metadata

tags:
  - brain-inspired
  - spiking-neural-network
  - biologically-plausible
  - modular-architecture
  - reinforcement-learning
  - vision-language
  - pytorch
  - curriculum-learning
  - cognitive-architecture
  - artificial-general-intelligence
license: mit
datasets:
  - mnist
  - imdb
  - synthetic-environment
language:
  - en
library_name: transformers
widget:
  - text: >-
      The first blueprint and the bridge to Neuroscience and Artificial
      Intelligence.
  - text: I’m sure this model architecture will revolutionize the world.
model-index:
  - name: ModularBrainAgent
    results:
      - task:
          type: image-classification
          name: Vision-based Classification
        dataset:
          type: mnist
          name: MNIST
        metrics:
          - type: accuracy
            value: 0.98
      - task:
          type: text-classification
          name: Language Sentiment Analysis
        dataset:
          type: imdb
          name: IMDb
        metrics:
          - type: accuracy
            value: 0.91
      - task:
          type: reinforcement-learning
          name: Curiosity-driven Exploration
        dataset:
          type: synthetic-environment
          name: Synthetic Environment
        metrics:
          - type: cumulative_reward
            value: 112.5

🧠 ModularBrainAgent: A Brain-Inspired Cognitive AI Model

ModularBrainAgent (SynCo) is a biologically plausible, spiking neural agent combining vision, language, and reinforcement learning in a single architecture. Inspired by human neurobiology, it implements multiple neuron types and complex synaptic pathways, including excitatory, inhibitory, modulatory, bidirectional, feedback, lateral, and plastic connections.

It’s designed for researchers, neuroscientists, and AI developers exploring the frontier between brain science and general intelligence.

🧩 Model Architecture

Total Neurons: 66
Neuron Types: Interneurons, Excitatory, Inhibitory, Cholinergic, Dopaminergic, Serotonergic, Feedback, Plastic
Core Modules:
- SensoryEncoder: Vision, Language, Numeric integration
- PlasticLinear: Hebbian and STDP local learning
- RelayLayer: Spiking multi-head attention module
- AdaptiveLIF: Recurrent interneuron logic
- WorkingMemory: LSTM-based temporal memory
- NeuroendocrineModulator: Emotional feedback
- PlaceGrid: Spatial grid encoding
- Comparator: Self-matching logic
- TaskHeads: Classification, regression, binary outputs

🧠 Features

🪐 Multi-modal input (images, text, numerics)
🔁 Hebbian + STDP local plasticity
⚡ Spiking simulation via surrogate gradients
🧠 Biologically inspired synaptic dynamics
🧬 Curriculum and lifelong learning capability
🔍 Fully modular: plug-and-play cortical units

📊 Performance Summary

Note: Metrics shown below are for illustrative purposes from synthetic and internal tests.

Task	Dataset	Metric	Result
Digit Recognition	MNIST	Accuracy	0.98
Sentiment Analysis	IMDb	Accuracy	0.91
Exploration Task	Gridworld Simulation	Cumulative Reward	112.5

💻 Training Data

MNIST: Handwritten digit classification
IMDb: Sentiment classification from text
Synthetic Environment: Grid-based exploration with feedback

🧪 Intended Uses

Use Case	Description
Neuroscience AI Research	Simulating cortical modules and spiking dynamics
Cognitive Simulation	Experimenting with memory, attention, and decision systems
Multi-task Agents	One-shot learning across vision + language + control
Education + Demos	Accessible tool for learning about bio-inspired AI

⚠️ Limitations

Early-stage architecture (prototype stage)
Unsupervised/local learning only (no gradient-based finetuning yet)
Synthetic data only for now
Accuracy and metrics not benchmarked on large-scale public sets

✨ Credits

Built by Aliyu Lawan Halliru, an independent AI researcher from Nigeria.
SynCo was created to demonstrate that anyone, anywhere, can build synthetic intelligence.

Almusawee
/

ModularBrainAgent