Measuring Robustness and Efficiency in a Connectome-Constrained Fly Vision Model
Field: Computer Vision Authors: Alexandre Garcia-Duran, Alejandro Rodriguez-Garcia, Manuel Molano-Mazón, Alexandre Hyafil, Srikanth Ramaswamy, Janne K. Lappalainen Published: 2026-11-10 arXiv: 2611.14498
Abstract
Biological visual systems compute reliably under tight metabolic and wiring constraints, whereas artificial vision systems consume far more energy and can be sensitive to adversarial perturbations. Whether networks with brain-like wiring are more robust at lower computational cost is unclear. The authors address this by benchmarking a task-optimized, continuous-time recurrent neural network of the fly visual system constrained by its synaptic connectome (FlyNet) against artificial neural networks. FlyNet is compared with CNNs, RAFT, and continuous-time recurrent neural networks without detailed connectome constraints (sCTRNN, CTRNN) on a collision-detection task within a differentiable 3D rendering pipeline.
Key Findings
- Adversarial vulnerability vs. corruption robustness: FlyNet is adversarially vulnerable but, relative to its compute, more robust to common video corruptions than RAFT, using up to five orders of magnitude fewer FLOPs.
- Efficiency, not robustness, from connectome constraints: Continuous-time RNNs without connectome constraints are more robust than FlyNet under almost all perturbations, suggesting the connectome's low edge count yields efficiency benefits rather than robustness.
- Recurrence and temporal context: Recurrent structure improves RAFT's robustness to white-box attacks, and temporal context improves robustness to common corruptions.
- Representation analysis: FlyNet's position and velocity representations degrade under adversarial textures but recover under counterfactual textures that restore correct classification while being equally far from the training distribution — ruling out "unfamiliar textures" as the failure cause.
- Reverse-phi illusion link: The reverse-phi illusion similarly disrupts these representations and flips the same low-level motion statistics as adversarial textures. Optimizing only this input statistic (without access to the network) is sufficient to induce classification failures, attributing both perturbations to the same circuit-level vulnerability.
Conclusion
Connectome-constrained models can transform adversarial failure mechanisms into testable hypotheses about biological perceptual vulnerabilities.
--- *Auto-collected on 2026-11-12*