Abstract

A central feature of perception is that our internal expectations to a large degree shape how we perceive the world. Predictive coding (PC) is a popular framework to explain cortical responses to the violation of our sensory expectations. In the first part of the talk, Prof. Máté Lengyel will introduce the theory of predictive coding, and clarify how PC and Bayesian inference can be distinguished, despite sharing core computational concepts and addressing an overlapping set of empirical phenomena. He will argue that predictive coding is an algorithmic/representational motif that can serve several different computational goals of which Bayesian inference is but one. In the second half of the talk, Daniel Kornai (1st year CBL PhD) will discuss results from three recent papers on the topic of predictive coding: Jiang and Rao (PLOS Comp. Biol, 2024) show how their implementation of a Dynamical Predictive Coding model qualitatively reproduces several phenomena in visual processing, including the emergence of spatio-temporal receptive fields, a temporal and representational hierarchy, and predictive/postdictive artefacts in human motion perception. https://doi.org/10.1371/journal.pcbi.1011801 A core hypothesis of predictive coding is that the content of error signals should reflect the difference between predicted and observed sensory input. Furutachi et al. (Nature, 2024) use calcium imaging data from layer 2/3 neurons in V1 to infer the content of the error signal induced by a violation of a learned sequence to test this claim. https://doi.org/10.1038/s41586-024-07851-w Predictive coding also implies that violations of learned sequences should induce a robust feedforward error signal that propagates up the visual processing hierarchy. Westerberg et al (preprint, 2024) use spiking data from mice and monkeys collected during an oddball task to investigate if the path of the error signal through various neuronal subpopulations involved in visual processing match the expectations of PC. https://doi.org/10.1101/2024.10.02.616378