Does the fly brain beat a simple straight-line formula on the eight element fractions?

EXPERIMENTS.md · lines 405–444

E12 — The control that sets the bar: composition-only regression

Date 2026-09-12 · Question Does the connectome beat a trivial baseline? · Provenance …/baseline_test.py

Method. Ridge regression from the eight element concentrations to the same objective used in E11, evaluated leave-one-out so the baseline cannot memorise the 64 points. Repeated with pairwise interaction terms added - still composition-only, no structure, no connectome.

Result.

Interpretation. An eight-parameter linear model beats the entire 166,700-neuron, 25.6-million-edge circuit by a wide margin, and adding pair terms widens it further.

E10 and E11 stand as stated - the measured wiring is in the causal path and does beat a rewired control - but they are a comparison between two things that are both worse than trivial. Chapter 11 anticipated precisely this: "if a ridge regression from standardised features already picks well, the recurrent network has to beat that and not uniform sampling." It does not.

What this does and does not rule out.

• It does not show the connectome is useless: the circuit here is untrained, and the whole point of the architecture is that KC->MBON plasticity should improve it. The learning loop has not been run under the new readout.
• It does show that the bar is 0.95, not zero, and every future claim must clear it. Beating uniform sampling, or beating a rewired control, is not evidence of value while a linear model on eight numbers does better than both.
• The objective is MACE's mixing enthalpy, which is largely a function of composition by construction. A composition-only model is therefore expected to do well on this target. A harder target where composition is not near-sufficient would be a fairer test - and choosing one is itself a design decision to make deliberately rather than by accident.