Fan L. Cheng¹, Isabella E. Rosario¹, Zitang Sun², Nikolaus Kriegeskorte¹; ¹Columbia University, ²Independent Researcher

Anomalous motion illusions—including peripheral drift illusion (e.g., Rotating Snakes), central drift illusion, and the Ouchi illusion—offer a powerful probe into the visual computations that transform small image shifts on the retina into vivid motion percepts. For the Rotating Snakes illusion, no single explanation has gained consensus. Proposed neural accounts include contrast-dependent differences in response timing within V1 and MT, nonlinear saturation of motion detectors, and biased local motion estimates arising from early visual processing. These mechanisms suggest that multiple neural computations may jointly contribute to the illusory percept. To examine this possibility, we tested whether modern neural network models can reproduce motion illusions and reveal their underlying computational ingredients. We evaluated ten motion-estimation models—spanning standard deep architectures and biologically inspired models—under simulated viewing conditions (static, onset, microsaccade, and saccade). Across models, the bio-inspired Dual model showed the closest—though still partial—alignment with human perception of Rotating Snakes, and its responses were strongest during microsaccade-like image shifts. In contrast, the Ouchi illusion is well accounted for by models with first-order motion-energy mechanisms, whereas central drift stimuli remain poorly captured by all models. Ablation analyses of the Dual model reveal that first-order mechanisms alone cannot reproduce Rotating Snakes: higher-order motion features increase alignment with expected rotational flow, and recurrence is essential for integrating local cues into a coherent global pattern. Thus, the candidate mechanism identified for Rotating Snakes goes beyond classic first-order latency accounts by requiring additional higher-order and recurrent processing. Our results show that video-computable neural networks can generate testable hypotheses about mechanisms of anomalous motion illusions. The framework highlights gaps between human and machine motion perception and suggests new directions for building models that more closely mirror the computations of the human visual system.