Margaret M Henderson¹ (), Sungjoon Park¹, Leila Wehbe¹, Michael J Tarr¹; ¹Carnegie Mellon University

Intermediate visual features may be sufficient to support certain types of object categorization, even in the absence of recognizable high-level properties. We hypothesize that the role of these features depends on the granularity of a task: fine-grained distinctions (types of birds) may require high-level, complex features, while coarser distinctions (animals vs. vehicles) may be accessible from low-level features alone. Differences in the diagnostic features relevant to these different tasks may also lead to task-dependent differences in how object images are cortically encoded during coarse and fine categorization. Across both behavioral and fMRI studies, we leveraged a computational texture synthesis procedure (Gatys et al.; 2015, NeurIPS) to generate “texturized” versions of target object images by matching their summary statistics at different layers of a deep neural network. This results in images that vary continuously in their feature complexity. Participants viewed these images and performed a 2-alternative forced choice task discriminating the image category at either a coarser (superordinate) or a finer (basic) level. We found that observers could behaviorally discriminate a subset of object categories at an above-chance level based on the simplest texturized images tested. Categorization of simple texture images was highest for coarse categories (vs. fine), natural objects (vs. artificial), and color images (vs. grayscale). In the brain, we used multivariate classification within higher visual cortex to demonstrate evidence for distributed neural representations of both coarse and fine object categories. As in the behavioral data, discriminability of these representations was lower for texturized as compared to original images, but coarse category information could be decoded with above-chance accuracy even from cortical responses to simple texturized images. Taken together, these results indicate that intermediate visual features contribute to object categorization in a manner that depends on task precision.

Acknowledgements: This study was funded by a grant from Apple Inc to MJT, and a postdoctoral fellowship from the Carnegie Mellon Neuroscience Institute to MMH.