One explanation does not fit all: an overview of AI explainability techniques

Speaker: Ronny Luss
Affiliation: IBM Thomas J. Watson Research Center

Abstract

As artificial intelligence and machine learning algorithms make further inroads into society, calls are increasing for these algorithms to explain their decisions. Whether you are the loan officer that needs to understand why an algorithm accepted a particular application or the applicant that wants to know what they could've done different to avoid their rejected application, the need for explanations is clear. This talk gives an overview to AI Explainability with a focus on some recent tools developed at IBM including attribution methods and contrastive explanations. I will discuss the application of such methods to different domains including image classification, text classification, and more recently to text generation, i.e., large language models. We further provide a taxonomy to help those requiring explanations figure out which explanation method will best serve their purposes.

Biography

Ronny Luss is a Research Scientist in Trustworthy AI at IBM Research where he has worked on projects across a multitude of industries and applications including product recommendations, advertising, insurance, and explainability, among others. Ronny has published articles in various machine learning and optimization journals and conferences, and holds a Ph.D. in Operations Research from Princeton University.