Máté Lengyel

 

Computational and Biological Learning Lab, Department of Engineering, University of Cambridge

 

The powers and perils of Bayesian inference in the brain: perception, learning, and decision making

 

It is widely believed that the nervous system develops internal models that are adapted to the statistical properties of the environment. However, there has been little, if any, direct evidence for an optimal representation of non-trivial stimulus statistics either at the behavioural let alone at the neural level. I will present results from a series of studies using an unsupervised visual learning paradigm involving high dimensional and complex stimulus statistics and show that humans form internal models of the visual environment in a statistically efficient way. Time permitting, I will also present multi-electrode data of evoked and spontaneous activity from awake behaving ferrets demonstrating that the visual cortex develops statistically optimal internal models of the visual environment. In the second part of my talk, I will present an analysis of the use of internal models for sequential decision making. I will show that even a statistically optimal internal model leads to suboptimal behaviour under realistic resource or time constraints, and that a statistically highly inefficient system based on episodic memories can guide adaptive behaviour more efficiently under ecologically relevant conditions. This offers a normative account of parallel memory systems in the brain.