Dear All,

On Thursday 17.10 we will host dr. Gergő Orbán, Computational Systems Neuroscience Lab, Wigner Institute. His lecture will be entitled: Normative theories of two-way interactions between episodic and semantic memories

Abstract

Acquiring internal models of the natural environment is central to efficient perception, and subsequent decisions. Generative models have been proposed as a normative framework to characterise these internal models. In this talk I will elaborate on a pair of complementary computational challenges when we consider internal models as the semantic knowledge we store about our experiences. First, I will consider how this internal model might support encoding specific experiences, also referred to as episodic memories. I argue that the internal model can be tightly linked to the normative theory of lossy compression, called rate distortion theory. Using tools borrowed from machine learning, deep generative models, I demonstrate that rate distortion theory can account for a wide range of memory distortions from domains as diverse as hand drawings, board games, and language memory. Second, I will discuss how episodic memory can provide support to updating the internal model. We highlight that Bayesian learning of internal models requires that alternative potential internal models are maintained in parallel for optimal acquisition of the internal model with the right structure. We show how an extension of Bayesian learning leads to the requirement of an episodic memory system. We show that this episodic memory-supported form of learning accounts for aspects of curriculum-dependence of learning in humans.

The lecture will take place at 3 p.m. in the CN lecture hall and it will be followed by a get together.

Hope to see you all there,

Tomasz Wypych