ConnectomeX: Connectome-Informed Computer Vision
Translating biological wiring motifs from the Drosophila visual connectome into scalable, parameter-efficient artificial neural networks.
Overview
Can biological connectomes provide reusable inductive biases for artificial intelligence? In ConnectomeX (FlyVision), we investigated whether the detailed circuit diagram of the Drosophila melanogaster visual system offers architectural blueprints that scale to modern computer vision benchmarks and clinical imaging tasks.
Approach
Rather than treating connectomics purely as an object of reverse engineering, FlyVision incorporates conserved circuit principles directly into artificial neural networks:
- ON and OFF pathways: Splitting visual inputs into rectified polarity channels inspired by early retinal and lamina processing.
- Recurrent population mixing: Multi-population graph interactions across recurrent stages.
- Lateral competition: Inhibitory dynamics and local competition that regulate representation capacity.
Results
FlyVision scales from small-scale vision benchmarks (MNIST, CIFAR-10) to ImageNet-1K, dermatological lesion classification, and 3D brain MRI age regression (BrainAGE). Across these domains, connectome-informed inductive biases achieve competitive accuracy while maintaining high parameter efficiency, operating at under 4M parameters compared to standard vision backbones.
Preprint
Preprint is available on arXiv: arXiv:2610.08418.