Skip to main content

How to Coach your Cognitive Assistant

Loizos Michael ( Open University of Cyprus )

During the dialectical process of human-human communication, a common assumption of a shared knowledge representation and reasoning structure would seem to underlie the efficient transfer of knowledge between interlocutors. We posit that a paradigm of “machine coaching” that explicates this shared structure would analogously facilitate human-machine interaction, especially in the context of designing personalized cognitive assistants for certain tasks. During a machine coaching session, a human communicates with a machine that seeks to learn and apply the human’s policy for a task, with the human asking the machine to explain its actions and inferences, while offering to the machine feedback to help it improve its approximation of the human’s target policy. We present a general framework for the design of machine coaching protocols that utilize formal argumentation for communication, and demonstrate a concrete protocol for which we can establish learning-theoretic guarantees on its efficacy and efficiency.

Speaker bio

Loizos Michael is an Assistant Professor at Open University of Cyprus, where he founded and directs the Computational Cognition Lab. He was educated at University of Cyprus, where he received a B.Sc. in Computer Science with a minor degree in Mathematics. He continued his education at Harvard University, where he received an M.Sc. and a Ph.D. in Computer Science.

His research focuses on the principled study of cognitive processes associated with individual or collective intelligence — such as learning, reasoning, sensing, communication, cooperation — and how those are used by humans and other organisms in everyday life. Emphasis is placed on the development of computational models for various aspects of cognitive processes, and the analysis of the formal implications that such models have. This computational view of cognition is complemented by simulations, real‐world experiments, and psychological studies, designed to validate the proposed models and to identify features thereof that warrant further study.

Share this: