Compression vs. Accuracy: Compact Models for Efficiency and Interpretation

Basic data for this talk

Information about the event

Abstract

Our surrounding world is inherently uncertain and relational. The field of Statistical Relational AI (StaRAI) has emerged to account for both. StaRAI explicitly encodes objects and relations in probabilistic models, which enables algorithms to exploit repeated structures, i.e., isomorphic subgraphs with matching associated probability functions, for efficiency gains during inference. While such repeated structures frequently occur in many practical applications, they are generally not explicitly represented in a learned model and thus cannot be exploited. It is therefore crucial to efficiently identify and compress these structures. Next to a significant reduction in storage requirements, dedicated inference algorithms can use these compressed structures for efficiency gains, yielding tractability in the number of random variables. This tutorial provides a look at recent advances in the task of computing a highly compressed model from a given propositional model. We consider how the compression can efficiently be realised. Furthermore, we discuss the approximation of a compressed representation, give error bounds for the induced approximation error as well as investigate how to obtain a compressed model for a given error bound. Additionally, for the trade-off between accuracy and compression, we also discuss a hierarchical approach and provide different points of view on the clustering in preparation for better interpretability.

Speakers from the University of Münster

Braun, Tanya	Junior professorship of practical computer science - modern aspects of data processing / data science (Prof. Braun)
Speller, Jan	Junior professorship of practical computer science - modern aspects of data processing / data science (Prof. Braun)