Skip to content

phase2_homeostasis_Claude_memory.md

Latest commit

 

History

History
39 lines (30 loc) · 2.39 KB

phase2_homeostasis_Claude_memory.md

File metadata and controls

39 lines (30 loc) · 2.39 KB

Phase 2 Working Memory: phase2_homeostasis

Fresh Start — Block 1

Context

Previous exploration (13 blocks, 104 iterations) used a supervised contrastive loss that read ground-truth metabolite type labels (x[:, 6]) during training. This constituted label leakage: embeddings separated because they were told the correct types, not because MLP_node discovered type-dependent regulation. The supervised contrastive loss has been removed from the code.

What was learned (retaining legitimate findings only)

Signal characteristics:

  • Homeostatic signal is ~1000x weaker than reaction dynamics
  • MLP_node starts at zero with near-zero gradients; standard LRs (1E-3) insufficient
  • lr_node_homeo must be higher than lr_emb_homeo (MLP_node must learn first)

Time step behavior (WITHOUT contrastive loss — from Blocks 1-2 before contrastive was added):

  • time_step=4: too short for homeostatic signal to accumulate
  • time_step=16: marginal signal
  • time_step=32: best balance of signal accumulation vs gradient noise (Block 1: 0.38)
  • time_step=64: strongest signal but noisier gradients, highly stochastic

Training strategies (legitimate):

  • Signal amplification (10x): makes homeostatic gradient comparable to reaction gradient
  • Offset penalty: suppresses constant-output solutions, forces slope learning
  • Gradient accumulation (4x): reduces variance from single-rollout BPTT
  • Gradient clipping: stabilizes long-rollout BPTT
  • Kaiming re-initialization of hidden layers: enables gradient flow through Tanh

Key open problem:

  • Without supervised contrastive loss, embedding separation depends entirely on MLP_node producing type-differentiated outputs. If MLP_node learns a single average slope for all metabolites, there is no gradient signal to push embeddings apart. This chicken-and-egg problem is the core challenge.

Established Principles

  • DO NOT use ground-truth labels (x[:, 6]) during training — label leakage
  • Embeddings must self-organize from learned MLP_node behavioral differences only
  • The previous "best results" (80-92% avg_slope_ratio) were achieved with supervised contrastive loss and are not valid baselines

Block 1 Plan

  • Start with legitimate strategies only: amplification + offset penalty + gradient accumulation + gradient clipping
  • Sweep lr_node_homeo across slots at different time_steps
  • Establish unsupervised baselines before attempting any auxiliary losses