Hannah Stone

Presentation Title

A Biologically Informed Model of V1 Predicts Phosphenes from Cortical Stimulation  

Abstract

Major research efforts over the past few decades have focused on developing visual cortical prostheses (VCPs). These devices aim to provide functional vision to blind individuals by electrically stimulating the visual cortex. However, the visual sensations created by VCPs are often distorted, and predicting what a patient will see from a given stimulation remains a major challenge. At its core, this problem reflects a gap in our understanding of how the brain interprets artificial V1 activity as visual perception. To this end, we developed a biologically informed, end-to-end model of human V1 that combines an encoding model with a learned decoder. We incorporate known spatial and functional properties of V1 to generate population responses to natural images, and use a trained linear decoder to reconstruct those images from V1 activity. Using this model, we simulate electrical stimulation of V1 and quantify the resulting phosphene predictions. The model reproduces several known properties of phosphenes, including increases in phosphene size with electrode eccentricity, electrode size, and stimulation strength, as well as changes in phosphene intensity as a function of eccentricity and stimulation strength. By combining a biologically informed encoder with a learned readout, the model not only predicts realistic phosphenes, but also explains how these properties arise from the interaction between V1 organization and decoding.