vit-beans-v3

Geometric Deep Learning with Cantor Multihead Fusion + Shatter-Reconstruct Training

This repository contains training runs using Cantor fusion architecture with:

• Pentachoron (5-simplex) structures for geometric routing
• CosineAnnealingWarmRestarts for exploration cycles
• GeometricCoalescenceLoss for shatter-reconstruct training

🚀 LR Boost + Geometric Coalescence

This run uses restart_lr_mult = 1.15x with GeometricCoalescenceLoss:

• LR boosts create aggressive exploration cycles
• Coalescence loss provides geometric scaffolding during weight thrashing
• Adaptive weighting: 0.1 → 0.8 during LR spikes
• Model reconstructs from geometric first principles when patterns shatter

Current Run

Latest: cifar100_weighted_ADAMW_WarmRestart_boost1.15x_coal0.5_20251124_152227

• Dataset: CIFAR100
• Fusion Mode: weighted
• Coalescence: λ=0.5 ✓
• LR Boost: 1.15x 🚀

---

Repository maintained by: @AbstractPhil