Does feeding every sensory cell wake more of the fly brain than smell alone?

encoding                  driven   active   outside MB   DNs active   DNa02   DN families
pn (current model)           808     2.7%       1.4%          28       no     7
orn:100                      100     0.3%       0.2%           1       no     1
orn:300                      300     0.5%       0.4%           2       no     1
orn:1000                   1,000     1.6%       1.2%           5       no     2
orn:all                    2,635     4.2%       2.9%           8       no     4
multi:100  (+JO +MeTu)     1,781     2.7%       2.7%         203       no     9
multi:1000                 2,681     4.0%       3.7%         206       no    10
multi:all                  4,316     6.6%       5.4%         211       no    10
afferent:all (broadcast)  17,479    25.1%      24.6%         376      YES    all

EXPERIMENTS.md · lines 10148–10184

E157 — How much input does the brain need? Active fraction against sensory drive

The operator's directive is that the whole head learns and forms its own latent states, and that how much of it engages depends on how much input it is given. Before any learning rule is wired brain-wide, that dependence has to be measured, because it decides the encoder.

Model. Shiu et al. (Nature 2024) propagate activity through the whole connectome with weights = synapse count x transmitter sign and one free gain, driving sparse sensory populations, and predict descending-neuron output at 91%. This uses the same construction at rate level rather than spike level: h_{t+1} = relu(g W_signed h_t + drive), W_signed from the asset's contacts and the source neuron's transmitter (acetylcholine +, GABA / glutamate / histamine -, modulators 0 for this purpose), g set so the largest eigenvalue is below one, drive on the chosen sensory neurons only, 8 steps. A neuron is "active" if its rate exceeds 1% of the maximum driven rate; a DN "fires" if active. Nothing here learns.

Sensory encodings compared, same composition, same gain:

A   the current model's input: 12 channels onto the 595 PNs (what MushroomBody does)
B   k olfactory receptor neurons (ORN, 2,635 total), k = 100, 300, 1,000, all
C   B plus mechanosensory JO (672) and compass input MeTu (1,009) carrying ladder state

Predicted:

1. Under A, fewer than 5% of the 164,506 neurons are ever active and the DNs that fire are those two hops from the mushroom body only.
2. Active fraction rises monotonically with k under B, and at k = 1,000 exceeds 30%.
3. Under C, DNs fire in every descending family (DNa, DNb, DNd, DNg, DNp), including DNa02, which under A never fires.
4. The curve saturates: doubling k past ~1,000 adds less than 5% more active neurons — the brain's capacity to be engaged by sensory drive has a knee, and that knee is the input size the encoder must reach.

Falsified if the active fraction under A is already above 20% — then the brain was engaged all along and the earlier "loaded, not used" reading was about the readout alone, not the dynamics; or if the curve under B is flat in k, which would mean propagation is governed by the recurrent gain and not by how many inputs are driven, and the encoder is not the lever.

EXPERIMENTS.md · lines 10186–10214

E157 addendum, before the run — three papers settle the encoder question and the design.

FlyGM (arXiv:2602.17997) does not assign features to neurons. Its observation — 741-d for walking, 104-d for flight — is compressed to a 32-d code and broadcast to every afferent neuron through a trainable gate; efferent = FlyWire's output flow class, decoded by an MLP; each neuron carries a trainable descriptor, one shared update MLP; imitation from an MLP expert then PPO; functional segregation into sensory / central / motor emerged inside. "How much input the fly needs" is therefore not a count of driven neurons: it is a low-d code reaching the whole afferent class, with the connectome doing the expansion.

Lappalainen et al. (Nature 634, 1132, 2024): 45,669 neurons, 1,513,231 connections, 734 free parameters — time constant and resting potential per cell type (65 each), one unitary synapse strength per type pair (604) — task-optimized on optic flow with threshold- linear differentiable dynamics; matched 26 studies. Per-type parameters suffice, and the strategy works best when connectivity is sparse, which it is here.

Costi, Hadjiivanov, Dold, Hale & Izzo (Biomimetics 10, 341, 2025): the entire fly connectome as a reservoir with a trained readout is markedly more overfitting-resilient than standard reservoirs and reaches under 2% error at lower regularisation; both the topology and the synaptic weights contribute (Morra & Daley 2022: 100x lower variance).

Design, revised: encoder = FlyGM-style broadcast of a ~32-d code (composition + ladder state) onto the whole afferent class; dynamics = signed connectome operator with per-type trainable gain and time constant (Lappalainen); readout = trained linear map over all neurons or all descending neurons (Costi); learning = the ladder's reward through policy gradient on the per-type parameters and the readout, plus three-factor plasticity at the modulator targets. E157 adds encoding D — broadcast to the full afferent class — and the prediction for it: the largest active fraction of all encodings, and DNs firing in every family. Predictions 1-4 stand.

EXPERIMENTS.md · lines 10296–10333

E157 result — how much input the brain needs, measured. Rate-level Shiu propagation on the whole asset, spectral radius 3,749, g = 2.4e-4 (rho 0.9), 8 steps, unit drive on the chosen neurons, active = above 1% of the largest driven rate.

encoding                  driven   active   outside MB   DNs active   DNa02   DN families
pn (current model)           808     2.7%       1.4%          28       no     7
orn:100                      100     0.3%       0.2%           1       no     1
orn:300                      300     0.5%       0.4%           2       no     1
orn:1000                   1,000     1.6%       1.2%           5       no     2
orn:all                    2,635     4.2%       2.9%           8       no     4
multi:100  (+JO +MeTu)     1,781     2.7%       2.7%         203       no     9
multi:1000                 2,681     4.0%       3.7%         206       no    10
multi:all                  4,316     6.6%       5.4%         211       no    10
afferent:all (broadcast)  17,479    25.1%      24.6%         376      YES    all

Prediction 1 confirmed: the current entry point engages 2.7% of the brain. Prediction 2 falsified: olfaction is monotone in k but reaches 1.6% at a thousand receptors and 4.2% with every one of them — nowhere near 30%. The olfactory pathway cannot engage this brain, however hard it is driven; it is a narrow channel into the mushroom body and the lateral horn and little else. Prediction 3 falsified on its named cell: JO + MeTu raise descending activity from 8 to ~210 neurons — mechanosensation reaches DNs in one hop, exactly as E156's edge counts said — but DNa02 does not fire under any olfactory-plus encoding. Prediction 4 not assessable: no knee, at a level that does not matter. Encoding D confirmed on both counts: driving the whole afferent class engages a quarter of the brain, fires 376 descending neurons in every family, and is the only encoding that reaches DNa02.

The answer. "How much input the fly needs" is not a number of olfactory neurons. It is the sensory surface: the FlyGM construction — one low-dimensional code broadcast to every afferent neuron through a gate — is the encoder, and the current 12-channel olfactory entry is why 97% of the brain was silent. Task 0 step 2's input side is decided.

Caveat, pre-registered before the check: the absolute fractions depend on the recurrent gain; at rho 0.9 the network is well sub-critical and 8 steps is shallow. The claim is the ordering pn < orn < multi < afferent and DNa02 reachable only by the broadcast. Predicted at rho 0.99: the ordering holds, the fractions rise by less than a factor of three, and DNa02 still fires only under afferent:all. Falsified if olfaction alone reaches DNa02 at rho 0.99, which would make the gain, not the encoder, the lever.

EXPERIMENTS.md · lines 10335–10339

E157 rho check — confirmed. At rho 0.99 (g = 2.6e-4 against 2.4e-4): pn 2.5%, orn:all 4.6%, multi:all 7.1%, afferent:all 26.1%; DN counts 25 / 10 / 217 / 391; DNa02 fires only under the broadcast, as before. The ordering holds, every fraction moved by less than 10% — well inside the "less than a factor of three" I allowed — and olfaction alone still does not reach the steering neuron. The recurrent gain is not the lever. The encoder is.