PaperlayerView exact revision on arXiv

Source-linked AI summary

Sensorimotor behaviour from an untrained brain-and-nerve-cord connectome

John R. Minnick, Jesus Gonzalez-Ferrer, Jinghui Geng, Jason K. Eshraghian, Mohammed A. Mostajo-Radji, David Haussler, Mircea Teodorescu

arXiv:2610.33881v1q-bio.NCcs.NE

TL;DR

How much behaviour follows from a connectome’s wiring remains an open question, and testing a nervous system as a controller requires a body and sensory feedback. The authors run an untuned spiking fly connectome in isolation and in a simulated body, where it navigates to odour sources and learns odour avoidance through synaptic depression.

  • Problem

    How much nervous-system function follows from wiring remains open, and the contribution of measured wiring in an untrained, embodied sensorimotor loop is unknown.

    Source: Introduction · p. 2

  • Method

    The authors build a 172,433-neuron spiking network from the BANC connectome with published model parameters unchanged, then test it alone and inside a simulated body using fixed sensory and motor coupling rules.

    Source: Introduction · p. 2, The connectome as a spiking network… · p. 2, Discussion · p. 9

  • Results

    The embodied network reached 30 of 32 odour sources versus 5.8 for a rewired network, while replayed commands reached 2% when sensory feedback was removed.

    Source: Discussion · pp. 9–10

  • Takeaways & Limitations

    Dopamine-gated depression of mushroom-body output synapses turns odour approach into avoidance, and associative depression depends on Kenyon-cell activity being odour-dependent.

    Source: Introduction · p. 2, Discussion · p. 11

    The embodied claims are scoped by two hand-written coupling rules with fixed constants, including a steering rule that reads valence as the balance between approach- and avoidance-driving MBONs.

    Source: Discussion · p. 10

Abstract

from arXiv · show

Connectomes now map the neurons and synapses of an adult fly's brain and nerve cord, but how much of behaviour follows from the wiring is unknown. Here we turn one such connectome into a spiking network of 172,433 neurons with a published brain model's untuned parameters, and run it alone and inside a simulated body. Two fixed rules join them: odours and looming objects drive sensory neurons, and motor neurons move the legs. Alone, the network passes a signal to the motor neurons it is known to drive when the pathway supplies about 2% or more of their input. In the body, the fly navigates around obstacles to 95% of the odour sources. A rewiring that keeps each neuron's connection counts and signs reaches 18%, and the fly's leg commands replayed with the senses off reach almost none, so navigation needs the measured wiring and sensory feedback. Dopamine-gated weakening of the mushroom body's output synapses turns approach to an odour into avoidance, which requires the pairing and is odour-specific once one olfactory synapse class is scaled. An untrained connectome thus supports a behaviour and its learned reversal, and the pathways it cannot drive mark what the diagram leaves out.

Introduction

The paper asks how much behaviour follows from measured neural wiring, especially in a closed sensorimotor loop, and tests an untuned fly connectome alone and in a body. The embodied network navigates to odour sources and can learn to avoid an odour.

Introduction

  • Earlier fly models reproduced heading dynamics [21], predicted feeding and grooming drivers [22], or identified walking circuitry [23], but wiring-to-function evidence remains limited.Other work found recurrent dynamics can be underdetermined [24, 20], and measured signal propagation can depart from anatomical predictions [25].Source: Introduction · pp. 1–2
  • Known behavioural roles—from descending-neuron actions and steering to looming-triggered escape and mushroom-body olfactory learning—provide tests of whether wiring reproduces biology [26, 27, 28, 29, 30, 31, 32, 33, 34].The paper tests pathways whose behavioural roles are established experimentally.Source: Introduction · p. 2
  • A closed-loop test is needed because an animal’s movements alter its sensory input, making behaviour an interaction among nervous system, body, and environment [35, 36].The paper asks what measured brain-and-nerve-cord wiring contributes when run untrained in a body.Source: Introduction · p. 2
  • 95% of 32 odour sources were reached by the embodied network, versus 18% after degree-and-sign-preserving rewiring and 2% with replayed motor commands without sensory feedback.The results link navigation to the measured wiring and sensory feedback.Source: Introduction · p. 2
  • Dopamine-gated depression of mushroom-body output synapses shifts odour approach to avoidance, with associative depression depending on Kenyon-cell activity that depends on the odour.The trained fly turns away from the odour.Source: Introduction · p. 2
  • The study builds a 172,433-neuron spiking network from the BANC connectome and keeps every parameter of the published brain model at its original value.The network is tested outside the body, in a simulated body, and during learning.Source: Introduction · p. 2

Results

The untrained connectome navigated to most odour sources, while rewiring or removing sensory feedback sharply reduced success. Its responses also reproduced some known pathways, and dopamine-gated changes to mushroom-body output supported learned avoidance under the tested conditions.

The connectome as a spiking network inside a body

  • 358 of 576 ordered region pairs had synapses, but only 139 of 552 pairs responded at 0.1 Hz per neuron or more when a region was driven.The response map followed excitatory-signed synapse counts across connected region pairs (ρ = 0.65).Source: The connectome as a spiking network… · p. 3
  • Seven of 15 tested pathways conducted, failed under degree-and-sign-preserving rewiring, and carried the sign measured in the animal.Every tested pathway supplying at least 2% of its target's input conducted at some drive; the seven matching pathways were sugar, bitter, LPLC2, LC4 with LPLC2, DNa02, PFL3 and DNg12.Source: Pathways of known behavioural role… · p. 4

The antennal lobe as read is bistable, and one synapse class sets the Kenyon-cell code

  • 68% of Kenyon cells were active for either odour as read, but scaling cholinergic local-neuron output to projection neurons produced a 6.8% median active share across eight odours.At gain 0.15, the code fell within the fly's reported 5–10% response range and the overlap was under 0.5.Source: The antennal lobe as read is bistable,… · p. 5
  • Odour generalisation tracked Kenyon-cell code overlap and was complete by three shared glomeruli; even disjoint odours showed 0.13 to 0.28 of the depression.In the matched variant, the overlap rose from 0.26 to 0.72 as shared glomeruli increased from zero to four.Source: The antennal lobe as read is bistable,… · p. 6
  • 30.25 of 32 odour sources (95%) were reached by the intact fly, compared with 5.80 (18%) by the rewired fly.Across 180 runs, the intact fly reached every source in 161 runs; no rewired fly reached as many sources as any intact fly.Source: The connectome in the body navigates… · p. 6

The labels are causal, the deficit localises, and the loop is required

  • The mushroom-body-lesioned fly reached none of three sources in 20 runs, while pinning the MBON valence scalar restored approach in lesioned and rewired flies.Pinning at 0.7 rescued the lesioned fly as fully as pinning at 1.44; a random lesion of the same size left approach mostly intact.Source: The labels are causal, the deficit… · pp. 6–7
  • 0.07 sources per replay run (2.0%) were reached when motor commands were replayed without sensory feedback, versus 3.29 (94.0%) in the closed loop.Rotation alone did not account for the difference: rotated closed-loop runs reached 3.29 sources against 3.36 unrotated.Source: The labels are causal, the deficit… · p. 7
  • After paired reinforcement, none of 20 flies reached the odour source, but reinforcement without odour also produced avoidance in the as-read network.In the matched variant, pairing produced avoidance while reinforcement alone left all 20 flies reaching the source; the as-read Kenyon cells fired even without odour input.Source: Dopamine-gated depression in the loop… · p. 8

Discussion

With fixed parameters and a fixed interface, the untrained connectome navigates to most odour sources, while rewiring or removing sensory feedback sharply reduces navigation. The model also links pathway failures to missing biological features and shows that odour-dependent Kenyon-cell activity is required for associative reversal.

Discussion

  • 30 of 32 odour sources are reached by the intact network, versus 5.8 after degree-and-sign-preserving rewiring and 2% with replayed motor commands.The comparison supports navigation from measured wiring and sensory feedback, while the fixed parameters and interface distinguish this from a trained controller.Source: Discussion · p. 9
  • The model’s pathway tests suggest targets are driven when a pathway supplies about 2% of input, while failures can reflect missing gap junctions or absent ongoing activity.The 2% boundary is empirical for this model’s synaptic weight and threshold, not a literature value.Source: Discussion · p. 9
  • The dynamical map is sparser than the anatomical map: 139 region pairs respond, compared with 358 connected pairs; mushroom-body removal eliminates approach while sparing walking and wall avoidance.A random lesion of the same size leaves approach mostly intact (2.55 of 3) at a cost to walking.Source: Discussion · pp. 9–10
  • The embodied conclusions are bounded by hand-written interface rules: the turn reads MBON valence, while receptor drive does not encode odour concentration and the body model sets gait.Shuffling the connectivity supplying valence reduces it from 1.41 to 0.11; the rescue localises the deficit to mushroom-body output but does not show that the connectome computes the turn.Source: Discussion · p. 10
  • Replayed commands reach 2% of sources when sensory feedback is removed, although the same commands reach 94% in closed loop.A 30° start rotation alone does not lower the closed-loop score; the replay result is independent of the odour rule.Source: Discussion · p. 10
  • In the antennal lobe, excitatory chemical coupling sustains a projection-neuron/local-neuron loop, making two thirds of Kenyon cells respond to every odour.Scaling that synapse class to 0.15 brings the code into the fly’s range; correcting the known-wrong APL signs lowers active Kenyon cells from 68% to 42%, still short of that range.Source: Discussion · p. 10
  • Depressing Kenyon-cell output synapses shifts MBON balance toward avoidance only when Kenyon-cell activity depends on the odour.Where codes depend on odour, generalisation follows code overlap; summing every MBON instead leaves behaviour unchanged because undriven MBONs set a floor.Source: Discussion · pp. 10–11
  • The model omits electrical synapses, neuromodulation and spontaneous activity, and represents one animal; these omissions offer testable additions against the same pathway measurements.The paper also predicts that steering depends on ongoing leg-network activity and that reinforcement without odour pairing devalues odour-independent Kenyon-cell activity.Source: Discussion · p. 11

Methods

The model simulates the neurons and synapses in the adult female fly’s brain and ventral nerve cord using the connectome and previously published neuron parameters. Input, transmitter signs, and experimental manipulations are defined from the connectome’s labels and predictions.

The connectome and the network model

  • The BANC connectome supplies 172,433 neurons, 11,752,828 connections and 35,733,096 synapses, with no wiring changes except specified manipulations.The connection list includes neurons with non-neuronal labels and self-connections; these are retained in the model.Source: The connectome and the network model · p. 11
  • Each neuron uses the leaky integrate-and-fire equations and parameters of Shiu et al., with no parameter retuning.The model uses the published synaptic weight as well as the other published parameters.Source: The connectome and the network model · p. 11
  • Connection strength scales with synapse count, while its sign follows the presynaptic neuron’s predicted transmitter; neurons without a prediction are assigned +1.The model signs acetylcholine, dopamine, serotonin, octopamine and tyramine as +1, and GABA, glutamate and histamine as −1.Source: The connectome and the network model · p. 12
  • Sensory inputs use independent Poisson spike trains, and cell-type labels define sensory, mushroom-body, descending and leg-motor populations.The rewired null randomizes postsynaptic partners while preserving each neuron’s in-degree, out-degree, connection signs and synapse counts.Source: The connectome and the network model · p. 12
  • The 24 regions are grouped from connectome labels, with central-complex, sensory and leg-motor groups further split by neuropil, cell class or segment.The antennal-lobe projection and local neurons, along with neurons matching no grouping rule, fall in the other central-brain group.Source: The connectome and the network model · p. 12

Stimulation outside the body

Outside the body, the authors measured regional responses by driving one region at a time and recording spikes across the network. Separate pulse experiments tracked odour and looming propagation and estimated traversal through the wiring.

Stimulation outside the body

  • Each of the 24 regions received 40 Hz input for 3 s in turn, and the mean response per neuron was measured in every region across three seeds.The displayed influence map transforms each row as log(1 + rate) and divides it by that row’s maximum.Source: Stimulation outside the body · p. 12
  • Odour and looming pulses were delivered from 18 to 78 ms after a 60 ms no-input period, with spikes counted in 6 ms bins.Odour drove receptor neurons at 400 Hz; looming drove LC and LPLC neurons at 300 Hz.Source: Stimulation outside the body · p. 13
  • A neuron counted as reached in the propagation snapshots when its rate in a bin was at least one tenth of its mean over the pulse.A separate information-flow analysis traversed neurons probabilistically according to the share of their input synapses from already traversed neurons, averaging distance over ten seeded traversals.Source: Stimulation outside the body · p. 13

Validation against pathways of known behavioural role

Pathway tests drove populations with known behavioural roles and read out identified target neurons or leg-motor pools. The protocol compared each condition with an undriven control and a rewired twin, then applied prespecified criteria for conduction and wiring specificity.

Validation against pathways of known behavioural role

  • Each test drove a population with a known behavioural role and read an identified target or motor pool, using nine seeds and matched undriven and rewired controls.Runs lasted 3 s, with spikes counted in 20 ms steps and target mean rate measured over the run.Source: Validation against pathways of known… · p. 13
  • A pathway conducted when its target exceeded control by at least 0.10 Hz and the row-specific guard; it was wiring-specific when the rewired twin’s effect was under half the intact effect.The guard was half the control rate, or three control standard deviations for sided rows.Source: Validation against pathways of known… · p. 13
  • Sided rows also had to match the sign measured in the animal, and sides with less than 0.10 Hz total leg-motor drive were not read.The gap-junction test hand-added a giant-fibre connection to the tergotrochanteral motor neuron at five synapse-equivalent strengths.Source: Validation against pathways of known… · p. 13

The odour probe

The odour probe measured neural activity after presenting paired combinations of five olfactory glomeruli to the untrained network. It also specified how local-neuron output scaling, odour training and follow-up tests were performed.

The odour probe

  • Odour presentations drove the receptor neurons of five glomeruli at 150 Hz for 1 s after a 1 s silent rest, following a 0.5 s warm-up.The main comparison paired A = D, DA1, DA2, DA3 and DA4l with B = DA4m, DC1, DC2, DC3 and DC4; three further pairs followed the alphabetical glomerulus list.Source: The odour probe · p. 13
  • The readouts were layer active fraction, overlap among the 5% most active Kenyon cells, and total MBON spikes during presentation.MBON output below 5 spikes before training was treated as unmeasurable.Source: The odour probe · p. 14
  • The matched variant scaled output synapses of 105 predicted-cholinergic antennal-lobe local neurons, using the largest gain meeting the prespecified Kenyon-cell activity and overlap criteria.The selected gain was g = 0.15, with 3 to 15% of Kenyon cells active and overlap at most 0.5.Source: The odour probe · p. 14
  • Training depressed Kenyon-cell-to-MBON synapses according to each cell’s normalized response to odour A, after which both odours were presented again.Additional tests varied receptor drive, odour breadth and background contingency; an APL-sign test was explicitly not pre-registered and marked exploratory.Source: The odour probe · p. 14

The body and its coupling to the network

The simulated body receives sensory input through fixed odour and looming rules, while network activity controls leg movement and steering. The loop runs in 50 ms steps across arenas with walls and odour sources.

The body and its coupling to the network

  • The body is a MuJoCo NeuroMechFly model stepped at 0.1 ms, using leg adhesion, contact sensors, and a hybrid gait controller driven by six leg amplitudes.The controller retains its central-pattern-generator network and retraction and stumbling reflexes.Source: The body and its coupling to the network · p. 14
  • Every 50 ms, odour asymmetry drives antenna receptor rates and upper-eye brightness drives LC and LPLC visual neurons before the network runs for 50 ms.The odour sensor values are combined with weights 9 and 1; visual drive is based on the inverse of mean brightness in the upper 45% of eye rows.Source: The body and its coupling to the network · p. 14
  • Smoothed leg-motor-pool counts set each leg’s amplitude, while MBON valence and odour asymmetry scale the legs to steer toward or away from odour.Positive valence turns toward the stronger odour side; negative valence turns away.Source: The body and its coupling to the network · pp. 14–15
  • When looming exceeds 0.05, a ramp suppresses odour steering and scales the legs opposite the stronger looming signal or more active DNa02.The ramp reaches 1 at 0.28, and side scalings are clipped between 0.3 and 1.Source: The body and its coupling to the network · p. 15
  • Nine arenas provide 32 odour sources across obstacle layouts, with a 20 s cap; the forage arena has three sources behind a wall and a 16 s cap.The learning arena has one source and 6 s episodes.Source: The body and its coupling to the network · p. 15

The nine-arena battery

The arena battery tests navigation across nine layouts, comparing intact, rewired, and mushroom-body-lesioned wirings against chance and source-informed reference controllers. Runs track source collection and movement measures.

The nine-arena battery

  • Each wiring condition runs in every arena for 20 seeds, yielding 180 runs per condition, with sources reached, straightness, median speed, and active-step fraction measured.An active step has speed above 1 mm/s.Source: The nine-arena battery · p. 15
  • Chance level uses a correlated random walk through the same geometries, while two reference controllers steer toward the nearest remaining source using source positions.When a wall blocks a step, one reference controller stays in place and the other moves along the wall.Source: The nine-arena battery · p. 15
  • The reference and chance comparisons match the intact fly’s speed, capture radius, time cap, and seed count; chance also reflects at walls.The reference controllers are given the position of every source.Source: The nine-arena battery · p. 15

Perturbations in the body

Body experiments perturb wiring, sensory input, motor commands, and descending-neuron drive to examine navigation and movement. Measures include source approach, wall proximity, path length, speed, and turning.

Perturbations in the body

  • Forage-arena tests compare intact, lesioned, rewired, and valence-pinned flies, plus random lesions and zeroed looming input, across six behavioural measures.Measures include sources reached, closest approach, DNa02 counts, wall proximity, path length, and leg-motor-pool rate; each condition has 20 completed runs.Source: Perturbations in the body · p. 15
  • Motor replay compares closed-loop runs with an open-loop run that feeds recorded leg amplitudes to a fly starting at a 30° rotated heading.The open-loop run ignores network output while replaying the recorded amplitudes step by step.Source: Perturbations in the body · p. 16
  • Descending-neuron tests deliver a 600 Hz, 2 s target drive after 1 s of baseline, then measure speed and net turn against a settling window.Targets include DNa01, DNa02, DNp09, a threat-response cluster, and non-locomotor controls; receptor drive is 120 Hz, with 40 and 0 Hz also used in one condition.Source: Perturbations in the body · p. 16
  • Graded removal zeroes synapses for random neuron sets or the most-connected candidates, while sparing the sensory, motor, MBON, and DNa02 sets used by the loop.Random removal spans 5–90%; targeted removal spans 1–20% and can be restricted to central-brain intrinsic neurons.Source: Perturbations in the body · p. 16

Learning in the loop

The learning setup weakens selected Kenyon-cell-to-MBON synapses using dopamine-linked eligibility traces, then tests pairing, odour exposure, and learning-rate conditions. The supplied passages also report a limitation of making all connections plastic under the summed readout.

Learning in the loop

  • Of 15,523 Kenyon-cell-to-MBON connections, 8,377 onto MBONs in dopamine-innervated aversive-learning compartments are plastic; the other 7,146 retain their weights.The plastic connections have a mean initial weight of 0.513 mV.Source: Learning in the loop · p. 16
  • Each plastic connection accumulates a decaying eligibility trace from Kenyon-cell spikes and, during reinforcement, depression proportional to that trace and its MBON’s spikes.The trace time constant is 200 ms; accumulated depression is subtracted at episode end and clipped between zero and the initial weight.Source: Learning in the loop · p. 16
  • With all connections plastic under the summed readout, learning lowers valence in both conditions but leaves 19 of 20 trained flies reaching the source.The passage attributes the remaining valence floor to MBONs not driven by the Kenyon cells; the matched-variant test specifies criteria for associative and expressed depression.Source: Learning in the loop · pp. 16–17
  • Training uses six episodes between naive and post-tests, with paired odour and reinforcement, unpaired reinforcement without odour, or no plasticity; the main learning rate is η = 0.02.The learning-rate ladder tests eleven η values from 10^-5 to 0.08, and outcomes include weight, valence, closest approach, and devaluation.Source: Learning in the loop · p. 16

Statistics and reproducibility

The paper defines the seed as the replication unit, treats arenas as repeated measures, and reports pre-specified statistical tests and simulation details. It also reports power estimates and sensitivity checks for selected analyses.

Statistics and reproducibility

  • Seeds are the replication unit, arenas are repeated measures, and all tests are two-sided.Replay uses an exact paired Wilcoxon signed-rank test; group comparisons use Mann–Whitney U tests with Cliff’s δ and bootstrap intervals.Source: Statistics and reproducibility · p. 17
  • Power for a 20-versus-20 Mann–Whitney test ranges from 0.19 to 0.97 across normal-shift effect sizes δ = 0.2 to 0.65.The reported powers for δ = 0.2, 0.35, 0.5 and 0.65 are 0.19, 0.49, 0.83 and 0.97, respectively.Source: Statistics and reproducibility · p. 17
  • The Fig. 3i correlation is Spearman’s ρ = 0.78 (P = 0.0045) over 11 types, and remains 0.71–0.88 when any one type is omitted.Without the largest type, ρ = 0.71 (P = 0.022); no multiple-comparison correction was applied, and Bonferroni correction across six learning tests changes no conclusion.Source: Statistics and reproducibility · p. 17
  • Sweeps’ cells, metrics, thresholds and predictions were written down before launch, except the explicitly marked exploratory APL run.Simulations ran in a specified software environment on two-core, 12 GiB pods in the Nautilus cluster; figures and statistics were built locally.Source: Statistics and reproducibility · p. 17

Funding

The authors report support from research funders and fellowships. They state that funders did not participate in the study’s design, analysis, publication decision or writing.

Funding

  • M.T. and D.H. disclose support from Schmidt Futures, the National Human Genome Research Institute and the National Science Foundation.Their disclosures also list additional support from the California Institute for Regenerative Medicine and the University of California Office of the President; D.H. additionally lists NIMH and NINDS support.Source: Funding · p. 22
  • M.A.M.-R. lists support from the National Science Foundation, NIMH, NINDS, the California Institute for Regenerative Medicine, the University of California Office of the President and the Brain and Behavior Research Foundation.J.G.-F. reports a QB3 Graduate Innovators Fellowship; J.R.M., J.G. and J.K.E. declare no relevant funding.Source: Funding · p. 22
  • The funders took no part in study design, analysis, the decision to publish or manuscript writing.The authors state that the content is their responsibility and may not represent the official views of the named agencies.Source: Funding · p. 22

Figure legends

The figure legends map the connectome and its influence, assay pathways and odour coding, and show embodied navigation, loop perturbations and learning protocols. They specify the visual encodings and conditions for each figure.

Figure legends

  • Figure 1 maps the connectome’s 24 regions and synapse counts, then shows row-normalized responses when each region is stimulated alone.The functional-influence matrix has stimulated regions in rows and responding regions in columns; each row is normalized to its maximum.Source: Figure 1 · p. 26
  • Figure 2 lays out identified driven pathways and readouts, including pathway input shares, target firing, rewired controls and registered behavioural outcomes.Panels include sugar responses with and without additional tastants and descending drives to leg motor neurons.Source: Figure legends · p. 24, Figure 2 · p. 27
  • Figure 3 compares the as-read antennal-lobe circuit with a variant scaling cholinergic local-neuron output, and tracks activity and odour-code overlap across layers.The legend also describes Kenyon-cell activity across odours, silencing local-neuron types and odour-drive conditions.Source: Figure legends · p. 24, Figure 3 · p. 28
  • Figure 4 contrasts intact, rewired and mushroom-body-lesioned flies across nine arenas, showing paths, occupancy, sources reached and walking measures.It includes 20 seeds per condition in each arena and 32 odour sources overall.Source: Figure 4 · p. 29
  • Figure 5 compares sensorimotor-loop perturbations, including closed-loop navigation against replayed motor commands with sensory feedback cut under a 30° yaw perturbation.The legend also describes paths, sources reached, measures relative to intact, and net turn and speed change.Source: Figure legends · p. 25, Figure 5 · p. 30
  • Figure 6 diagrams dopamine-gated plasticity at Kenyon-cell-to-MBON synapses and contrasts paired odour-plus-dopamine training, unpaired dopamine and no plasticity.The protocol places six training episodes between naive and post-training tests.Source: Figure legends · p. 25, Figure 6 · p. 31

Supplementary Information

Supplementary tests show that reducing summed MBON output does not reverse odour approach; learned avoidance requires a signed balance readout and reinforcement confined to aversive compartments.

Supplementary Discussion. The learning experiment with the valence read as the summed MBON count

  • 19 of 20 trained flies still reached the source under summed valence, despite valence falling from 1.44 to 0.24.The median closest approach was 0.23 mm after training versus 0.20 mm before training; 20 of 20 flies reached the source without plasticity.Source: Supplementary Discussion. The learning… · p. 32
  • Training reduced MBON spikes per step from 155 to 26, but ten MBONs carried 87% of the residual output.Six of those ten took at most a quarter of their input from Kenyon cells, which limits what this plasticity can lower.Source: Supplementary Discussion. The learning… · p. 32
  • In the matched variant, pairing lowered synaptic weight to 0.182 mV and valence to 0.25, while reinforcement alone left them at 0.495 mV and 1.42.The paired fly still reached the source, consistent with the summed-readout floor.Source: Supplementary Discussion. The learning… · p. 32
  • Odour and looming pulses reached leg motor pools within 24 ms in the network alone.A separate wiring traversal tracked the order of pulse arrival, although only 4% of odour-reached and 2% of looming-reached neurons spiked in the silent network.Source: Figure 1 · p. 34, Figure 2 · p. 35
  • 42.5% of synapses carried inhibitory signs, and influence correlated more strongly with excitatory-signed counts (ρ = 0.65) than with counts regardless of sign (ρ = 0.56).Region pairs with influence but no direct synapse represent relayed routes, including receptor pathways to the mushroom body and lateral horn.Source: Figure 3 · p. 36
  • Only the signed balance readout with reinforcement confined to aversive compartments produced learned turning away.With summed readout and restricted reinforcement, flies still reached the source; with balance readout and reinforcement everywhere, they neither approached nor avoided.Source: Supplementary Discussion. The learning… · p. 33, Figure 4 · p. 37
Loading 2610.33881v1…