacy. This is what Theorem 6 is all about. \n\nTheorem 6 is nontrivial. In particular, ``if all sub-routines are $(\\varepsilon, \\delta)$-DP, then the entire procedure is $(\\varepsilon, \\delta)$-DP'' mentioned by the reviewer, is not true for arbitrary multi-stage algorithms. \nOne of the most important contributions of this paper is pointing out that multiple sample splitting, a key feature of our CATE estimation procedure, is essential for a simple yet tight analysis of differential privacy. \nMore precisely, multiple sample splitting enables the usage of the parallel composition theorem introduced in Smith et al. (2021). \nWithout sample splitting, we will need to pay additional sequential composition cost in the differential privacy analysis. For example, if composed naively without sample splitting, the privacy loss of the corresponding DR-learner would be $3 * \\varepsilon$ instead of $1 * \\varepsilon$. The simplicity of the DP guarantee proved in Theorem 6 is the result of careful algorithm design together with the utilization of state-of-the-art DP analytical tools, not because the algorithm and analysis is trivial. We will edit the paper to make this more clear. \n\nFinally, we prove the DP guarantee in the f-DP framework and then specialize to $(\\varepsilon, \\delta)$-DP framework mainly because f-DP framework enables a tighter analysis than $(\\varepsilon, \\delta)$-DP. \nSince f-DP and the related Gaussian differential privacy framework covers a variety of DP frameworks like, e.g. $(\\varepsilon, \\delta)$-DP and Renyi-DP, and have recently gained increasing popularity in the differential privacy community, the way we present the DP analysis is also meant to be friendly to researchers used to various DP analytical tools. \n\n2. \"The numerical experiment section lacks sufficient detail...\"\n\nDetails of our experimental setup, including choices of hyperparameters, dataset specifications, sample sizes, values of $\\varepsilon$, and metrics tracked, are all reported in Sections 5.1.1, 5.1.2, and 5.1.3.\n\n\"The authors did not provide any details on the DP-nuisance parameter estimation algorithms...\"\n\nThe details of the machine learning method used for all stages of the CATE estimator -- DP-EBM -- are introduced in Section 4.1 of our paper. Algorithm 2 step 2 explicitly specifies the method we use for DP-nuisance parameter estimation with DP-EBM. \n\n3. \"The potential research questions suggested in this paper are only based on one simulation study...\"\n\nWe cannot agree with the reviewer's claim. Section 5 of our paper presents experiments of the proposed algorithms on six different DGPs. \nOne of the six datasets is a real dataset from Arceneaux et al. (2006), which studies the effect of paid get-out-the-vote calls on voter turnout. \nThe other five datasets are simulated datasets. \nOverall, the six datasets cover the DGPs with binary outcomes or continuous outcomes, independent or correlated covariates distribution, simple or complicated baseline control outcome functions, randomized control trials or complicated treatment assignment regimes, and smooth or nonsmooth CATE functions. \nMore detail can be found in table 1 and the paragraph at the bottom of page 9 of the paper. \nIn this sense, our simulation study challenges the CATE estimation algorithm in many different aspects. \nFor each of the datasets, we also consider seven different sample sizes. \n\nUnlike evaluation of a prediction algorithm, evaluation of CATE estimation procedure often requires the knowledge of the ground truth CATE function (Caveat: There is early literature on causal evaluation, but results there are still at an early stage.). \nThis intrinsic challenge implies that researchers will need to use simulated datasets or semi-synthetic datasets in which they know the true CATE functions for evaluation. \nGiven this situation, we collect datasets and DGPs used in recent papers Nie and Wager (2021) and Kennedy (2020) in the CATE literature for our data experiments. \nOur experiments are on par with the state-of-the-art in the CATE literature.\n4. \"The insights offered by the experiments are not very pedagogical...\"\n\nOur experiments deliver three takeaways as summarized on page 12. \nThe reviewer seems to miss the first two takeaways, both of which are important for characterizing the effect of differential privacy on CATE estimation. The reviewer only comments on the third takeaway and unfortunately misinterprets it.\n\nWe restate the first two takeaways here:\n\n(1) As expected, stronger privacy constraints lead to noisier and jumpier estimated CATE functions. \n\n(2) Most of the additional accuracy loss due to privacy can be attributed to a significant increase in variance, while bias increases at a much slower rate. This is important because bias is a much stronger concern than variance --- in CATE estimation which often deals with critical policy, economic or healthcare issues, it is much worse to confidently believe a wrong thing than to have too much uncertainty.\n\nOur third takeaway in the paper is not saying more complicated and data-hungry methods can only outperform simple methods when the sample size is large. \nInstead, our third takeaway in the paper is about how differential privacy, the new element we introduce to the CATE estimation problem, changes the tradeoff between more complicated vs simpler CATE estimation methods. \nWe restate the third takeaway in the paper here ``Flexible, multi-stage CATE learners incur larger accuracy loss from differential privacy than less flexible, single-stage CATE or ATE learners. \nIt takes a significantly larger sample size to exploit treatment effect heterogeneity under strong privacy constraints''. \nWe consider this takeaway instructive for shaping researchers' intuition on how differential privacy affects a spectrum of CATE estimation procedures differently. \nExperiment results behind this takeaway are presented in Figure 2 on Page 11. The text following Figure 2 presents the discussion with more detail.\n\n5. \"The type of differential privacy that the authors tackled is black-box...\"\n\nThe $(\\varepsilon, \\delta)$-DP and $f$-DP definitions of differential privacy used in this paper are standard and have been discussed extensively in the literature. \nThis paper states both definitions formally in definitions 1, 2, and 3. \nThese definitions are meant to quantify the worst-case privacy loss of data of each individual, which includes its covariates, treatment status, and potential outcomes. \nThe mathematical definition and the interpretation of differential privacy this paper considers are both transparent and well studied. \n\n\"A most natural DP-related setting is what if some of the key confounders have to be taken to be differentially private...\"\n\nThe reviewer suggests an alternative definition of privacy, which could be interesting, but alternate definitions of DP fall well outside the scope of this paper. \nMore specifically, the reviewer mentions a notion of confounder-specific privacy. \nWe are not aware of any definition addressing confounder-specific privacy in the differentially privacy literature. \nIt is not clear to us against what kind of privacy attacks this notion of privacy is robust. \nIn case this newly proposed notion of privacy makes sense, the DP guarantee of the CATE estimation algorithms in our paper will naturally be an upper bound of the kind of privacy loss the reviewer is concerned about. \n\n\"A more concrete example will be more useful...\"\n\nWe completely agree that a more concrete example that illustrates the theoretical and empirical significance of the work would be useful and we will add such an example to the final draft. Thank you for suggesting this.\n", "role": "user" }, { "content": "Title: Thank you for your rebuttal\nI want to thank the authors for the comprehensive rebuttal to my concerns. The rebuttals are convincing. Now I do agree now with the authors and other referees that this paper addresses quite important and novel questions in causal inference. As a result, I'll increase my evaluation to \"marginally above the acceptance threshold\".", "role": "assistant" } ]
50eDSLScz7r	CLeaR 2022 Conference	Reviewer_KuUF	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_KuUF, and the following are the paper you need to review:\n```<<50eDSLScz7r>>```", "role": "user" }, { "content": "Review Content: Summary: Develops an algorithm for discrete causal bandits that does not assume prior knowledge about the causal structure.\nMain review: This paper looks good enough to be accepted. That said, the clarity should be improved. \n\nPlease reword and be more specific when presenting definitions: “If we do not only observe the rewards but also some additional variables, and the rewards and these variables are generated according to a causal model, then the rewards are no longer independent, and this is called a Causal Bandit problem.”\n\nBe more specific about which previous works you are referring to. “In this case, the previously cited works on this topic have shown that one can this additional structure to improve our performance.”\n\nHow is the context variable different from the regime indicator variable F_t used by Philip Dawid? https://arxiv.org/pdf/2004.12493 (See section 2). Note that while this reference is relatively new the use of such variables can be found in older papers by Dawid and in Pearl’s Causality book.\n\nIn the introduction, the authors state, “we develop a framework for Causal Bandits where everything is learned from scratch.” This is an exaggeration. Later on in section 3.1 they state, “we introduce a novel information sharing estimator that relies on much less specific knowledge about the causal graph”. Please state explicitly the information being assumed about the causal graph.\n\nTypos and spelling mistakes: “then much less restrictive then specifically focussing on” \n\nThe algorithm is presented in the appendix. It should probably be moved to the main paper.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to review\nDear reviewer,\n\nThank you for your positive feedback. The points of clarity raised will greatly benefit the readability of the work!\n\nTo answer your questions:\n\nQ: “How is the context variable different from the regime indicator variable F_t used by Philip Dawid? https://arxiv.org/pdf/2004.12493 (See section 2). Note that while this reference is relatively new the use of such variables can be found in older papers by Dawid and in Pearl’s Causality book.”\n\nA: It is not different and is used here just as a convenient modeling choice. The proper citations will be added.\n\nQ: “In the introduction, the authors state, “we develop a framework for Causal Bandits where everything is learned from scratch.” This is an exaggeration. Later on in section 3.1 they state, “we introduce a novel information sharing estimator that relies on much less specific knowledge about the causal graph”. Please state explicitly the information being assumed about the causal graph.”\n\nA: In the theoretical analysis, a given separating set is assumed. In the presented algorithms and experimental results, however, separating sets are found purely from data and thus the statement in the introduction is not an exaggeration. We will clarify the simplifying assumption used in section 3.1.\n", "role": "user" }, { "content": "Thank you for the authors' feedback. It adequately addresses the points I raised. I stand by my favorable recommendation for this paper.", "role": "assistant" } ]
50eDSLScz7r	CLeaR 2022 Conference	Reviewer_7S3C	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_7S3C, and the following are the paper you need to review:\n```<<50eDSLScz7r>>```", "role": "user" }, { "content": "Review Content: Summary: An O.K. paper\nMain review: In this paper, the authors proposed a new bandit algorithm that can more efficiently estimate the causal effect via learning the causal graphs. The simulation results on the Sachs data show that there is improved cumulative regret compared to classical bandit approaches. I took a brief look at the proof and did not find any problem. Overall I think it is an O.K. paper and may be of interest to the CLeaR audience. Nevertheless, after carefully reviewed this manuscript, I still have the following questions:\n\n1. Theorem 2 only shows that there is variance reduction by given a valid separating set. However, in practice all the separating sets are learned from the algorithm and the causal discovery algorithm may not be guaranteed to find the valid set for separation. Is there any theoretical analysis to leverage the estimation errors from causal discovery algorithms into the regret or variance bound?\n\n2. In Section 5.2, the authors showed improved cumulative regret bound on Sachs data set. This means that when the causal graph is easy to learn, the proposed method can achieve a relatively nicer performance compared to orthodox bandit algorithms. I wonder how the performance would be when we instead apply it on a data set with more complicated causal graph. For example, when it is applied to a data with a dense causal graph or with a causal graph containing latent confounders, would it still perform well than the ordinary bandit methods? Or it would be the other way around?\n\nOverall I think it is an O.K. paper, but considering the problems on theoretical analysis and real world applications mentioned above, the contribution seems rather marginal.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Author response to review\nDear reviewer,\n\nThank you for your feedback and the time you spent carefully reviewing this work. To answer your questions:\n\nQ: “Theorem 2 only shows that there is variance reduction by given a valid separating set. However, in practice all the separating sets are learned from the algorithm and the causal discovery algorithm may not be guaranteed to find the valid set for separation. Is there any theoretical analysis to leverage the estimation errors from causal discovery algorithms into the regret or variance bound?”\n\nA: This is an excellent question and something we have also been thinking about. A natural way to derive theoretical guarantees for a causal bandit that relies on causal discovery would start from uniform consistency of the causal graph discovery. This can only be guaranteed under strong, unrealistic conditions (Robins et al., Biometrika 90(3) 2003). On the other hand, we believe that when using our proposed direct testing method, uniform consistency of the conditional independence test is not required to derive theoretical guarantees for a causal bandit that relies on direct testing to estimate separating sets. The reason is that even though an arbitrarily large number of samples is needed to distinguish a weak dependence from an independence, for the causal bandit it is not as crucial to distinguish these two as for estimating the true causal graph. Indeed, a weak dependence will only result in a small error in the estimated payoff (but could lead to a globally different causal graph in the worst case). So we do believe that our work may open up a way to derive theoretical guarantees for causal bandits under mild assumptions. However, deriving such results is nontrivial and we consider this beyond the scope of this work.\n\nQ: * “In Section 5.2, the authors showed improved cumulative regret bound on Sachs data set. This means that when the causal graph is easy to learn, the proposed method can achieve a relatively nicer performance compared to orthodox bandit algorithms. I wonder how the performance would be when we instead apply it on a data set with more complicated causal graph. For example, when it is applied to a data with a dense causal graph or with a causal graph containing latent confounders, would it still perform well than the ordinary bandit methods? Or it would be the other way around?”*\n\nAs can be seen in Mooij et al. (2020 JMLR), different causal discovery algorithms yield very different outputs on the Sachs dataset, and the resulting causal structures are not simple. If a dataset is generated by a dense SCM, separating sets would be very rare or not exist at all. As can be seen in figure 2(c) in the paper under review, if this is the case, the false positive rate of causal discovery leads to a small decrease in performance compared to the ordinary bandit method. The separating set property is still valid in the presence of confounders or cycles, so these do not pose a problem, as long as a causal discovery method is chosen that can handle these structures.\n", "role": "user" }, { "content": "Thank you for the authors' feedback. It adequately addresses the points I raised. I continue to recommend this work positively.", "role": "assistant" } ]
50eDSLScz7r	CLeaR 2022 Conference	Reviewer_nePs	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_nePs, and the following are the paper you need to review:\n```<<50eDSLScz7r>>```", "role": "user" }, { "content": "Review Content: Summary: This paper considers a causal bandit problem where a separating set between intervention and reward can be utilized through information sharing among different arms. With reduction in variance, the proposed causal bandit algorithm outperforms those without taking advantage of causal information obtainable through data and conditional independence tests.\nMain review: I have reviewed several causal bandit papers and found that some of the papers contribution is simply using the parents of reward variable Y to improve the performance. In my review, I wrote that there may be other separating set which can yield better variance reduction. Hence, I was somehow expected this idea being realized.\n\nThe authors solution is simple and clean. The use of context variable (~ regime indicator) is also a nice addition.\n\n\nSome questions:\nWhy does the algorithms in the appendix does not check a separating set on the fly?\n\nDo we have to perform causal discovery every iteration or once in a while (depending on p-value one may skip a few iteration)? Do we have an online version of conditional independence test or causal discovery algorithm that can be utilized in this setting?\n\nThe result with direct independence testing is better than those with causal discovery algorithm. Would it be desirable to employ some kind of “optimism in the face of uncertainty” for a “separating set”?\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is confident but not absolutely certain that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Author response to review\nDear reviewer,\n\nThank you for your very positive feedback! Funny coincidence that you also reviewed previous works on this precise topic. Indeed using the parents of the reward variable turns out to not be the ideal choice in general. \n\nTo answer your questions:\n\nQ: “Why does the algorithms in the appendix does not check a separating set on the fly?”\n\nA: Both algorithms discover separating sets on line 4 by invoking disc_sep_set, where how this is done exactly is left implicit to whatever causal discovery algorithm is used. This will be better explained in section 4.\n\n Q: “Do we have to perform causal discovery every iteration or once in a while (depending on p-value one may skip a few iteration)?”\n\nA: In our experiments, we run causal discovery initially, and then every time the number of data points increased by 10%. This will be made clear in the paper. (Add comment with potential of s values?)\n\nQ: Do we have an online version of conditional independence test or causal discovery algorithm that can be utilized in this setting?”\n\nAs far as we are aware this does not yet exist, so for now we use a standard conditional independence test. This is an exciting area for new research however!\n\nQ: “The result with direct independence testing is better than those with causal discovery algorithm. Would it be desirable to employ some kind of “optimism in the face of uncertainty” for a “separating set”?”\n\nA: We think exactly the reason direct independence testing performs better is that it is more optimistic with regard to separating sets. In sparse graphs, this seems to be a sensible strategy.\n", "role": "user" }, { "content": "Title: .\nThanks for the response. It seems that there are several interesting research directions.", "role": "assistant" } ]
59PvrMnEZm7	CLeaR 2022 Conference	Reviewer_uVEy	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_uVEy, and the following are the paper you need to review:\n```<<59PvrMnEZm7>>```", "role": "user" }, { "content": "Review Content: Summary: This paper considers a special case of nonparametric generalised method of moments in high dimensions, where the conditioning variable has a small intrinsic dimension. In particular, the authors propose an estimator that solves a locally weighted empirical conditional moment equation with sub-sampled kernel or sub-sampled K-NN weights. The authors show that the estimator can achieve the optimal convergence rate depending only on the small intrinsic dimension but not the high dimension where the covariates belong to, which escapes the well-known curse of dimensionality. Furthermore, the authors provide an explicit form of the asymptotic variance of their estimator, which helps construct the confidence intervals. The paper also proposes an adaptive data-driven approach for choosing the sub-sample size and proves that the estimator with the selected sub-sample size achieves the optimal rate. \nMain review: Nonparametric estimation of conditional moment models in high dimensions has been widely and well studied in the literature. In recent years, many works have been dedicated to establishing the large sample properties for the case where the conditioning variable has a small intrinsic dimension and found that the optimal convergence rate depends only on $d$; see the references in the paper and see also Jiao et al. (2021) and the references therein. The K-NN method of approximating the geodesic distance on the lower intrinsic dimensional manifold is commonly used (see Tenenbaum et al., 2000; Kpotufe, 2011; Chen and Muller, 2012 among others). The authors generalise the asymptotic results of Wager and Athey (2018), Athey et al. (2019) for sub-sampled random forests and that of Fan et al. (2019) for sub-sampled 1-NN estimator in the high dimension setting to their sub-sampled K-NN generalised method of moments in the low intrinsic dimension regime. Inspired by the ideas in Kpotufe (2011), the authors propose a data-driven method to choose the tuning parameter.\n\nThere are two main contributions of the paper: (1) the explicit form of the asymptotic variance of the estimator and (2) the data-driven tuning parameter that can guarantee the optimal convergence rate of the estimator. Nevertheless, I have the following major concerns:\n\n1. The motivation for using the sub-sampling and fixing the value of $k$ of the K-NN is unclear. Without the sub-sampling, the estimator is essentially a uniform kernel estimator, where the kernel shrinkage acts like the bandwidth of a standard kernel estimator and depends on the sample size $n$ and the K-NN parameter $k$. However, it is known that if $k$ satisfies certain rate conditions, such an estimator can also achieve the optimal convergence rate of a nonparametric estimator, and one may also incorporate kernel functions other than a uniform one to the K-NN to improve the practical performance; see Bickel and Li (2007); Kpotufe (2011); Padilla et al. (2020) for example. Then, what is the advantage of using the sub-sampling?\n2. The authors treat the K-NN parameter $k$ as a constant in the paper. As I mentioned in the first question, the kernel shrinkage parameter should depend on $k$. What would the asymptotic results of the estimator be if we allow $k$ to depend on the sample size $n$?\n3. In practice, $k$ has to be chosen to construct the estimator. Could the authors discuss how to determine $k$ and how does $k$ affect the practical performance?\n4. In the theorems in the paper, are $\\lambda$, $\\psi_{\\max}$ and $p$ constants lie in $(0,\\infty)$? If so, taking $B=1$ and $s=n$, which corresponds to the K-NN estimator without sub-sampling, in the upper bound of the square root of the mean squared error of the estimator, we have $O\\\\{\\lambda^{-1}\\psi_{\\max}\\sqrt{n^{-1}ps\\log\\log(pn/s)}\\\\}=O(1)$. This may not converge to 0. Why is that? Please correct me if I misunderstand anything.\n5. In the paper, the authors provide the explicit form of the asymptotic variance and say that if they have an estimate of the variance, they can build plug-in confidence intervals based on the asymptotic normality result. However, the exact estimator of the variance is not stated. Since this is one of the main contributions of the paper, the authors should be clear here. Based on the formula of the variance, many different estimation methods can be used. Do they all have similar theoretical and practical performance?\n6. In the simulation, the authors consider only one setting: one model, one sample size, one number of dimensions of the covariates and one intrinsic dimension. This does not reflect the theories well. \n (a) The sample size is set to $n=20000$ while the intrinsic dimension $d$ is only 2. Usually, with $d=2$, a sample size of $n=5000$ should be large enough to obtain a good performance. The authors may consider $n=1000,2000$ and $5000$ to present the convergence of the estimation and the coverage of the confidence intervals.\n (b) It would also be interesting to see a comparison between the proposed sub-sampled method and the K-NN estimator without sub-sampling but with $k$ chosen by cross-validation. \n7. Although the authors mention that the conditional moment models are related to heterogeneous treatment effect problems, the connection of the results to causal learning and reasoning is not straightforward. Indeed, in the treatment effect problems, other issues need to be considered in the estimation, such as confounding. Therefore, the estimation and inference results in the paper cannot be directly applied. This concerns if the paper is suitable for the conference.\n\nReference\n\nAthey, S., Tibshirani, J. and Wager, S. (2019) Generalized random forests. The Annals of Statistics, 47, 1148-1178.\n\nBickel, P. J. and Li, B. (2007) Local polynomial regression on unknown manifolds. Complex datasets and inverse problems, Institute of Mathematical Statistics, 177-186.\n\nChen, D. and Muller, H.-G. (2012) Nonlinear manifold representations for functional data. The Annals of Statistics, 40, 1-29.\n\nFan, Y., Lv, J. and Wang, J. (2018) Dnn: A two-scale distributional tale of heterogeneous treatment effect inference. arXiv preprint arXiv:1808.08469.\n\nJiao, Y., Shen, G., Lin, Y. and Huang, J. (2021) Deep nonparametric regression on approximately low-dimensional manifolds. arXiv preprint arXiv:2104.06708.\n\nKpotufe, S. (2011) K-NN regression adapts to local intrinsic dimension. In Advances in Neural Information Processing Systems, 729-737.\n\nTenenbaum, J. B., de Silva, V. and Langford, J. C. (2000) A global geometric framework for nonlinear dimensionality reduction. Science, 290, 2319-2323.\n\nPadilla, O. H. M., Sharpnack, J. and Chen, Y. (2020) Adaptive nonparametric regression with the K-nearest neighbour fused lasso. Biometrika, 107, 293-310.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is absolutely certain that the evaluation is correct and very familiar with the relevant literature\n", "role": "assistant" }, { "content": "Title: Response to Reviewer uVEy\nWe would like to thank the reviewer for their insightful review. Below we address your questions:\n\n1. It is correct that sub-sampling is not required for achieving optimal convergence rate in nonparametric setting. However, sub-sampling is required to make the $k$-NN a regular asymptotic normal estimator for inference. Without sub-sampling this estimator can achieve asymptotic normality but it would be irregular. For the same reason, Fan et. al. (2019) and Athey et. al. (2019) considered subsampled kernel estimators to achieve asymptotic normality, while Bickel and Li (2007), Kpotufe (2011) and Padilla et al. (2020) that study $k$-NN without subsampling do not address inference. \n\n2. You are right that the kernel shrinkage depends on both values of $k$ and $s$ (See Lemma 10). As stated in the paper, our results in Theorems 6 and 7 are valid for all sub-sampled kernel estimators, including $k$-NNs with $k$ depending on the sample size $n$. Once the conditions required for kernel shrinkage and incrementality rates for the $k$-NN established in Lemma 10 and 13 are satisfied, all the results in Theorems 12, 14, and 15 remain valid. For example, our convenient choice of $\\delta=s^{-1/d}$ requires that $s^{-1/d} \\leq \\frac{1}{2}\\exp(-k/2)$. One can incorporate these conditions and allow for settings where $k$ can get larger as $n$ gets larger. However, to keep the presentation simpler, and also keeping the adaptive parameter tuning problem simpler (tuning one parameter instead of two), we decided to mainly focus on the case where $k$ is constant.\n\n3. As mentioned in the answer to Q2, we decided to keep $k$ fixed and provided data-drive adaptive way for tuning $s$. These two may be related, but sub-sampling allows for inference and that is our main focus in this paper.\n\n4. Our main focus is on situations where $s << n$, where sub-sampling allows inference. In this case, we need to control for the error that is induced by sampling. This would contribute an additional term of order $\\tilde{O}(s/n)$ which as you mentioned does not exist in the case where $s=n$.\n\n5. This is a great question and it can be interesting follow-up work. However, studying different estimators for estimating the variance and their performances is outside of the scope of this work.\n\n6. We would be happy to provide results with smaller values of sample size $n$ in the final version. \n\n7. The conditional moment equation formulation in (1) allows for approaches that can handle situations with confounding. For example, instrumental variables is a case where the conditional moment restriction might hold for some set of exogenous (i.e. unconfounded) instruments $X$, even though the function $\\psi$ takes as input data $Z$ that may suffer from confounding. \nWith that said, even if this paper only accounted for the case with observed confounders, we would be in good company: Athey et. al. (2019) and Fan et. al. (2019) both considered cases with observed confounding variables as well. ", "role": "user" }, { "content": "Title: Response to the authors\nThank you for answering my questions. Your responses resolve all of my main concerns, and I can now update the score to 7: Good paper, accept.", "role": "assistant" } ]
59PvrMnEZm7	CLeaR 2022 Conference	Reviewer_QJkx	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_QJkx, and the following are the paper you need to review:\n```<<59PvrMnEZm7>>```", "role": "user" }, { "content": "Review Content: Summary: The paper provided an algorithm for estimating conditional moments based on a sub-sampled ensemble of the kNN Z estimator, and provided bound for estimation error. Theorems were also provided for adaptively choosing hyper parameter s to achieve the desired estimates. The reviewer find the paper clear and informative.\nMain review: \nThe reviewer is familiar with conditional moment models and their application, familiar with intrinsic dimensions, but only have textbook knowledge about non-parametric estimation.\n\nHere are my detailed comments:\n1)\tThe problem can use more motivation, e.g. in which application area is it more common to have high dimensional set of conditioning variables? In many cases, one can estimate the causal structure and given the estimated causal structure, only a smaller number of variables need to be conditioned on.\n2)\tSimulations: Is there any reason that the specific function family/data generation function was chosen? It would also be interesting to see numerical illustration/demonstration of the bound.\n3)\tIt would be helpful if the author can comment of the complexity of the estimation.\n4)\tAssumptions: Can the authors comment of some of the assumptions with respect to testability or how likely they are going to be met? In other words, guidance for when the methodology is applicable for people who are consider to use it on real world data.\nRating: 8 accept, good paper\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Response to Reviewer QJkx\nWe would like to thank the reviewer for their insightful review. Below we address your questions:\n\n1. We provide some examples of high dimensional spaces with low intrinsic dimensions in Appendix B. As mentioned there, one example is when the conditioning variable is an image, i.e., we want to do inference based on images. You can find other examples there too. You are correct that in many cases, one can estimate causal effects only using a small number of conditioning variables.\n\n2. Please see Appendix C where we explain the data generating process. We would be happy to add numerical illustrations of our theoretical bounds for some simple examples.\n\n3. The complexity of estimation can be decomposed into two parts: (1) calculating the weights using $B$ different sub-samples of size $s$ for different points and (2) solving the weighted generalized moment equation. The first part is usually not that difficult and complexity is $O(B C_{\\text{$k$-NN}})$ where $C_{\\text{$k$-NN}}$ is the complexity of $k$-NN algorithm executed on $s$ samples (this is generally $O(sd)$ unless approximate algorithms are used or further structure is assumed). The difficulty for solving the weighted generalized moment equation depends on the structure of this function. For example, in the case of convex moment, this can be solved very efficiently.\n\n4. Our assumptions stated in Assumptions 1 and 2 are mainly conditions on the score and moment function that need to be estimated. These assumptions are satisfied for many well-known moment functions used in practice (regression, quantile regression, etc.). Our assumption on the intrinsic dimension corresponds to having a high-dimensional feature space which is less complex locally. For the examples of feature spaces with small intrinsic dimension, see Appendix B. Our paper basically provides general estimation and inference guarantees for any combination of these two.", "role": "user" }, { "content": "I'm grateful for the authors' response. The reply answers almost all of my questions. I maintain my positive recommendation for this work.", "role": "assistant" } ]
59PvrMnEZm7	CLeaR 2022 Conference	Reviewer_HtcP	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_HtcP, and the following are the paper you need to review:\n```<<59PvrMnEZm7>>```", "role": "user" }, { "content": "Review Content: Summary: The paper considers estimators defined as solution to a conditional moment equation when the conditioned covariates are of high dimension but low \"intrinsic dimension\". Estimator based on a sub-sampled k-NN weighted estimating equation is considered and its properties are studied.\nMain review: ## Technical quality\nThe paper establishes the rate and asymptotic normality of the estimator. In particular, under an unknown intrinsic dimension, a data-driven procedure is shown to be adaptive. \n\nIn terms of its technical strengths and weaknesses in the context of the literature, I am unable to provide an informed assessment as I am not familiar with related works. \n\n## Clarity\nThe paper seems quite dense and I find it very hard to follow. After all, there is not much room left after introducing the definitions / conditions and stating the results. \n\nThere are a few broken sentences in the first paragraph of page 3. \n\n## Originality\nThe authors seem to be able to build upon previous works and generalize their results, e.g., the definition of \"intrinsic dimension\" and asymptotic normality of the sub-sampled k-NN estimator.\n\n## Significance\nAs authors mentioned in the Introduction, I presume that the results here are relevant to several problems in causal inference. But the impact is not immediately clear to me. Perhaps the authors can comment on the application of their results.\nRating: 6 marginally above the acceptance threshold\nConfidence: The reviewer is willing to defend the evaluation, but it is quite likely that the reviewer did not understand central parts of the paper\n", "role": "assistant" }, { "content": "Title: Response to Reviewer HtcP\nWe would like to thank the reviewer for their insightful review. Below we address your questions:\n\n1. The paper seems quite dense and I find it very hard to follow. After all, there is not much room left after introducing the definitions/conditions and stating the results. There are a few broken sentences in the first paragraph of page 3. \nResponse. We will fix the broken sentences in the first paragraph of page 3 and try to edit some of the text to make it less dense.\n\n2. As the authors mentioned in the Introduction, I presume that the results here are relevant to several problems in causal inference. But the impact is not immediately clear to me. Perhaps the authors can comment on the application of their results.\n\nResponse. Our paper is the first to show the possibility of achieving non-parametric inference for generalized moment equations with rates depending only on intrinsic dimension. This to some extent, can justify the success of non-parametric methods in causal inference. Our paper is mostly there to explain WHY existing methods work, rather than suggest new methods.\nAs we noted in the first paragraph, most forms of regression are special cases of Generalized Method of Moments, and find wide application in economics, statistics, public policy, political science, bio-statistics etc...", "role": "user" }, { "content": "I'm grateful for the authors' response. It adequately addresses the points I raised. I continue to recommend this work positively.", "role": "assistant" } ]
5zbRCWR3qQ	CLeaR 2022 Conference	Reviewer_6giK	[ { "content": "\nYou are an experienced academic paper reviewer. You will receive a conference paper in the field of computer science or AI. Please analyze the paper content carefully and output your review content.\nMore importantly, the author will further explain or refute the review comments you gave, and you need to further discuss with the author based on the author's response. \nThe purpose of this rebuttal and discussion is:\n- Clarification: During the review process, the reviewer (you) may not accurately understand the contribution or method of the paper due to unclear expression or misunderstanding of certain details. Through rebuttal, the author has the opportunity to clarify these misunderstandings and further illustrate the effectiveness of their innovations or methods.\n- Respond to criticisms and suggestions: The reviewer (you) may make criticisms or suggestions on the paper, and the author can use the discussion session to provide specific responses, explain why certain criticisms may not apply, or how to improve and optimize certain aspects of the paper.\nFinally, you and the author will work together to enhance the persuasiveness of the paper and improve the quality of the paper.\n", "role": "system" }, { "content": "Now your reviewer ID is assigned as Reviewer_6giK, and the following are the paper you need to review:\n```<<5zbRCWR3qQ>>```", "role": "user" }, { "content": "Review Content: Summary: The paper proposes a category theory formalization of causal model equivalence, i.e., when one model is an abstraction of another one.\nMain review: Strengths:\nThe paper uses category theory to formalize the idea of causal model equivalence which is rather elegant and consistent with previous works in category theory applied to causal modeling.\nThe theory derivation seems sound.\nThis model equivalence formalizes the notion of abstraction and thus has applicability to more fundamental ideas like coarse-graining and emergence.\n\nWeaknesses:\nThe categorical formulation only applies to discrete causal models, whereas in causal ML practice, people more often work with continuous causal models.\n\nThe categorical formulation only covers Bayesian networks and not full causal models: In Bayesian networks, only the conditional probabilities of a variable given its parents are known, but not the deterministic function mechanism and the noise specific to this variable, i.e., only P(A\|PA_A) is modeled and not the combination of deterministic function and noise A := f(PA_A, N_A).\nAccording to the ladder of causation from Pearl, this makes Bayesian networks only applicable at level 2 of the ladder (interventional), they cannot answer counterfactual (level 3) because counterfactual queries involve knowing the noise and the deterministic function.\n\nIt has the strengths and weaknesses of applied categorical approaches: fairly elegant construction that seem to reveal profound structure in the problem of interest, the promise of lifting and porting properties from other areas into the problem of interests via the categorical abstraction. Yet, it appears difficult to find concrete, directly applicable findings. The criteria in section 6. are not easily convertible into operational algorithms.\nWhat does this newly presented derivation allow us to do? What applicable insights does it bring?\n\nI don't think the limitations described above are enough to reject the paper, because building such an understanding of the causal problems is important and might turn out to be useful. In particular, this specific paper on the equivalence of causal models also tackles one particularly important aspect: causal model abstraction which is very related to coarse-graining and emergence.\nRating: 7 accept, but needs minor improvements\nConfidence: The reviewer is fairly confident that the evaluation is correct\n", "role": "assistant" }, { "content": "Title: Reply to Reviewer 6giK\nThank you for the encouraging comment on the value of our work. \n\nAlthough the present manuscript focuses on discrete cases, we believe that our framework can be extended to continuous cases, given that the category of more general measurable spaces and the Markov kernels between them have categorical structures similar to Stoch. Exploring this possibility will be future work. \n\nWe would like to note that our categorical framework CAN separate function and noise, and thus address counterfactual (level 3) inferences. This is easily done by making all box deterministic functions and adding a “noise string” to each box (or more straightforwardly, by constricting a string diagram category from a graph that has explicit noise nodes). We did not mention this in the manuscript because it doesn’t change our theoretical results. \n\nRegarding the lack of operational algorithms, we’d like to stress that this is a proof-of-concept paper, with the view to formulating the concept of the “sameness” of causal structures in mathematically rigorous terms. This is important, because it tells us the identity of causal models — i.e., what we are actually trying to infer with causal discovery algorithms. Prior to any scientific measurement one must know what one is trying to measure. In the case of variables this is specified by identifying their scale, which amounts to specifying admissible transformations---i.e., natural transformations. In this paper we provided a similar criterion for causal models, and thus made it explicit what are those things that are to be discovered by several causal inference algorithms.\n\nMoreover, given our criterion, it is rather straightforward to come up with methods to check whether two models are equivalent. One could check the commutativity (def. 1) by some divergence measure (e.g. KL-divergence) or search coarse-grained models that satisfy causal homogeneity (def. 7). But due to a space constraint, we decided to focus on the conceptual side in this paper. \n", "role": "user" }, { "content": "Title: Reply to authors\nThank you to the authors for their answer.\n\nI think it is correct that the framework can accomodate counterfactuals (level 3). Having many deterministic functions might bring some possible simplifications and new problems with the noises that will be, in practice, unobserved.\n\nOverall, I am updating my score a little bit higher as I think the paper deserves to be accepted\n", "role": "assistant" } ]

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.