Syed Suleman Abbas Zaidi¹ (), Kohitij Kar²; ¹Massachusetts Institute of Technology, ²York University

Visual perception in artificial and biological systems often diverges in the processing of complex stimuli. Geirhos et al. (2019) underscored this distinction, demonstrating that unlike convolutional neural networks (CNNs) trained on ImageNet, human vision is biased by shape over texture in object recognition. To delineate the mechanistic differences in visual perception between CNNs and biological systems, we probed the object shape versus texture biases of rhesus macaques—an animal model where finer-grained neural measurements are feasible. We trained two macaques on binary object discrimination tasks using the Microsoft COCO dataset across ten object categories. They were subsequently tested on cue-conflict images (from Geirhos et al. 2019), wherein images featured either texture-shape congruence or conflict – designed to assess whether macaques exhibit shape-bias like humans. Our results revealed a nuanced perceptual strategy in macaques. Consistent with previous studies, we observed high accuracy when the images contained no shape-texture conflicts –indicating that macaques are adept at shape-based recognition, with performances ranging from 0.80 to 0.89 across shapes. However, the introduction of conflicting textures led to variable outcomes. In particular, the accuracy for recognizing 'Elephant' shapes with 'Chair' textures dropped sharply to 0.14, highlighting a substantial influence of texture on the recognition process. The performance gradient across various shape-texture pairings suggests a complex interplay in the macaques' visual processing, differing significantly from the consistent human shape bias reported earlier. Next, we asked how these behavioral biases were driven by activity in the macaque IT cortex (critical for object recognition). We observed a significant alignment (consistency of 0.36) between neural activity and cue-conflict confusion pattern. In conclusion, our results reveal that macaques' reliance on shape versus texture is context-dependent and less robust than in humans. These insights motivate further exploration of the factors influencing distinct perceptual biases and the evolution of visual processing across species.

Acknowledgements: Google Research, CFREF, Brain Canada, SFARI