Connectome-Constrained Spiking Neural Networks: Olfactory Classification Study

Moberg et al., 2026

Abstract

Artificial neural networks achieve remarkable performance through learned weight distributions over largely arbitrary topologies, yet biological neural circuits evolved under strict metabolic and behavioral constraints over millions of years. Whether this evolved wiring encodes computational priors that confer advantages in learning efficiency, generalization, or representational sparsity over randomly initialized architectures remains an open question. We present a controlled comparison between a Spiking Neural Network whose recurrent synaptic topology is constrained to the Drosophila melanogaster olfactory circuit as mapped by the FlyWire connectome (Shiu et al., 2024) and three baselines: a degree-matched randomly-wired SNN, a sparsity-matched MLP, and a fully-connected parameter-matched MLP. All models are trained on the DoOR olfactory receptor response dataset for multi-class odor identity classification using surrogate gradient descent. We hypothesize that biologically-derived synaptic topology encodes an inductive bias toward more efficient or stable olfactory representations than randomly-initialized architectures of equivalent capacity. We evaluate convergence rate, peak classification accuracy, spike sparsity, and training stability across five-fold cross-validation and multiple random seeds. All code, connectome extraction pipelines, and trained weights are released to support reproducibility and extension to other FlyWire circuit subgraphs.

EXPERIMENT DATA REPORT

Connectome-Constrained Spiking Neural Networks: Olfactory Classification Study

Results summary prepared for co-author review — March 2026

What this experiment asked

Does using the actual wiring diagram of a fruit fly's nose as the architecture of an AI network help it learn better than a randomly-wired network of the same size?

How it was tested

Four AI models were trained to identify odors from receptor response data. All models had roughly the same number of trainable parameters (~1.25M). The only difference between the two key models was whether their internal connections matched real fly brain anatomy or were randomly shuffled.

Dataset

DoOR — a database of olfactory receptor responses compiled from multiple published studies. 24 receptor measurements per odor, multi-class classification task.

Protocol

5-fold cross-validation × 3 random seeds = 15 independent runs per model, 60 total training runs. Results reported as mean ± standard deviation.

Plain Language Summary

Primary finding (null)

The biologically-wired model (ConnectomeSNN) performed no better than the randomly-wired model (ShuffledSNN) in classification accuracy. The difference (45.5% vs 45.4%) is smaller than the measurement error and is not statistically meaningful.

Secondary observation

The biologically-wired model showed higher average training duration (27 epochs vs 24.7) but also higher variability. When compared run-by-run, it trained longer in only 7 of 15 matched runs — not a consistent effect (p = 0.696).

✓

Robust finding

Both spiking neural network models naturally settled into ~80% spike sparsity — meaning about 80% of neurons were silent at any given moment. This mirrors biological olfactory circuits and emerged without being explicitly programmed.

Architecture comparison

Standard feedforward networks (MLPs) outperformed both spiking models by ~3–4 percentage points and converged faster. This is expected — MLPs are optimized for this type of static classification task.

The Four Models

ConnectomeSNN

Spiking network whose internal connections are fixed to match real fly brain anatomy (FlyWire connectome). The primary model — what the hypothesis is about.

SNN

ShuffledSNN

Identical spiking network but connections are randomly shuffled. Same number of connections per neuron, different wiring. The topology control.

SNN

SparseMLP

Standard neural network with the same density of connections as the connectome models, but no spiking dynamics. Isolates the effect of sparsity from topology.

DenseMLP

Fully connected standard neural network. The conventional deep learning baseline. Matched in total parameter count to the other models.

Summary Results

All values: mean ± standard deviation across 15 runs. Test accuracy is proportion correct (0–1 scale, so 0.455 = 45.5% correct). Higher is better for accuracy; lower is better for stopped epoch (faster convergence).

Model	Type	Test Accuracy	Stopped Epoch	Spike Sparsity	Parameters
ConnectomeSNN Biological topology	SNN	45.5% ± 1.2%	27.0 ± 8.8	80.2% ± 0.6%	1,235,892
ShuffledSNN Random topology	SNN	45.4% ± 1.3%	24.7 ± 7.1	80.3% ± 0.7%	1,235,892
SparseMLP Sparse standard network	MLP	48.9% ± 1.1%	19.3 ± 5.2	—	1,235,892
DenseMLP Fully connected baseline	MLP	49.2% ± 0.9%	18.7 ± 4.8	—	1,235,180

ConnectomeSNN

Biological topology • SNN

Test Accuracy 45.5% ± 1.2%

Stopped Epoch 27.0 ± 8.8

Spike Sparsity 80.2% ± 0.6%

Parameters 1,235,892

ShuffledSNN

Random topology • SNN

Test Accuracy 45.4% ± 1.3%

Stopped Epoch 24.7 ± 7.1

Spike Sparsity 80.3% ± 0.7%

Parameters 1,235,892

SparseMLP

Sparse standard network • MLP

Test Accuracy 48.9% ± 1.1%

Stopped Epoch 19.3 ± 5.2

Spike Sparsity —

Parameters 1,235,892

DenseMLP

Fully connected baseline • MLP

Test Accuracy 49.2% ± 0.9%

Stopped Epoch 18.7 ± 4.8

Spike Sparsity —

Parameters 1,235,180

Run-by-Run Comparison: ConnectomeSNN vs ShuffledSNN

Each row is one matched training run (same data fold, same random seed). This table is the basis for the sign test reported in the paper.

Sign test result: ConnectomeSNN stopped later (trained longer) in 7 of 15 matched runs. Under the null hypothesis of no effect, we'd expect ~7–8 by chance. One-sided sign test p = 0.696 — not significant. The biological topology does not systematically produce longer training.

Run	Fold	Seed	Connectome Epochs	Shuffled Epochs	Connectome Accuracy	Shuffled Accuracy	Conn Trained Longer?
1	0	7	34	26	46.7%	46.6%	Yes
2	0	1010	34	37	47.9%	48.1%	No
3	0	2013	23	28	46.3%	47.0%	No
4	1	7	26	18	45.3%	44.5%	Yes
5	1	1010	26	21	44.6%	44.0%	Yes
6	1	2013	18	20	47.1%	47.5%	No
7	2	7	28	20	46.4%	45.5%	Yes
8	2	1010	16	22	45.2%	45.6%	No
9	2	2013	23	23	46.1%	45.6%	Tie
10	3	7	41	30	46.6%	45.8%	Yes
11	3	1010	30	35	44.8%	46.3%	No
12	3	2013	44	24	46.8%	45.8%	Yes
13	4	7	19	24	45.9%	45.8%	No
14	4	1010	21	42	44.8%	45.2%	No
15	4	2013	17	20	43.0%	43.6%	No

Run 1 (Fold 0, Seed 7)

Connectome Epochs 34

Shuffled Epochs 26

Connectome Accuracy 46.7%

Shuffled Accuracy 46.6%

Conn Trained Longer? Yes

Run 2 (Fold 0, Seed 1010)

Connectome Epochs 34

Shuffled Epochs 37

Connectome Accuracy 47.9%

Shuffled Accuracy 48.1%

Conn Trained Longer? No

Run 3 (Fold 0, Seed 2013)

Connectome Epochs 23

Shuffled Epochs 28

Connectome Accuracy 46.3%

Shuffled Accuracy 47.0%

Conn Trained Longer? No

Run 4 (Fold 1, Seed 7)

Connectome Epochs 26

Shuffled Epochs 18

Connectome Accuracy 45.3%

Shuffled Accuracy 44.5%

Conn Trained Longer? Yes

Run 5 (Fold 1, Seed 1010)

Connectome Epochs 26

Shuffled Epochs 21

Connectome Accuracy 44.6%

Shuffled Accuracy 44.0%

Conn Trained Longer? Yes

Run 6 (Fold 1, Seed 2013)

Connectome Epochs 18

Shuffled Epochs 20

Connectome Accuracy 47.1%

Shuffled Accuracy 47.5%

Conn Trained Longer? No

Run 7 (Fold 2, Seed 7)

Connectome Epochs 28

Shuffled Epochs 20

Connectome Accuracy 46.4%

Shuffled Accuracy 45.5%

Conn Trained Longer? Yes

Run 8 (Fold 2, Seed 1010)

Connectome Epochs 16

Shuffled Epochs 22

Connectome Accuracy 45.2%

Shuffled Accuracy 45.6%

Conn Trained Longer? No

Run 9 (Fold 2, Seed 2013)

Connectome Epochs 23

Shuffled Epochs 23

Connectome Accuracy 46.1%

Shuffled Accuracy 45.6%

Conn Trained Longer? Tie

Run 10 (Fold 3, Seed 7)

Connectome Epochs 41

Shuffled Epochs 30

Connectome Accuracy 46.6%

Shuffled Accuracy 45.8%

Conn Trained Longer? Yes

Run 11 (Fold 3, Seed 1010)

Connectome Epochs 30

Shuffled Epochs 35

Connectome Accuracy 44.8%

Shuffled Accuracy 46.3%

Conn Trained Longer? No

Run 12 (Fold 3, Seed 2013)

Connectome Epochs 44

Shuffled Epochs 24

Connectome Accuracy 46.8%

Shuffled Accuracy 45.8%

Conn Trained Longer? Yes

Run 13 (Fold 4, Seed 7)

Connectome Epochs 19

Shuffled Epochs 24

Connectome Accuracy 45.9%

Shuffled Accuracy 45.8%

Conn Trained Longer? No

Run 14 (Fold 4, Seed 1010)

Connectome Epochs 21

Shuffled Epochs 42

Connectome Accuracy 44.8%

Shuffled Accuracy 45.2%

Conn Trained Longer? No

Run 15 (Fold 4, Seed 2013)

Connectome Epochs 17

Shuffled Epochs 20

Connectome Accuracy 43.0%

Shuffled Accuracy 43.6%

Conn Trained Longer? No

Fold = which partition of odor classes was held out. Seed = random initialization number (7, 1010, or 2013). Same fold + seed = identical data split and starting conditions for both models.

Glossary

Connectome

A complete map of every neuron and synapse in a nervous system. The FlyWire connectome maps the fruit fly (Drosophila melanogaster) brain at single-synapse resolution.

Spiking Neural Network (SNN)

An AI model that mimics biological neurons — neurons only "fire" when their activation crosses a threshold, producing sparse, event-driven computation rather than continuous signals.

MLP (Multilayer Perceptron)

A conventional fully-connected neural network. The standard architecture used in most deep learning applications. No spiking dynamics, no biological constraints.

Surrogate gradient

A mathematical trick that allows spiking networks to be trained with backpropagation, which normally requires smooth differentiable functions. Spikes are binary (on/off), which backprop cannot handle directly.

Spike sparsity

The fraction of neurons that are silent (not firing) at any given moment. Higher sparsity = fewer neurons active. Biological Kenyon cells show ~90–95% sparsity; these models showed ~80%.

Early stopping

Training is halted when validation accuracy stops improving for 5 consecutive epochs. "Stopped epoch" is how many epochs the model trained before this triggered.

Cross-validation (5-fold)

The dataset is divided into 5 groups. Each group takes a turn as the test set while the other 4 train the model. Gives 5 independent accuracy estimates; reduces dependence on any one data split.

Sign test

A simple statistical test that asks: if there were truly no difference, how likely is it to see this many outcomes in one direction just by chance? p = 0.696 means 69.6% chance of seeing this result randomly — not significant.

DoOR dataset

Database of Odorant Responses — a compiled dataset of olfactory receptor neuron measurements from Drosophila melanogaster, aggregated from multiple published studies.

Degree-preserving shuffle

The connections in the ShuffledSNN were randomized, but each neuron kept the same number of inputs and outputs as in the real connectome. Only the specific wiring pattern changed, not the connection density per neuron.

Summary of results from Connectome-Constrained Spiking Neural Networks: Olfactory Classification Study by Moberg et al., 2026

The full research paper detailing our hypothesis, methodology, and interpretation is currently being edited. Follow MIRE or check back here for updates when the preprint is published.