Does feeding the input to every sensory cell make the whole-brain search better?

MBON 95.0%     LAL 4.28%     DNa 0.72%     DNa02 5.1e-4
E161 (olfactory input, same readout):  MBON 99.62%   LAL 0.37%   DNa 0.011%   DNa02 1.2e-5

arm                                                        AUC_Q           finds          best
whole brain, olfactory input, learned readout   (E159)    2.61 +- 0.48    7.0 +- 2.0     -90
whole brain, afferent broadcast, learned readout (E163)  11.57 +- 8.87   29.0 +- 17.0   -111
CEM, 300 rounds                                  (E136)   13.57 +- 3.79   34.5 +- 8.5    -78

EXPERIMENTS.md · lines 10555–10582

E163 — Task 0, step 2b: the afferent broadcast as the brain's input

E157 measured the current olfactory entry engaging 2.7% of the brain and never reaching DNa02, and the FlyGM construction — one low-dimensional code broadcast to the whole afferent class — engaging 25% and firing 376 descending neurons in every family. E159 then showed that reading LAL + DNa with a learned readout lifts the whole brain from 0.15 to 2.61 AUC_Q even under the olfactory input. Step 2b changes the input: the composition is projected by a fixed random map onto all 17,479 afferent neurons (ORN, JO, LC, MeTu, OCG, lamina) instead of onto the 595 PNs through the channel code. Readout, plastic site and reward stay as in E159. FlyGM's encoder is trained; this first version is a fixed random projection, the reservoir choice, and that difference is named so it is not mistaken for the paper's.

Predicted, whole brain, E159's protocol (200 x 4 x 2, no novelty):

1. AUC_Q above E159's 2.61 and above 4.0 — a quarter of the brain now feeds a readout already shown to work on 2.7% of it.
2. The learned readout's weight share on LAL + DNa exceeds what E161 measures under olfactory input, since those populations are now driven.
3. The mushroom-body asset cannot run this encoder (it has no afferent class beyond PNs), so the control is E159's whole-brain arm, not the subset — stated so the comparison is not misread as whole-vs-subset.

Falsified if AUC_Q is at or below 2.61 — then engaging more of the brain adds nothing a readout can use, the fixed random projection is the suspect (FlyGM's is trained), and the next step is a trained encoder, not a bigger one. Falsified downward past E145's 0.15 if the broadcast destroys the olfactory code the KC→MBON plasticity depends on — a sparsity calibration failure, which the KC active fraction will show directly.

EXPERIMENTS.md · lines 10776–10782

E163, seed 0 readout (arm still running). With the composition broadcast onto 17,479 afferent neurons instead of the 595 PNs, the learned readout after 799 lessons carries:

MBON 95.0%     LAL 4.28%     DNa 0.72%     DNa02 5.1e-4
E161 (olfactory input, same readout):  MBON 99.62%   LAL 0.37%   DNa 0.011%   DNa02 1.2e-5

LAL + DNa's share rose from 0.4% to 5.0% — twelve-fold — and DNa02 from 1e-5 to 5e-4.

EXPERIMENTS.md · lines 10783–10785

E163 prediction 2 holds on seed 0: once the regions are driven, the delta rule puts weight on them. Whether that weight helps is prediction 1 (AUC_Q above 2.61 and above 4.0), which waits on the arm's second seed and its result row.

EXPERIMENTS.md · lines 10817–10838

E163 result — confirmed, with the spread stated as the headline. Whole brain, the composition broadcast onto 17,479 afferent neurons, learned LAL+DNa+MBON readout, E159's protocol (200 x 4 x 2, no novelty):

arm                                                        AUC_Q           finds          best
whole brain, olfactory input, learned readout   (E159)    2.61 +- 0.48    7.0 +- 2.0     -90
whole brain, afferent broadcast, learned readout (E163)  11.57 +- 8.87   29.0 +- 17.0   -111
CEM, 300 rounds                                  (E136)   13.57 +- 3.79   34.5 +- 8.5    -78

Prediction 1 confirmed: 11.6 is above 2.61 and above 4.0 — x4.4 on AUC_Q, x4 on finds, 21 meV/atom deeper. The error bar is the honest part: two seeds at roughly 2.7 and 20, so one seed matched E159 and the other quadrupled it. The mean is not a fluke of averaging — both seeds are at or above E159 — but the size of the gain is not yet a number; it is a range, and more seeds are the only way to narrow it. Prediction 2 confirmed on both seeds: LAL + DNa share 5.0% and 2.7% against 0.4% under olfactory input; DNa02 5e-4 and 7e-4. Prediction 3 stands as stated: the control is E159's whole-brain arm, not the subset.

With E161 and E163 together the story is now one sentence: the whole brain was silent because its input reached 2.7% of it, and once the sensory surface is driven the learned readout uses the regions the olfactory input never reached. This is the first result in which the whole brain beats every previous fly arm, and the first time "use the whole head" has a number attached that is not a deficit.