Inference via Abstraction: A framework for efficient Bayesian cognition

We propose a novel framework for efficient Bayesian cognition called Inference via Abstraction (IvA), which learns to approximate complex world models with simpler abstractions that capture main dependencies, but leverage structure in the prior distribution for efficient inference. We instantiate IvA with a combination of probabilistic graphical models and deep neural networks. We hypothesize that this combination is key to efficient Bayesian inference in complex environments. We will make core methodological contributions that demonstrate IvA for perceptual inference tasks, and partner with experimental colleagues to test IvA as a model of human inference using both behavioral and neural measurements.