Runnan Cao¹ (), Jie Zhang¹, Jie Zheng², Yue Wang¹, Peter Brunner³, Jon Willie³, Shuo Wang^1,3; ¹Washington University in St. Louis, ²University of California Davis, ³Washington University in St. Louis

A critical question in cognitive neuroscience is how unified category representations emerge from visual inputs that undergo high-dimensional visual changes, such as visual variations in shape, color, and texture. Two distinct hypotheses have been proposed. On the one hand, non-human primate studies suggest that single neurons in the inferotemporal cortex, which is homologous to the human VTC, encode specific feature axes. On the other extreme, neurons in the human MTL are reported to encode abstract concepts related to individual persons or places, demonstrating a highly selective and sparse code. Little is known about how the perceptual-driven representations in the VTC are transformed into memory-driven representations in the MTL. It is also unclear whether the human VTC encodes visual objects using an axis-based code. To address these questions, we recorded intracranial EEG activity across the VTC and MTL areas, along with single-neuron activity in the MTL. Indeed, the human VTC exhibited strong axis-based feature coding, as shown in primate studies. By constructing a neural feature space using the VTC neural axes, we observed that MTL neurons encode receptive fields within the VTC neural feature space, exhibiting region-based feature coding. This region-based coding of MTL neurons at the low-dimensional feature space, which was derived from the VTC, may serve as an intermediate step that connects dense visual representations to sparse semantic representations. Notably, we uncovered the physiological basis for this coding transformation, showing that the VTC axis-coding channels exhibited stronger synchronization with the MTL. These findings propose a new computational framework for object recognition, advancing our understanding of both visual perception and visual memory.

Acknowledgements: This research was supported by the NIH (K99EY036650, R01MH129426, R01MH120194, R01EB026439, U24NS109103, U01NS108916, U01NS128612, R21NS128307, P41EB018783), AFOSR (FA9550-21-1-0088), NSF (BCS-1945230), and McDonnell Center for Systems Neuroscience.