Hugo Caselles-Dupré
Date:
Speaker
Hugo Caselles-Dupré is a Post-Doc at ISIR (Sorbonne University) under the supervision of Olivier Sigaud and Mohamed Chetouani, studying the teachability of autonomous agents. In 2021, he graduated with a PhD from the Flowers Laboratory of ENSTA ParisTech and INRIA (supervised by David Filliat and Michaël Garcia-Ortiz). His research focuses on Machine Learning for Artificial Agents. He studies how to artificial agents in situated environments can create their own perception and learn using a social partner.
Abstract
Learning from demonstration methods usually leverage close to optimal demonstrations to accelerate training. By contrast, when demonstrating a task, human teachers deviate from optimal demonstrations and pedagogically modify their behavior by giving demonstrations that best disambiguate the goal they want to demonstrate. Analogously, human learners excel at pragmatically inferring the intent of the teacher, facilitating communication between the two agents. These mechanisms are critical in the few demonstrations regime, where inferring the goal is more difficult. In our work, we implement pedagogy and pragmatism mechanisms by leveraging a Bayesian model of Goal Inference from demonstrations inspired by works in cognitive sciences and developmental psychology by M. Ho and H. Gweon. We also show how these mechanisms can help avoid misunderstanding the teacher’s intentions in the case of language instructions, by solving referential ambiguities that are inherent to the use of language.