UCPH Statistics Seminar: Jack Brady

Title: Towards a Theory of Compositional Generalization in Representation Learning

Speaker: Jack Brady from Max Planck Institute for Intelligent Systems

Abstract:

Generalizing to situations which consist of novel combinations of known concepts is a hallmark of natural intelligence, yet remains elusive for current machine learning models. Recently, however, a theoretical understanding of how such compositional generalization can be realized in machines has begun to emerge. In this talk, we will present this developing theory of compositional generalization from the ground-up, first providing a rigorous formalization of the problem and then discussing a series of results in this direction. We will then discuss how these results can be leveraged empirically as well as their practical implications. Finally, we will give an overview of open problems and future directions.

Bio:

Jack Brady is a third year PhD student at the Max Planck Institute for Intelligent Systems in Tübingen, Germany, supervised by Wieland Brendel and Thomas Kipf. His research aims to develop a mathematical understanding of key aspects of natural intelligence, such as compositional generalization and building abstract world models, and to leverage these insights to endow machine learning models with similar capabilities.