Dagmawi N. Wube¹ (), Najib J. Majaj², Timothy D. Oleskiw^1,2; ¹Department of Computer Science, University of Regina, ²Center for Neural Science, New York University

Planar shape, or the silhouette contour of a solid body, carries rich information important for object recognition, including both local (curvature) and global shape cues. While neurons selective for shape are found within the intermediate cortical area V4 of primates, it remains unknown how populations of these neurons contribute to our perception of objects in natural environments. Recently, we have used a unique array of shape stimuli that dissociates local and global cues to identify V4 neurons that preferentially respond to natural shape features (Oleskiw et al., 2023). To investigate the role of this tuning toward a population-level object code, we analyzed simultaneous activity recorded from a 96-channel ‘Utah’ array implanted in area V4 of a juvenile macaca nemestrina observing natural and synthetic shapes. Linear classifiers trained to identify these stimuli decode natural shapes from our V4 population twice as accurately (Mann-Whitney, p < .05) than synthetic shapes that lack natural cues. Furthermore, a correlation analysis suggests that the V4 population activity is robust to trial-to-trial variability: the decoding of synthetic shapes, including those with identical local curvature statistics, is significantly more sensitive to shared neural noise. To explore the computations underlying a population shape code for natural objects, we repeated our analysis in silico with a variety of artificial neural networks (ANNs) trained for object recognition. Interestingly, while many convolutional networks, including AlexNet and VGG, contain intermediate layers from which natural shapes are decoded somewhat more effectively than synthetic controls, modern transformer architectures (i.e., ConvNeXT) readily utilize natural shape features, demonstrating a 43.5% and 63.1% (p < .01) increase in classification accuracy for local and global cues, respectively. Together, our findings reinforce the notion that modern ANNs are compelling models for understanding the integration of natural shape features seen in V4 neurons.

Acknowledgements: Faculty of Graduate Studies and Research, University of Regina