Umael Qudrat¹, Shirin Taghian Alamooti¹, Judith Fan², Kohitij Kar¹; ¹York University, ²Stanford University

Visual abstraction is the process of distilling complex visual scenes into their essential components. Such abstraction is exemplified by the production and recognition of sketches, which can be effective in conveying the content of a scene while omitting many details. To what degree does visual abstraction also manifest in non-human primates, and what neural computations are responsible? To answer this question, we measured how well macaque monkeys (N=2) could identify the visual concept conveyed in human-drawn sketches and evaluated how well their behavior and ventral stream neural responses could be predicted by an artificial neural network (i.e., AlexNet). Both macaques performed a sketch-recognition task using stimuli (1000 images, 10 object categories) from the Google Quick Draw dataset. In each trial, we briefly presented (100 ms) a sketch, followed by a choice screen where monkeys selected which of two object images the sketch represented. Both monkeys achieved accuracies exceeding 70%, demonstrating that even simple sketches convey sufficient information for robust object identification by non-human primates. The monkeys’ image-by-image recognition accuracies significantly correlated with those predicted by AlexNet (R = 0.42, p<0.001). This correlation matched the monkeys’ noise ceiling (~0.4), indicating that AlexNet strongly approximates the abstraction strategies employed by the primate visual systems, given intrinsic variability in the current dataset. We also recorded population activity (384 sites) from inferior temporal (IT) cortex as monkeys viewed line drawings and sketches. Using linear classification on IT responses, we found that distributed neural activity patterns strongly predicted (accuracy~0.83, chance-level=0.5) object identity, providing evidence that IT cortex encodes the abstract visual features underlying sketch recognition. These findings establish macaques as a powerful model for investigating the neural computations that support sketch recognition.

Acknowledgements: KK was supported by funds from the Canada Research Chair Program, the Canada First Research Excellence Funds (VISTA Program), Google Research Award, and a NSERC-DG. JEF is supported by NSF CAREER #2047191, NSF DRL #2400471, and a Stanford Human-Centered AI Institute Hoffman-Yee award.