Addressing Computational Challenges in Bayesian Perception


Bobby Ikonomov

The Bayesian model of perception assumes that the brain uses prior information based on past experiences in order to process sensory stimuli in a fashion similar to Bayesian inference. However, existing Bayesian inference methods often involve computational challenges that limit their applicability. This project aims to improve these existing inference methods, develop new ones, and use them to further our understanding of perception. Currently, we are working on a procedure that combines top-down processing in the form of a generative model (a differentiable renderer), bottom-up processing via deep learning-based recognition models, and approximate Bayesian inference through the use of Approximate Bayesian Computation.


Supervisors: Michael Gutmann & Chris Williams