AIAI Seminar - 18/10/21 - Siddharth Narayanaswamy

Title: A Partial View of xAI

Abstract:

An interesting challenge faced by data-driven approaches of recent years relates to their 'explainability'. While they often match (and sometimes exceed) human performance, they are typically opaque to human understanding of rationale, i.e. how they reached their decisions.

Standard approaches to xAI primarily tackle rationale by attempting to 'translate' the black box into an explainable form---broadly, through surrogate models or sensitivity analyses---which are often only able to provide simplistic explanations (heat maps, categorical labels), and can lack context (mainly targetting expert users). In some cases, the translation itself can involve black boxes, which further exacerbates issues.

Building on some of my prior work on human-like learning and reasoning with abstract symbolic representations of data, I have been exploring an alternate approach to xAI that seeks to establish rationale by 'opening up' the black box---through learning interpretable structured decompositions of data. The key insight is that interpretable representations can help establish common ground with humans by leveraging broad (not task specific) constraints from cognitive science, and enable more human-understandable symbolic reasoning and processing of decisions. In this talk, I will outline the benefits of such an approach, describe some recent approaches towards achieving this goal, and set out a plan for further research exploring structured representations of data.

Bio:

Siddharth N. (Sid) is a Reader in Explainable AI in the School of Informatics at the University of Edinburgh. Prior to this, he was a Senior Researcher in Engineering at the University of Oxford and a Postdoctoral Scholar in Psychology at Stanford. He obtained his PhD from Purdue University in Electrical and Computer Engineering. His research broadly involves the confluence of machine learning, computer vision, natural-language processing, cognitive science, robotics, and elements of neuroscience, leading towards a central research goal to better understand perception and cognition with a view to enabling human-intelligible machine intelligence