Statistical Relational AI - Exploiting Symmetries

Basic data for this talk

Information about the event

Abstract

In recent years, a need for compact representations of databases has become apparent, for example in natural language understanding, machine learning, and decision making, accompanied by a need for efficient inference algorithms. This contrasts with the claim to represent large numbers of individuals, objects, and relations among them, as well as uncertainties about attribute values, object identities, or even the existence of individuals, objects, and relations. From this area of tension, many advances have emerged in the field of probabilistic relational modeling for artificial intelligence, also known as statistical relational AI (StaRAI), as addressed in David Poole’s Keynote Talk at KR2020. A central concept that many approaches in StarAI have in common is the exploitation of symmetries, for example caused by the indistinguishability of certain individuals or objects. In this tutorial, we give a general introduction into StaRAI and show how StaRAI advances long standing problems in the field of AI by exploiting symmetries. In a first focus topic, we address symmetries in probabilistic graphical models. We have a look at identifying most likely sources of events while avoiding a combinatorial explosion when domain sizes increase by using representatives for indistinguishable objects. Further, symmetries in temporal inference allow for efficient inference even for huge domain sizes by approximating symmetries over time, while propositional approaches already struggle with small domain sizes. In a second focus topic, we investigate symmetries in conditional knowledge bases and in probability distributions with maximal entropy with the overall goal of lifted inference at maximum entropy.

Speakers from the University of Münster

Braun, Tanya

Junior professorship for practical computer science - modern aspects of data processing / data science (Prof. Braun)