Proceedings of the Annual Meeting of the Cognitive Science Society

This paper describes the principles of a research programme for cognitive science that exploits recent developments in machine learning (ML) to generate novel hypotheses about the structure of human cognition. Current debate over the interpretability and explainability algorithms usually focuses on the properties of the algorithms themselves in virtue of which they are either interpretable or explainable. However, we argue that there is value in conceptualizing these categories as inherently psychological constructs. Given certain mathematical features of machine learning algorithms – specifically, that many useful ML algorithms are members of Rashomon sets – it is possible to exploit their utility to reason using a principle of parsimony about the inferential structure of certain human cognitive tasks. Algorithms that do something the human mind can do, and are both interpretable and explicable could be, we shall argue, inferential homologues of certain core cognitive processes. We illustrate this proposal with an example drawn from clustering models used in exploratory data analysis, and then conclude with a discussion of some of the philosophical limitations of our proposal.