diff options
| -rw-r--r-- | README.md | 55 |
1 files changed, 49 insertions, 6 deletions
@@ -6,11 +6,9 @@ algorithmic hypothesis is narrow: a mixed apical compartment should teach from the component that is unexpected given the same neuron's ordinary somatic state, rather than from raw apical activity. -External A6000 collaborators running the complete 81-cell matched crossover -should begin with -[`COLLABORATOR_ONBOARDING.md`](COLLABORATOR_ONBOARDING.md). It contains the -method boundary, current positive and negative results, environment/data -setup, frozen matrix, restart policy, and the single entry-point command. +The archived 81-cell digital crossover packet is documented in +[`COLLABORATOR_ONBOARDING.md`](COLLABORATOR_ONBOARDING.md), including its +method boundary, completed cells, environment, and restart policy. For hidden population `l`, the implemented rule is @@ -21,6 +19,46 @@ r_l = a_l - a_hat_l Delta W_l = eta (r_l * local postsynaptic gain) h_(l-1)^T. ``` +## Current paper direction + +The current paper studies predictable error in a local teaching channel. Its +three experimental parts are: + +1. add the same innovation subtraction to digital Dual Propagation, + Equilibrium Propagation, standard coupled learning, and overclamped coupled + learning; +2. scale a digital CLLN from 32 to 2,048 learnable edges under fixed component + imperfection, comparing clean learning, same-RMS zero-mean noise, raw + imperfection, static calibration, and SDIL; +3. test the update in the nonlinear CLLN with published component-error scales, + nonideal local correlated-double-sampling, and a SPICE check of the local + sample/hold/subtract primitive. + +The current one-line method is + +```text +r = measured local teaching signal - instruction-off local prediction +Delta w = eta * r * local eligibility. +``` + +The predictor can be a fitted local function or a same-state neutral sample. +It reads no BP gradient, device-error constant, downstream weight, or global +task loss. Neutral measurements and their cost are counted. + +The completed transfer evidence is: Dual Propagation recovers from 9.40% to +82.92% validation accuracy against 82.86% clean; EP recovers from 31.38% to +74.52% against 76.26% clean; and the nonlinear 4-by-4 CLLN recovers from +25.86% mean classification error to 0%. The first fully confirmed digital +CLLN size contains 40 tasks and three device draws: at 32 edges, clean and +SDIL have 0% error, raw imperfection has 34.69%, same-RMS noise has 0.73%, and +static calibration has 6.15%. The remaining five sizes are running under the +frozen protocol. + +Start with [`THREE_PART_EVIDENCE.md`](THREE_PART_EVIDENCE.md) for the evidence +map and [`CLLN_SCALING.md`](CLLN_SCALING.md) for the frozen scaling protocol. +The current hardware-realistic figure is +[`figure_physical_hardware_evidence.pdf`](results/figs/figure_physical_hardware_evidence.pdf). + The standard-ResNet stability branch additionally uses a fast paired neutral observation to project any *remaining* affine neutral residual off current soma before plasticity. This local microphase is explicitly counted; it is not @@ -33,7 +71,12 @@ novel. The candidate contribution is the neutral-period, per-cell innovation operation under mixed apical traffic, together with its theory, causal audit, and scaling behavior. See `NOVELTY.md` for the exact prior-art boundary. -## Current audited evidence +## Previous clean-scaling program (retained audit trail) + +The results below remain reproducible and useful as baseline evidence. Clean +ResNet scaling in this program comes from the reciprocal Kolen--Pollack +backbone, so it is not used as evidence that residualization itself causes +scaling in the current paper. - On flattened CIFAR-10, SDIL loses only `0.214 +/- 0.349` accuracy points from hidden depth 5 to 60, while DFA's early-layer teaching alignment falls from |