Does learning at the fly's learning site close the gap to the simple formula?

EXPERIMENTS.md · lines 566–608

E15 — Does the plasticity help? Learning, frozen and shuffled-reward arms

Date 2026-09-12 · Question E12 left open whether KC->MBON plasticity - the point of the architecture - closes the gap to the composition baseline. · Provenance …/learning_test.py

Method. MushroomBody with the channel encoder and compartment readout, learning rate 0.05, six epochs over the 64 candidates in random order. Reward is the standardised negative mixing enthalpy (lower dH is better), minus a running baseline. Objective computed with MACE-MPA-0 averaged over 8 occupancies, spanning -100.7 to +181.4 meV/atom. Three arms: learning, frozen (no update), and shuffled rewards (reward paired with the wrong candidate).

Result.

Bar to beat: composition-only ridge, leave-one-out, 0.948 (E12).

Interpretation. The frozen arm is exactly unchanged, so the measurement is clean. The learning arm improves where the shuffled arm does not, with near-identical weight motion (5.236 against 5.239), so the rule is directionally real - it depends on the pairing between action and outcome, which is the minimum a learning rule must do.

But the effect is 0.01 against a gap of 0.37. Plasticity does not close it, and at this rate could not.

The pathology is the same one, and it survived every fix. Weight norm moves by 5.24 while the function moves by 0.01 - the same signature as the original null, where ‖dw‖ of 111 moved every selection probability by at most 0.0015. It has now survived a named readout, an anatomy-derived valence assignment, an element-channel encoder, and plasticity confined to the real learning site. Large weight motion with negligible functional change is the persistent finding of this project, not an artefact of any one design choice.

Caveats. One run per arm, so the 0.01 learning-minus-shuffled difference has no error bar and should not be called significant without repeats across seeds. Six epochs and one learning rate; no sweep was done. The objective is MACE-MPA-0's mixing enthalpy, which is largely a function of composition by construction (E12).