Was the learning rule the reason the brain failed to choose, or would backpropagation do better?

EXPERIMENTS.md · lines 13–39

E1 — Does the learning rule matter? Backprop against Forward-Forward

Date 2026-09-11 · Question Is the null result of the three published runs a credit-assignment failure? · Provenance forager/baselines.py::BackpropGraph, config configs/rhea_wholebrain_conditions_bp_a100.json

Method. BackpropGraph differentiates the same per-neuron objective through the three settling steps with autograd, so derivatives flow between neurons within a decision; LocalFF holds presynaptic activity fixed and does not. Two fields differed from the published flagship run — method and run_name. Graph, candidates, seed (17), encoder, softmax temperature, exploration rate and budgets were identical. Run on a Colab A100.

Result.

• Update directions align at cosine 0.9916 at whole-brain scale, ‖dw‖ 28.80 (FF) against 28.86 (BP), measured on 163,972 nodes / 6,143,838 edges.
• The live run reproduced the Forward-Forward screening sequence in 39 of 40 picks. The only difference was position 38: HfTaTi-033 against HfNbTaW-034.

Interpretation. The null is not a credit-assignment failure. Gradients flowing between neurons change one candidate in forty, which with 15% uniform exploration is indistinguishable from no change. The damage is upstream of the learning rule.

Caveat. The run's exported artifacts were lost when the Colab session handle expired; the screening sequence was read from the live job log. The 10 DFT numbers it would have produced were 16-atom cells, whose value is limited by E3.