Characterizing representational geometry and learning across primate visual streams

The National Eye Institute awarded The Trustees of Columbia University in the City of New York $235,650 on March 1, 2026, under the Vision Research program (CFDA 93.867) to characterize representational geometry and learning across primate visual streams.

The research examines how the primate visual system balances competing demands—representing objects and places in formats abstract enough for recognition under novel conditions while maintaining specific contextual associations. The work addresses this gap by contrasting joint encoding of multiple objects in inferotemporal cortex and parahippocampal cortex of common marmoset monkeys. Aim 1 investigates representational geometry across cortical space, hypothesizing that representations of different objects in inferotemporal cortex will be comparatively independent while parahippocampal cortex will more nonlinearly encode groups of objects. Artificial neural networks trained on various loss functions, architectures, and training diets will model these phenomena. Aim 2 investigates how representational geometry varies over familiarization with a novel virtual environment, hypothesizing that neural representations of different views will acquire the same correlation structure as the images themselves in underlying latent factor space. The research will suggest new neural network architectures for modeling vision and help predict effects of stroke or lesion on different brain regions.

Performance occurs in New York, New York, with a period of performance from March 1, 2026, through August 31, 2028.

Abstract

The primate visual system must balance competing demands, including representing objects and places in a format that is both abstract enough to be recognizable under novel conditions, but also specifically bound to other objects and places via contextual associations. How the visual system achieves this balance remains to be understood, as current artificial neural network (ANN) models fail to exhibit this same robustness and flexibility. One notable difference between anns and the brain is that the former preponderantly model the visual system as a single feedforward hierarchy, while ample evidence suggests that the latter in reality includes numerous parallel hierarchies with possibly complementary propensities for abstraction versus granularity. Aim 1 addresses such variations in representational geometry over cortical space by contrasting joint encoding of multiple objects in inferotemporal cortex (IT) and parahippocampal cortex (PHC) of common marmoset monkeys. I hypothesize that representations of different objects in IT will be comparatively independent of each other, while PHC will more nonlinearly encode groups of objects. Artificial neural networks trained on various loss functions, architectures, and training diets will then be used to attempt to model these phenomena. Aim 2 will investigate variation in representational geometry over the course of familiarization with a novel virtual environment over several days. I hypothesize that neural representations of different views of the learned environment will acquire the same correlation structure as the images themselves in the underlying latent factor space. This research will suggest new, non-standard neural network architectures for modeling vision, help US make more precise predictions about the effects of stroke or lesion on different parts of the brain, and better understand the underpinnings of disorders like attention deficit hyperactivity disorder, autism, frontotemporal dementia, and face and object agnosias. The training plan described in this proposal will afford me proficiency in new experimental and theoretical methodologies, including large-scale, multi-area electrophysiological recording, the use of common marmosets as a model organism for visual neuroscience, deep neural network modeling, and advanced data analysis techniques for understanding the structure of neural representations. I will receive expert training in electrophysiology and machine learning from my sponsor DR. Elias issa, as well as in theoretical neuroscience and advanced neural data analysis methods from my CO-sponsor DR. Stefano fusi. This work will be conducted at columbia university's zuckerman institute, a world-class neuroscience research and training institution whose faculty specialize in approaches ranging from molecular to systems to theoretical.