Skip to content

Latest commit

 

History

History
205 lines (136 loc) · 11.4 KB

File metadata and controls

205 lines (136 loc) · 11.4 KB

Encephagen: Research Development Report — Day 5

Date: April 5, 2026 Author: edvatar (toroleapinc) Repo: https://github.com/toroleapinc/encephagen


1. Summary

Day 5 executed all 4 phases of the revised plan, hit a fundamental wall, and built a working demo anyway.

Phase A (T1w/T2w gradient): Timescale hierarchy emerges (r=-0.45, p<0.0001) but is identical for connectome and random — it's driven by the gradient, not topology.

Phase B (Innate dynamics): Stimulus doesn't propagate beyond visual cortex. The wall: all dMRI-derived long-range connections are excitatory. At high coupling → saturation. At low coupling → no propagation. Added feedforward inhibition (3x stronger onto inhibitory neurons) which improved FC-FC to 0.30+ but didn't fix propagation.

Phase C (Learning scaffold): Both connectome and random at chance (~36%) on stimulus-action task. E-prop doesn't produce meaningful learning at this scale with 60 episodes.

Phase D (Demo): Built python demo.py — interactive brain with 16K neurons, 80 regions, T1w/T2w gradient, conduction delays, CPG walking, stimulus response. It works.


2. Phase A: T1w/T2w Timescale Gradient

Implementation: Downloaded HCP S1200 T1w/T2w myelination map, parcellated onto 80 AAL2 cortical regions. Tau_m ranges from 10ms (Heschl/auditory) to 30ms (temporal pole).

Result: Timescale-rate correlation r=-0.445 (p=3.5e-5). Regions with shorter tau_m fire faster. Regional CV increased 24% (0.082 vs 0.066). FC-FC(emp) maintained at 0.43.

But: Hierarchy is IDENTICAL for connectome and random (both r≈-0.45). The gradient drives hierarchy, not topology. This means regional heterogeneity is necessary but not sufficient — topology still doesn't differentiate.


3. Phase B: Innate Dynamics — The Wall

Test: Flash 100mV visual stimulus, measure response propagation to temporal, parietal, frontal regions.

Result: Visual cortex responds (+12% at gc=5, +10% at gc=10). All other regions show 0% change. Signal is TRAPPED in visual cortex.

Root cause: All between-region connections are excitatory (dMRI can't distinguish neurotransmitter type). At high coupling, every region is driven to saturation by global excitation — no room for stimulus-evoked changes. At low coupling, the stimulus increment is too small relative to local dynamics.

Feedforward inhibition fix: Made long-range connections 3x stronger onto local inhibitory neurons. This improved FC-FC (gc=5 → 0.304, passes benchmark!) and prevented saturation. But stimulus STILL doesn't propagate — the feedforward inhibition that prevents saturation also blocks signal cascade.

Fine temporal analysis: At 2ms resolution, the network shows ~50Hz oscillatory fluctuations. But single-trial evoked responses are buried in spontaneous noise.

This is the fundamental dMRI wall: Without inhibitory long-range connections (which require neurotransmitter identity that dMRI doesn't provide), you can't have BOTH balanced dynamics AND stimulus propagation.


4. Phase C: Learning Scaffold

Task: 3-way stimulus-action mapping with delayed reward. 60 episodes, e-prop learning.

Result: Connectome 36% vs Random 37% (chance=33%). p=0.47. Neither brain learns. 0/7 FDR-significant.

Interpretation: The task is too hard for e-prop at this scale. Can't compare learning speed when neither agent learns. Need either simpler task, more episodes, or stronger learning rule.


5. Phase D: Interactive Demo

Built demo.py with full pipeline:

  • 16,000 spiking neurons, 80 regions
  • T1w/T2w timescale gradient (10-30ms)
  • HCP structural connectivity with conduction delays
  • Feedforward inhibition for balanced dynamics
  • Visual/auditory/somatosensory stimulation
  • CPG walking modulated by motor cortex
  • FC-FC(empirical) = 0.28-0.43 depending on gc

Commands: look, sound, touch, walk, status, rest, quit


6. Cumulative Findings (31 Experiments)

What works:

  1. SC-FC validation — simulated FC matches real fMRI at r=0.42 (neurolib80, gc=10)
  2. Timescale hierarchy — T1w/T2w gradient produces Murray et al. 2014 hierarchy
  3. Feedforward inhibition — prevents saturation, improves FC-FC
  4. Interactive demo — brain responds to stimuli, modulates CPG
  5. E-prop learning — produces weight changes, but doesn't solve cognitive tasks at this scale

What doesn't work:

  1. Structural advantage — 0/4 significant on any cognitive measure with validated dynamics (Exp 29)
  2. Stimulus propagation — signal trapped in stimulated region, doesn't cascade
  3. Learning — e-prop at 16K neurons with 60 episodes doesn't learn 3-way classification
  4. Connectome as learning scaffold — no learning speed difference

The wall:

dMRI tractography provides excitatory-only, undirected, macro-scale routing. Without:

  • Inhibitory long-range connections (needs neurotransmitter ID)
  • Directed connections (dMRI is undirected)
  • Synaptic-level specificity (dMRI resolves regions, not neurons)

...you cannot build a connectome-dominant model with differentiated dynamics.


7. What's Next

Option 1: Work around the dMRI wall

Add ESTIMATED inhibitory long-range from neuroanatomy literature. Not just BG→thalamus, but:

  • ~30% of between-region connections as inhibitory (from known long-range GABAergic projections)
  • Thalamic reticular → thalamus (requires subcortical parcellation)
  • Cortical SST+ long-range inhibition This requires switching to a parcellation WITH subcortical regions AND continuous weights.

Option 2: Multi-scale approach

Use Wilson-Cowan at regional level (which produces realistic SC-FC and propagation) coupled with spiking neurons in select regions (PFC for working memory, V1 for perception). This is the Arbor-TVB approach.

Option 3: Switch to a task where structure already helps

BT-SNN (Zhao et al. 2024) showed connectome topology helps for RL tasks when regions have different properties. Adapt their approach: use the connectome as architecture for a spiking RL agent on MuJoCo tasks, with region-specific tau_m from T1w/T2w.

Option 4: Developmental approach

Start with noisy connectome, let STDP/e-prop refine it. Compare final learned connectivity to real connectome. If learning converges toward the real pattern, that proves the connectome is an attractor of learning dynamics.


8. Files Changed

File Change
src/encephagen/gpu/spiking_brain_gpu.py T1w/T2w gradient, feedforward inhibition
src/encephagen/connectome/bundled/neurolib80_t1t2_*.npy HCP T1w/T2w myelination data
experiments/30_innate_dynamics.py Stimulus propagation + oscillation tests
experiments/31_learning_scaffold.py Learning speed comparison
demo.py NEW: Interactive brain demo


9. Experiment 32: The Newborn (Closed-Loop Brain-Body)

Goal: Replicate what the worm/fly projects did — innate animalistic behavior from structure alone. A newborn doesn't think, it MOVES: reflexes, spontaneous movement, righting response.

Setup: Brain→CPG→Body→Brain closed loop. Inverted pendulum physics with random perturbations. Somatosensory input (tilt, velocity) → brain → motor cortex → CPG modulation → hip torques → body.

Result: All 20 runs (10 connectome, 10 random) survive full 10s. Identical metrics: tilt_std=0.010, torque_diversity=0.016. 0/5 significant.

Why: The body physics is too stable — the damped pendulum converges to equilibrium regardless of brain activity. The CPG locks to constant output instead of oscillating. The brain's influence is too small to matter.

Lesson: Need a properly UNSTABLE body (MuJoCo Walker2d) where the brain must actively balance, not a self-stabilizing pendulum. The worm's body is inherently unstable in fluid — that's what makes the connectome's motor patterns matter. Our pendulum is passively stable.


10. Summary: 32 Experiments, 5 Days

Experiment Finding Significant?
1-4 Wilson-Cowan: degree→hierarchy, wiring→FC Yes (network science)
5 Spiking hierarchy matches Murray 2014 Yes (but WC disagrees)
6 STDP habituation Yes
7-9 Embodied learning INVALIDATED (motor death) Negative
10 Pendulum learning: 5 approaches failed Negative
11 Spontaneous body twitching Descriptive
12-13 Brain+CPG+body: 0.98 Hz gait Yes
14 Crawling worm Yes
15-18 Cognitive functions (condition, discrim, memory, integrated) Yes (but Hebbian)
19-20 Walker2d body control Marginal
21 Connectome vs random (Hebbian) CV only (p=0.0002)
22 Connectome vs random (e-prop) Conditioning (p=0.011, parameter-dependent)
23 Discrimination analysis Entropy artifact (d=-9)
24 Full biophysical model ALIF reverses advantage
25 SC-FC validation FAILS at gc=0.15 Exposed bad parameters
26 SC-FC tuning gc=0.20 passes (r=0.388)
27 Validated connectome vs random (tvb96) 1/5 structure, 3/5 random wins
28 tvb66 tuning WIP
29 Validated (neurolib80) connectome vs random 0/4 significant
30 Innate dynamics: stimulus propagation Signal doesn't propagate (dMRI wall)
31 Learning scaffold Neither learns (both at chance)
32 Newborn closed-loop Body too stable to differentiate
33 Walker2d brain control Brain helps (210 vs 119 baseline) but no structural advantage

Net result: 0 significant structural advantages on validated dynamics. The macro-scale dMRI connectome does NOT provide measurable cognitive, learning, or behavioral advantage over random wiring.

Root causes identified:

  1. All-excitatory long-range connections (dMRI can't distinguish E/I)
  2. Regions are interchangeable (tau_m gradient helps but doesn't differentiate topology)
  3. E-prop doesn't learn at this scale (task too hard)

What works: SC-FC validation (r=0.42), timescale hierarchy (r=-0.45), feedforward inhibition, CPG locomotion, interactive demo, brain-controlled Walker2d (78% improvement over baseline).


11. Experiment 33: Brain-Controlled MuJoCo Walker2d

Goal: Wire the heterogeneous brain to a properly unstable body. BT-SNN approach: connectome as RL architecture. Does structure help motor control?

Setup: 16K neuron brain (T1w/T2w gradient, neurolib80, delays, feedforward inhibition) → sensory input from Walker2d obs (17-dim) → somatosensory regions → brain routing → motor cortex (6 groups) → 6 joint torques. E-prop learning from reward. 5 runs × 20 episodes per condition.

Results:

Metric Connectome Random Baseline p
Mean survival (steps) 212.5 209.7 119 0.15
Mean reward 332.5 329.5 0.42

0/5 FDR-significant. Both brains keep Walker2d alive 78% longer than zero-action baseline. The brain IS controlling the body — processing sensory input and producing useful motor output. But the connectome provides no advantage over random wiring.

Key finding: The spiking brain with T1w/T2w gradient CAN control a MuJoCo body. This is the "animalistic behavior" we were targeting — the brain produces movement that keeps an unstable biped upright. But it works equally well with any wiring pattern.


Day 5. 33 experiments. The brain is alive and controls a body (210 vs 119 baseline). The connectome provides the scaffold but not the specificity. At macro-scale dMRI, random wiring works just as well.