# Garcia-Duran and coauthors link fly visual-model failure to a perceptual illusion [Canonical reading](https://escapemeridian.com/readings/garcia-duran-and-coauthors-link-fly-visual-model-failure-to-a-dbcc6ab9) The authors of an arXiv paper compare FlyNet, a task-optimized, continuous-time recurrent neural network constrained by the fly visual system’s synaptic connectome, with artificial neural networks on a collision-detection task. They report that FlyNet is vulnerable to adversarial perturbations but, relative to its compute, more robust to common video corruptions than RAFT, while using up to five orders of magnitude fewer FLOPs. Continuous-time recurrent neural networks without detailed connectome constraints were more robust than FlyNet across almost all tested perturbations. The authors suggest that connectome constraints benefit efficiency through low edge count rather than robustness. They also report that adversarial textures and the reverse-phi illusion disrupt FlyNet’s position and velocity representations through the same low-level motion statistic; optimizing that statistic alone induced classification failure. Escape Meridian’s reading: the authors present this model vulnerability as a testable hypothesis about biological perception, not evidence that it occurs in biological flies. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv q-bio.NC · Article Original source date: Nov 10, 2026 (2026-11-10). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-11T22:18:59.996Z. First observed: 2026-11-11T05:01:20.154Z. Retrieved: 2026-11-11T22:17:45.215Z. - [Measuring Robustness and Efficiency in a Connectome-Constrained Fly Visual System Model: Linking an Adversarial and a Perceptual Illusion](https://arxiv.org/abs/2611.14498) ## Limits The evidence supplied is the arXiv abstract, not the paper's full methods or results. The reported comparisons concern a collision-detection task and the tested perturbations; the abstract does not establish performance on other tasks. The abstract proposes testable hypotheses about biological perceptual vulnerabilities but does not report whether the identified vulnerability occurs in biological flies. ## Open question Does the circuit vulnerability identified in FlyNet also occur in the biological fly visual system? ## Source-stated The authors benchmarked a connectome-constrained FlyNet against artificial neural networks on a collision-detection task. > We 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. - [Supporting source](https://arxiv.org/abs/2611.14498) ## Source-stated The authors report that FlyNet was more robust to common video corruptions than RAFT relative to its compute, while using up to five orders of magnitude fewer FLOPs. > FlyNet is adversarially vulnerable but, relative to its compute, more robust to common video corruptions than RAFT, while using up to five orders of magnitude fewer FLOPs. - [Supporting source](https://arxiv.org/abs/2611.14498) ## Source-stated Continuous-time recurrent neural networks without detailed connectome constraints were more robust than FlyNet across almost all tested perturbations; the authors say connectome constraints benefit efficiency through low edge count over robustness. > The CTRNNs remain more robust than FlyNet across almost all tested perturbations, suggesting that connectome constraints benefit efficiency, through their low edge count, over robustness. - [Supporting source](https://arxiv.org/abs/2611.14498) ## Source-stated The authors report that adversarial textures and the reverse-phi illusion disrupt FlyNet's position and velocity representations through the same low-level motion statistic. > The reverse-phi illusion similarly disrupts these representations and inverts the same low-level motion statistic as the adversarial textures. - [Supporting source](https://arxiv.org/abs/2611.14498) ## Escape Meridian interpretation The authors present connectome-constrained models as a source of testable hypotheses about biological perceptual vulnerabilities, not as evidence that the model's vulnerability occurs in biological flies. > Connectome-constrained models can thus surface adversarial failure mechanisms as testable hypotheses about biological perceptual vulnerabilities. - [Supporting source](https://arxiv.org/abs/2611.14498) ## Source and contributor context Garcia-Duran, Alexandre, Rodriguez-Garcia, Alejandro, Molano-Mazón, Manuel, Hyafil, Alexandre, Ramaswamy, Srikanth, Lappalainen, Janne K. · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Neural network A system of connected simple calculations, loosely inspired by the brain, that learns patterns from data. - [Definition reference](https://developers.google.com/machine-learning/glossary#neural-network) ## Floating-point operations A count of basic arithmetic steps; used to measure how much computation went into training a model. - [Definition reference](https://en.wikipedia.org/wiki/FLOPS) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 33d96823-6b88-4b1e-8d7c-105a87f81358. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-11T22:18:59.996Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 561db2b98ac06e7809850a223f54f1ce684d02b609d47529574fd953a853d22f. Source text hash: ac40ff190dfb85a35664f78201f0e60100bde01b5364c9cedaefa7315d020283. ## Revision and source-check history ### Version 1 — 2026-11-11T22:18:59.996Z The authors of an arXiv paper compare FlyNet, a task-optimized, continuous-time recurrent neural network constrained by the fly visual system’s synaptic connectome, with artificial neural networks on a collision-detection task. They report that FlyNet is vulnerable to adversarial perturbations but, relative to its compute, more robust to common video corruptions than RAFT, while using up to five orders of magnitude fewer FLOPs. Continuous-time recurrent neural networks without detailed connectome constraints were more robust than FlyNet across almost all tested perturbations. The authors suggest that connectome constraints benefit efficiency through low edge count rather than robustness. They also report that adversarial textures and the reverse-phi illusion disrupt FlyNet’s position and velocity representations through the same low-level motion statistic; optimizing that statistic alone induced classification failure. Escape Meridian’s reading: the authors present this model vulnerability as a testable hypothesis about biological perception, not evidence that it occurs in biological flies. ## Original source arXiv q-bio.NC · Article Original source date: Nov 10, 2026 (2026-11-10). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-11T22:18:59.996Z. First observed: 2026-11-11T05:01:20.154Z. Retrieved: 2026-11-11T22:17:45.215Z. - [Measuring Robustness and Efficiency in a Connectome-Constrained Fly Visual System Model: Linking an Adversarial and a Perceptual Illusion](https://arxiv.org/abs/2611.14498) ## Limits The evidence supplied is the arXiv abstract, not the paper's full methods or results. The reported comparisons concern a collision-detection task and the tested perturbations; the abstract does not establish performance on other tasks. The abstract proposes testable hypotheses about biological perceptual vulnerabilities but does not report whether the identified vulnerability occurs in biological flies. ## Open question Does the circuit vulnerability identified in FlyNet also occur in the biological fly visual system? ## Source-stated The authors benchmarked a connectome-constrained FlyNet against artificial neural networks on a collision-detection task. > We 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. - [Supporting source](https://arxiv.org/abs/2611.14498) ## Source-stated The authors report that FlyNet was more robust to common video corruptions than RAFT relative to its compute, while using up to five orders of magnitude fewer FLOPs. > FlyNet is adversarially vulnerable but, relative to its compute, more robust to common video corruptions than RAFT, while using up to five orders of magnitude fewer FLOPs. - [Supporting source](https://arxiv.org/abs/2611.14498) ## Source-stated Continuous-time recurrent neural networks without detailed connectome constraints were more robust than FlyNet across almost all tested perturbations; the authors say connectome constraints benefit efficiency through low edge count over robustness. > The CTRNNs remain more robust than FlyNet across almost all tested perturbations, suggesting that connectome constraints benefit efficiency, through their low edge count, over robustness. - [Supporting source](https://arxiv.org/abs/2611.14498) ## Source-stated The authors report that adversarial textures and the reverse-phi illusion disrupt FlyNet's position and velocity representations through the same low-level motion statistic. > The reverse-phi illusion similarly disrupts these representations and inverts the same low-level motion statistic as the adversarial textures. - [Supporting source](https://arxiv.org/abs/2611.14498) ## Escape Meridian interpretation The authors present connectome-constrained models as a source of testable hypotheses about biological perceptual vulnerabilities, not as evidence that the model's vulnerability occurs in biological flies. > Connectome-constrained models can thus surface adversarial failure mechanisms as testable hypotheses about biological perceptual vulnerabilities. - [Supporting source](https://arxiv.org/abs/2611.14498) ## Source and contributor context Garcia-Duran, Alexandre, Rodriguez-Garcia, Alejandro, Molano-Mazón, Manuel, Hyafil, Alexandre, Ramaswamy, Srikanth, Lappalainen, Janne K. · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Neural network A system of connected simple calculations, loosely inspired by the brain, that learns patterns from data. - [Definition reference](https://developers.google.com/machine-learning/glossary#neural-network) ## Floating-point operations A count of basic arithmetic steps; used to measure how much computation went into training a model. - [Definition reference](https://en.wikipedia.org/wiki/FLOPS) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 33d96823-6b88-4b1e-8d7c-105a87f81358. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-11T22:18:59.996Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 561db2b98ac06e7809850a223f54f1ce684d02b609d47529574fd953a853d22f. Source text hash: ac40ff190dfb85a35664f78201f0e60100bde01b5364c9cedaefa7315d020283. ### passed — 2026-11-11T22:18:59.996Z — anthropic/claude-opus-5.5 Publication: 33d96823-6b88-4b1e-8d7c-105a87f81358 attribution — pass: The reading correctly credits the six-author arXiv paper, led by Garcia-Duran. The title names him and coauthors, and findings are framed as what the authors report or suggest. The 'testable hypotheses' framing is the authors' own conclusion, and the EM-labeled sentence keeps it attributed. > Authors:Alexandre Garcia-Duran, Alejandro Rodriguez-Garcia, Manuel Molano-Mazón, Alexandre Hyafil, Srikanth Ramaswamy, Janne K. Lappalainen summary — pass: The reading faithfully states the core findings: an efficiency advantage relative to compute, CTRNNs being more robust, and the adversarial and reverse-phi link. It keeps hedges such as 'relative to its compute', 'almost all' and 'suggest'. Technical terms are kept exact. Suggested term-library tooltips: connectome, RAFT, FLOPs, adversarial perturbation, reverse-phi illusion and CTRNN. The reading omits the counterfactual-texture control and the RAFT recurrence finding, but neither omission changes the meaning. > Optimizing this input statistic alone, without access to the network, suffices to induce classification failure, attributing both perturbations to the same circuit vulnerability. date — pass: The original publication date, 2026-11-10, matches the arXiv v1 submission date. The feed date of 2026-11-11 is correctly not used. > \[Submitted on 10 Nov 2026\] support — pass: Each claim matches the abstract verbatim in substance. The limitations correctly note that the evidence is only the abstract and that the results are specific to one task. They also note that no biological validation is reported. The question follows from the authors' hypothesis framing without asserting an answer. No benefits or costs are invented. > Connectome-constrained models can thus surface adversarial failure mechanisms as testable hypotheses about biological perceptual vulnerabilities. media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct way to reach the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/dbcc6ab9d5b3a8443daca33cdc5a2563261bb9773de6461bb132a9100c26e906/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/garcia-duran-and-coauthors-link-fly-visual-model-failure-to-a-dbcc6ab9.json)