Akshay V Jagadeesh¹, Sohrab Najafian¹, Michael J Arcaro², Margaret S Livingstone¹; ¹Harvard Medical School, ²University of Pennsylvania

Neurons in primate visual cortex encode visual information within localized regions of visual space, i.e., receptive fields. In early visual areas, spatial receptive fields are well-characterized using simple stimuli like bars or gratings. However, in high-level areas such as inferior temporal (IT) cortex, the complex feature selectivity of neurons makes mapping their spatial receptive fields more challenging. To understand how spatial and featural selectivity are integrated in IT cortex, we measured multiunit electrophysiological responses throughout macaque IT cortex. Monkeys passively viewed a set of naturalistic images, including faces, face parts, hands, objects, and scenes, presented in a grid spanning the central 17 degrees of visual space. Using a Gaussian receptive field model, we estimated reliable receptive field positions for face-selective units and observed shifts in receptive field position based on stimulus type. These shifts were explained by spatially non-uniform contributions of face parts (e.g., eyes, mouth), emphasizing the necessity of jointly modeling spatial and featural selectivity for accurately characterizing IT neuron response properties. To address this, we developed a novel patch-based deep neural network model that simultaneously estimates spatial and featural tuning of neurons. When tested on neural responses to natural scene images, this model identified diverse receptive field structures with spatially-dependent feature tuning, evidenced by differential feature encoding across receptive field subregions. Even for neurons in anterior IT cortex, classically thought to exhibit position invariance, the model uncovered differential tuning for specific face features at distinct spatial locations, such as preference for eyes in the upper visual field and mouths in the lower visual field. This approach demonstrates evidence for precise spatial coding of features in high-level visual cortex and provides a formal modeling framework to characterize the interaction between spatial and featural coding in brain areas with complex tuning properties.