FlyVision, a computer-vision model built around the organization of the Drosophila visual circuit, matched a much larger standard network on a basic image test. On MNIST, it reached 99.34% top-1 accuracy and 99.33% macro-F1 with 80,608 parameters when trained from scratch. An ImageNet-initialized ResNet18 reached 99.38% accuracy with 11,181,642 parameters.
That result was part of a wider modeling study. It combined deterministic MNIST and CIFAR-10 benchmarks, ImageNet-1K pretraining, biomedical classification, a hierarchical chest-radiography evaluation and a volumetric T1-weighted MRI extension called BrainAGE. The checkpoint-compatible FlyVision implementation uses parallel ON/OFF processing, recurrent graph-mixing stages, adaptive global pooling, an embedding layer and a task-specific prediction head.
Transfer did not always help
The transfer pattern was not uniform when models moved between datasets. On CIFAR-10, the reported transfer difference relative to scratch baselines was -0.41 percentage points for FlyVision, while direct ImageNet-initialized ResNet18 showed a 2.40-point difference. The comparison indicates that transfer behavior varied by architecture and initialization route.
On ImageNet-1K validation, FlyVision Base reached 60.79% top-1 accuracy with 1.8 million parameters, while FlyVision Large reached 66.25% with 3.7 million. A Large local-k7 model with a learned low-frequency branch reached 66.53%, below ResNet18's 69.25% with 11.7 million parameters. In this comparison, the smaller parameter counts came with lower top-1 accuracy.
The biomedical tests split the difference
The skin-disease test used a 22-class, image-level split of 11,506 training images, 2,042 validation images and 1,538 test images; patient or lesion grouping was unavailable. On that split, FlyVision Large achieved 63.78% accuracy, 60.23% macro-F1 and 95.28% macro-AUROC with 2.99 million parameters. ResNet18 reported higher point estimates of 66.19%, 61.96% and 95.96% with 11.19 million parameters.
Chest radiography gave FlyVision higher point estimates in the held-out comparison. The release contained 21,165 image-mask pairs, split into 14,815 training images, 3,175 validation images and 3,175 test images under a deterministic class-stratified 70%/15%/15% allocation. ImageNet-pretrained FlyVision Large reached 92.76% accuracy and 92.85% macro-F1. Its 95% bootstrap interval for accuracy was 91.83% to 93.66%, while FlyVision Base reached 92.60% accuracy and 92.75% macro-F1. ResNet18 reached 91.56% accuracy and 91.98% macro-F1. The radiography intervals were estimated from 300 bootstrap resamples.
Adding a hierarchy did not improve the final four-class point estimates. The flat model scored 92.60% accuracy and 92.75% macro-F1. A two-stage hierarchy scored 91.53% and 91.68%, while a multi-head hierarchy scored 91.75% and 92.17%.
A useful benchmark result, not a clinical verdict
BrainAGE extended the test to scan-level prediction from T1-weighted MRI. The cohort comprised 2,851 scans from 13 source datasets and used participant-grouped 70:15:15 partitioning, leaving 433 held-out scans. Three-axis fusion produced a mean absolute error of 5.98 years, with a 95% bootstrap interval of 5.54 to 6.46 years. RMSE was 7.67 years, R2 was 0.868 and Pearson correlation was 0.931. The BrainAGE intervals were estimated from 1,000 bootstrap resamples.
The pattern is narrower than a claim that FlyVision is better everywhere. It matched or exceeded ResNet18 in some reported tests and used fewer parameters in comparisons that provided parameter counts, but it trailed on ImageNet and the skin-disease split. The study measures model performance on named datasets, not patient outcomes. It also does not show that any particular Drosophila-inspired feature caused the result, because matched component-level ablations were not reported.
That distinction matters most for the biomedical results. The skin-disease and radiography evaluations were image-level, and the review reports no patient-, site- or scanner-separated testing. BrainAGE used participant-grouped splitting, but the supplied evaluation does not establish whether its scan-level performance would hold in independent cohorts or deployment settings.
Reproducibility is another caveat. The authors say ConnectomeX source code and manuscript-reproduction resources are proprietary, while supporting data may be requested from the corresponding author where permitted. The document is a preprint identified as arXiv:2610.08418v1, dated 6 Oct 2026.
Paper data and sources
Original title: From the Drosophila Visual Connectome to General-Purpose Computer Vision
Authors: Zongyu Li, Akito Yamauchi, Huaizhi Liu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-10-06
DOI: Not available
Original paper · Full text