PaperlayerView exact revision on arXiv

Source-linked AI summary

Structure alone supports efficient visual computation in the Drosophila visual system

Eudald Correig-Fraga, Roger Guimerà, Marta Sales-Pardo

arXiv:2610.10023v1q-bio.NCcond-mat.dis-nncs.LG

TL;DR

Whether measured synaptic wiring determines neural computation remains a central question. The paper couples the adult Drosophila connectome to a faithful eye model and compares visual-task performance across randomized wiring variants. The biological network performs best among biologically constrained alternatives, while less constrained rewiring achieves higher accuracy at greater wiring cost.

  • Problem

    A central connectomics question is how much neural computation is determined by measured synaptic wiring.

    Source: §I · p. 1

  • Method

    The authors compare visual-task performance in a connectome-based model and randomized variants with differing wiring constraints.

    Source: §I · p. 2

  • Results

    Biologically plausible randomizations preserving wiring budget and global length distribution perform close to, but consistently below, the biological connectome, while less constrained rewiring achieves higher accuracy with greater wiring cost.

    Source: §III · p. 8

  • Takeaways & Limitations

    Connectome-only models provide a framework for linking neural structure, computation, and biological constraint.

    Source: §III · p. 8

Abstract

from arXiv · show

Understanding the extent to which measured synaptic wiring determines computation remains a central challenge. Here, we couple the proofread adult Drosophila melanogaster connectome to an anatomically faithful model of its eye. Visual information is inputted in the eye model, then passed to the connectome, and finally read from a Kenyon-cell-centered linear decoder. This creates a connectome-only model in which the anatomical graph and eye geometry are fixed and only scalar synaptic gains and neuronal thresholds may be learned. The model supports multitask vision, including color discrimination, shape classification, and numerical discrimination that follows a ratio-dependent scaling characteristic of approximate number perception. To test whether precise connectivity is consequential under wiring economy, we compare the biological graph to randomized ensembles that increasingly preserve biological synaptic constraints. At matched wiring cost, the biological network consistently yields higher accuracy, whereas less constrained rewiring surpasses it at the cost of inflated wiring. These findings indicate that the measured connectivity and eye geometry jointly set efficient operating points for visual computation.

I. INTRODUCTION

The paper tests whether measured fly-brain wiring and eye geometry support visual computation under wiring economy. It couples a connectome-only model to a realistic eye model and compares biological wiring with randomized networks under varying anatomical constraints.

I. INTRODUCTION

  • The study asks whether measured synaptic wiring supports non-trivial visual computation under biophysical constraints, including energetic costs.The authors frame brain architecture as supporting multiple complex tasks with a limited set of neurons.Source: §I · p. 1

A. A connectome-only model to assess visual task performance

  • The model couples a connectome message-passing network to a compound-eye model and reads activity from Kenyon cells.The eye model maps visual inputs to retinal activations; the connectome graph and eye geometry are fixed, with scalar gains on existing synapses as the trainable connectome parameters.Source: A. A connectome-only model to assess… · p. 2

B. Randomized connectome ensembles to explore the relationship between wiring and task performance

  • Randomized networks are compared across wiring constraints, using summed soma-to-soma distance weighted by synapse count as a proxy for wiring cost.The ensembles range from unconstrained rewiring and cost-matched pruning to distance-binned and neuron-specific binning schemes.Source: B. Randomized connectome ensembles to… · p. 4
  • In the biological and more biologically plausible networks, activity spreads more gradually than in the least constrained ensembles, which activate nearly all neurons, including Kenyon cells, by the second step.The authors suggest that fewer long-range synapses limit spread and may contribute to task-performance differences.Source: B. Randomized connectome ensembles to… · p. 4

C. Connectome models can perform different visual tasks

The models are trained with fixed anatomical wiring and evaluated on color, shape, and numerosity discrimination. They perform nearly perfectly on color, less strongly on shape, and show ratio-dependent numerosity performance in area-controlled stimuli.

C. Connectome models can perform different visual tasks

  • Models use three message-passing steps, then average Kenyon-cell activity and pass it to a single linear decision unit; all variants share the training protocol.The task set includes color discrimination, shape discrimination, and numerosity tasks.Source: C. Connectome models can perform… · pp. 4, 6
  • The figure reports activity spreading more gradually in biological and biologically constrained networks than in unconstrained networks, where the whole brain becomes active almost immediately.It tracks active-neuron fractions, mean distance of active neurons, and propagation patterns.Source: Figure 2 · p. 5
  • All models reach nearly 100% accuracy on color discrimination, and the biological connectome reaches 92% even without training.The task classifies images by circle color despite variation in circle size and position.Source: C. Connectome models can perform… · p. 6
  • Shape accuracy is 64% for the biological model and 70% for models in the unconstrained ensembles.The passage characterizes shape recognition as substantially more difficult than color discrimination.Source: C. Connectome models can perform… · p. 6
  • Numerosity discrimination asks which color has more dots; its area-controlled condition equalizes the total surface area of the two colors.This condition prevents total surface area from serving as a cue.Source: C. Connectome models can perform… · p. 6
  • Across models, numerosity accuracy is above chance and increases with the ratio of the more to less numerous dot sets.For the biological model, accuracy rises from 64% at r = 1.5 to 85% at r = 5.0.Source: C. Connectome models can perform… · p. 6

D. Wiring constraints limit model accuracy and the biological connectome performs best among biologically-plausible models

Accuracy varies with the wiring constraints imposed on randomized connectomes. Less constrained networks perform best, but the biological connectome performs best among the more biologically plausible alternatives at comparable wiring economy.

D. Wiring constraints limit model accuracy and the biological connectome performs best among biologically-plausible models

  • Unconstrained and connection-pruned networks achieve the highest accuracies across tasks, with more long-range synapses that allow activity to spread farther and reach more neurons in fewer steps.The authors suggest that access to more neurons for computation helps explain their higher accuracy.Source: D. Wiring constraints limit model… · p. 6
  • The biological connectome performs close to but consistently above the synapse-bin and neuron-bin ensembles, while unconstrained and connection-pruned networks achieve higher accuracy.The latter networks contain more long-range synapses; the connection-pruned ensemble matches biological total wiring while changing the synaptic-length distribution.Source: D. Wiring constraints limit model… · p. 6, Figure 3 · p. 7
  • The authors interpret the biological connectome as an efficient evolutionary solution that maximizes accuracy under biological constraints and performs systematically better than nearby random alternatives.Source: D. Wiring constraints limit model… · p. 7

III. DISCUSSION

The connectome-only model supports non-trivial visual computations, including number comparisons, and the biological connectome performs near the accuracy optimum under biologically realistic wiring constraints. The discussion presents connectome-only models as tools for linking neural structure, computation, and biological constraints.

III. DISCUSSION

  • The model identifies differences in the numbers of differently colored points even though it is not trained with point number as an explicit target.The authors present this as evidence that measured wiring and eye geometry support visual computation requiring abstraction.Source: §III · pp. 7–8
  • Under biologically realistic wiring constraints, the biological connectome consistently outperforms randomized variants and operates near the accuracy optimum at its native wiring cost.Less constrained rewiring can achieve higher accuracy, but increases average connection length and total wiring cost.Source: §III · p. 8
  • Unconstrained long-range wiring raises energetic and developmental costs, making the biological connectome a parsimonious performance-cost trade-off.The discussion links these costs to axonal length, membrane area, signal propagation, circuit-generation rules, and axon guidance.Source: §III · p. 8
  • Randomized connectomes provide a way to test which wiring features are functionally consequential by comparing architectures that preserve or violate anatomical constraints.The authors position connectome-only models as benchmarks and explanatory tools for linking neural structure, computation, and biological constraint.Source: §III · p. 8
  • The results delineate a regime in which measured wiring and eye geometry support non-trivial visual computation, while offering a starting point for progressively richer models.The proposed extensions could add explicit neuronal or synaptic dynamics, such as graded responses or firing-rate dynamics.Source: §III · p. 8

IV. METHODS

The model uses the proofread adult fly connectome as its anatomical graph and a spatially and spectrally organized eye model to supply retinal activity. Neural activity propagates through the graph in discrete steps using synaptic inputs and neuronal activation rules.

A. Connectome Data

  • The anatomical graph uses the version 783 proofread adult Drosophila whole-brain connectome, with 139,255 neurons and 54.5 million synaptic connections.Neurons are nodes, directed edges represent connected pairs and carry synapse multiplicity; analyses used the right-eye projection and downstream circuitry unless specified otherwise.Source: A. Connectome Data · p. 9

B. Biologically faithful eye front end

  • The eye model maps image input to individual photoreceptor activations while preserving ommatidial spatial organization and receptor spectral sensitivities.The eye model’s output is a vector of retinal-neuron activations that drives the neural model.Source: B. Biologically faithful eye front end · p. 9

C. Neural network architecture and propagation

  • Neural activity propagates through a message-passing network on the connectome, with each neuron a node and each synaptic connection a directed edge.In the implementation, neurons reset between steps and compute new activations from incoming inputs; the message contribution is linear in presynaptic activity and anatomical synapse count, with learnable gains and neuronal thresholds.Source: C. Neural network architecture and… · pp. 9–10

D. Synaptic and neuronal plasticity

The model can vary synaptic gains and neuronal activation thresholds while keeping the connectome architecture fixed. The main-text experiments use the edges-only configuration.

D. Synaptic and neuronal plasticity

  • Learning can adapt synaptic gains, neuronal activation thresholds, or both while keeping the underlying connectome architecture fixed.Gains modulate anatomical connection strength, while thresholds control individual neurons’ excitability.Source: D. Synaptic and neuronal plasticity · p. 11
  • The four configurations train the classifier alone, edges alone, thresholds alone, or both edges and thresholds alongside the classifier.Edges-only learns one gain per nonzero directed connection; thresholds-only learns one threshold per retained neuron.Source: D. Synaptic and neuronal plasticity · p. 11
  • Main-text experiments use edges-only, optimizing one gain per nonzero directed connection while topology and neuronal properties remain fixed.Threshold-learning and transmitter-specific sign variants are described in the Supplemental Material and are not used in the principal biological-versus-randomized connectome comparisons.Source: D. Synaptic and neuronal plasticity · p. 11

E. Readout locus

The model reads out Kenyon-cell population activity as its stimulus representation, then uses a linear classifier for behavioral predictions.

E. Readout locus

  • After N propagation steps, the model uses mushroom-body Kenyon-cell population activity as its final stimulus representation.The paper identifies Kenyon cells as widely considered principal decision-making neurons in the Drosophila brain.Source: E. Readout locus · p. 11
  • A simple linear classifier converts the Kenyon-cell activity into a behavioral prediction.The stated choice keeps the model as biologically constrained as possible.Source: E. Readout locus · p. 11
  • An auxiliary class-unit variant for multi-class tasks is described in the Supplemental Material but is not used in the principal comparisons.The variant introduces NKenyon × C auxiliary class units.Source: E. Readout locus · p. 11

F. Randomised-connectome ensembles

The study compares the biological connectome with randomized networks that preserve different aspects of its wiring. The ensembles range from unconstrained rewiring to variants that preserve connection-length structure.

F. Randomised-connectome ensembles

  • The four randomized configurations preserve the neuron set and use the same stimuli and train/test splits as the biological connectome.One graph instance was used per randomization strategy.Source: F. Randomised-connectome ensembles · p. 12
  • Unconstrained rewiring preserves each neuron’s out-degree but ignores spatial constraints, producing networks with greater total wiring length and mean synapse length than biological values.Targets are reassigned randomly, and synapse counts are redistributed across the new targets.Source: F. Randomised-connectome ensembles · p. 12
  • Connection-pruned rewiring removes costly connections from an unconstrained network until total wiring length matches the biological value, while mean connection length remains higher.Connections contributing most to wiring cost are preferentially removed.Source: F. Randomised-connectome ensembles · p. 12
  • Synapse-bin rewiring preserves the global connection-length distribution and total wiring cost while randomizing which neuron pairs are connected.Neuron pairs are grouped into distance bins, and synapses are shuffled only within each bin.Source: F. Randomised-connectome ensembles · p. 12
  • Neuron-bin rewiring preserves each neuron’s outgoing connection-length distribution, in/out-degree, and neighbors while altering synapse counts between already connected pairs.It is the most constrained randomized null model described, short of the observed biological connectome.Source: F. Randomised-connectome ensembles · p. 12

G. Wiring-cost metrics

The section quantifies wiring cost using neuron locations and synapse counts, with total wiring length and synapse-weighted mean connection length capturing complementary aspects of network structure.

G. Wiring-cost metrics

  • Total wiring length quantifies physical network cost using neuronal soma distances and synapse counts.The distance is Euclidean, and the synapse count is the number of connections between neurons.Source: G. Wiring-cost metrics · pp. 12–13
  • Synapse-weighted mean connection length measures the average spatial extent of synaptic connections.Source: G. Wiring-cost metrics · p. 13
  • Synapse-length histograms are annotated with both total wiring length and mean connection length.Source: G. Wiring-cost metrics · p. 13

H. Computational tasks and stimuli

The study tests color, shape, and numerical discrimination with synthetic images sampled according to the eye’s retinotopy, using controls to limit reliance on simple visual cues. Models share the same data splits and training protocol, with regularization used to mitigate overfitting.

H. Computational tasks and stimuli

  • Synthetic 512 × 512 images were rendered according to eye retinotopy, with 5,000 training, validation, and test images per class and controls against trivial cues.Source: H. Computational tasks and stimuli · p. 13
  • Color discrimination classifies yellow versus blue stimuli matched for area and center position, with an additional test across visual-field sectors.Source: H. Computational tasks and stimuli · p. 13
  • Shape recognition classifies circles versus stars, testing generalization from left-field training stimuli to right-field test stimuli.This retinotopic-sector protocol tests whether representations are position-invariant.Source: H. Computational tasks and stimuli · p. 13
  • Numerical discrimination uses more-versus-less judgments across eight Weber ratios, with area-controlled results reported as the direct measure of discrimination.The stimuli also prevent reliance on total area and trivial dot-size correlations.Source: H. Computational tasks and stimuli · p. 13

I. Training protocol and regularisation

  • Models used cross-entropy loss and AdamW for up to 100 epochs, with early stopping monitored on training measures and regularization including neuron, Kenyon-cell, and selected cell-type dropout.Biological and randomized ensembles used the same data splits.Source: §I · p. 14

J. Propagation and spatial-reach metrics

The section defines activity-spread and spatial-reach measures for message passing, and documents the software, data, and code resources used for reproducibility. The authors report no conflict of interest.

J. Propagation and spatial-reach metrics

  • Propagation metrics include the fraction of active neurons at each step and the analogous fraction restricted to Kenyon cells.A neuron is counted as active when its updated activation exceeds a small tolerance.Source: J. Propagation and spatial-reach metrics · p. 14
  • Mean distance from input measures the synapse-weighted mean Euclidean distance between active neurons and retinal input neurons at each step.Metrics are computed under the default training configuration and averaged across connectome ensembles.Source: J. Propagation and spatial-reach metrics · p. 14

K. Software and reproducibility

COMPETING INTERESTS

DATA AVAILABILITY

  • Processed biological and randomized graphs, retinal-neuron coordinate mapping, prediction files, and numerical-discrimination tables are available on Zenodo under CC BY 4.0.Source: DATA AVAILABILITY · p. 15
Loading 2610.10023v1…