Ann Hermunstad: Inductive biases for rapid learning

Janelia Research Campus

Abstract

Animals can learn to modify their behavior from limited experience in a new setting. This rapid flexibility is enabled in part by smart inductive biases that capture existing knowledge about predictable structure in an animal’s surroundings and its relationship to those surroundings. Using navigational learning tasks for flies and mice, I will show how behavioral data constrains compact generative models of action selection and rapid learning. In flies, we map this model onto genetically-specified circuit architectures that combine multiple evolving internal representations to guide behavior, thereby translating abstract inductive biases into concrete mechanistic implementations. Finally, I will discuss the implications of behavioral and circuit level constraints on rapid learning and behavioral flexibility in biological and artificial agents more broadly.

 

Guests are welcome!

 

Organized by

Henning Sprekeler / Margret Franke