Friday, 10th March 2023 - 11am Emile van Krieken : Seminar

Title:  Optimisation of Neurosymbolic Learning Systems

Abstract:

Recent work in Neurosymbolic Learning studies how to combine neural networks with background knowledge. This talk gives an overview of my PhD work, which explores the optimisation dynamics of Neurosymbolic Learning systems. We find two challenges for Neurosymbolic Learning: Firstly, the discrete-continuous gap, i.e. how to optimise deep systems with continuous and discrete reasoning. Secondly, the symbol grounding problem, i.e. how to use deep learning to give real-world meaning to symbols. We study how ideas from fuzzy and probabilistic logics tackle these challenges. We find that the gradients of these logics perform symbol grounding. Furthermore, we study how ideas from gradient estimation can help scale probabilistic logics and develop A-NeSI, a highly scalable approximate inference method for probabilistic logics.

Bio:

Emile van Krieken is a PhD student in the Learning and Reasoning group at the Vrije Universiteit Amsterdam. His research combines symbolic reasoning and machine learning, or "Neurosymbolic Learning". He studied fuzzy logics for Neurosymbolic reasoning and developed Storchastic, a general framework for stochastic automatic differentiation. More recently, he introduced A-NeSI, a highly scalable neurosymbolic method that uses neural networks for symbolic inference.