Skip to content

README.md

Latest commit

 

History

History
114 lines (81 loc) · 4.21 KB

README.md

File metadata and controls

114 lines (81 loc) · 4.21 KB

Signed Event Dynamics

Small experimental local learning system based on signed state-transition events.

This repository contains the public-facing v9.1 proof-of-concept from cap-flow. It learns on sklearn digits without backpropagation, using local credit assignment over sparse signed cells and state-transition events.

This is not a claim that the system replaces neural networks. The goal is narrower: document a reproducible local-learning mechanism and the diagnostic result that improved it.

Current Finding

Learning from errors is not enough;
fragile correct decisions near the border also need local reinforcement.

On sklearn.datasets.load_digits, v9.1 correct_border prob=0.80 improves over the clean v9 hard_scale baseline across 10 seeds:

mode final_acc peak_acc final_gap border_after_acc peak_drop
v9 full 0.9145 0.9209 0.0151 0.9164 0.0064
v9.1 correct_border 0.80 0.9469 0.9497 0.0282 0.9473 0.0028

See TECHNICAL_NOTE.md for the mechanism, diagnostics, limitations, and reproduction commands.

Install

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run

Cheap smoke check:

PYTHONPATH=lab python lab/v9_diagnostics.py --cells 12 --epochs 1 --seeds 1 --modes full,weights_only,event_only,correct_border --correct-border-prob 0.80

Focused confirmation chunk:

PYTHONPATH=lab python lab/v9_diagnostics.py --cells 64 --epochs 12 --seeds 2 --seed-start 0 --modes full,correct_border --correct-border-prob 0.80

Canonical 10-seed reproduction:

PYTHONPATH=lab python lab/v9_diagnostics.py --cells 64 --epochs 12 --seeds 10 --seed-start 0 --modes full,correct_border --correct-border-prob 0.80

This run can take a while on a laptop or small VM. On the capture machine it took roughly 13 minutes for both modes.

Expected canonical output is captured in RESULTS_v9_1_digits_10seeds.txt.

Exploratory second-dataset follow-up:

PYTHONPATH=lab python lab/v9_diagnostics.py \
  --dataset breast_cancer \
  --cells 32 \
  --epochs 6 \
  --seeds 10 \
  --seed-start 0 \
  --modes full,correct_border,correct_border_weights_only,correct_border_event_only \
  --correct-border-prob 0.80

This follow-up is public-reproducible but remains exploratory, not a benchmark claim. Expected output is captured in RESULTS_breast_cancer_10seeds.txt.

Mechanistic checks:

PYTHONPATH=lab python lab/test_v6_core.py
PYTHONPATH=lab python lab/test_v9_core.py

Repository Layout

lab/v4_core.py          shared data/loading helpers from the earlier transition-event line
lab/v6_core.py          signed event cell baseline
lab/v9_core.py          clean v9 hard-scale control
lab/v9_diagnostics.py   diagnostic runner and v9.1 correct_border mode
lab/cap_flow_v9.py      v9 configs and simple probe entry point
TECHNICAL_NOTE.md       public technical note for the v9 -> v9.1 finding
RESULTS_*.txt           captured canonical reproduction output

Limitations

  • This is a small sklearn digits proof-of-concept.
  • It does not establish broad scaling behavior.
  • It is not intended as a general ML replacement.
  • On digits, the short-horizon accuracy gain appears to be carried mostly by the plastic weight path; the event-score path is live but weaker under the current setup.
  • v10 is intentionally postponed until diagnostics justify a class-level tissue.

Current Open Questions

  • What does the event-score path contribute beyond plastic weight updates?
  • Does correct_border remain positive across additional splits or small datasets?
  • When the gain appears beyond digits, does it come from near-border samples or from wider non-border decision-field changes?

A reproducible exploratory breast_cancer follow-up is included in RESULTS_breast_cancer_10seeds.txt. It is not a benchmark claim and does not replace the canonical digits reproduction; see TECHNICAL_NOTE.md.

Status

v9   = clean hard_scale baseline
v9.1 = confirmed frontier candidate
v10  = postponed