Sampling-based inference under hierarchical probabilistic models to study perception and neural dynamics in the visual cortex

JOSEFINA CATONI; RODRIGO ECHEVESTE

Montevideo

Conferencia; Latin American Meeting In Artificial Intelligence, Khipu 2023; 2023

UDELAR

The Bayesian theory of visual perception in Neuroscience assumes the brain performs probabilistic inference to estimate probability distributions over variables of interest given an observed stimulus. To understand how this process could take place we train neural networks to perform Bayesian inference under hierarchical generative models of perception. Inference in these networks is done sampling, employing the dynamics of recurrent modules. These types of networks constitute useful models of cortical circuits of perception, being able to capture not only mean responses and neural variability but also cortical dynamics, including: oscillations, transient responses, and temporal cross-correlations. We focus on visual inference from natural images, and aim to learn both a generative model for these images and an inference model. As a first step we are training Variational Autoencoders whose modules will serve as silver-ground truth for our recurrent ones. These models will be applied to the study of neurotypical visual perception, and also to improve the understanding of the link between physiology and sensory perception in individuals with autism.