Biography

Olivier Sigaud is professor in Computer Science at University Pierre et Marie Curie, Paris VI. He is centrally interested in the use of machine learning for animal decision making, human motor control and developmental robotics. He is or has been involved in the MACSi (http://macsi.isir.upmc.fr/), CODYCO (https://www.codyco.eu/) and DREAM (www.robotsthatdream.eu/) projects.

He also holds a PhD in philosophy since 2004. For more information, see http://people.isir.upmc.fr/sigaud.

Abstract

Towards developmental discovery of objects in dynamical scenes

To act efficiently in an open environment, a robot needs to understand the surrounding world in terms of objects, properties, affordances, etc. A long-standing goal of developmental robotics consists in endowing a robot with the capability to learn the corresponding representations. So far, a good deal of the Intrinsically Motivated Open-ended Learning literature has been focused on designing architectures encompassing several developmental learning components such as goal exploration processes, multi-task learning, curriculum learning based on learning progress, reinforcement learning, etc. In parallel, deep learning tools such as CNNs or autoencoders have led to unprecedented performance in object recognition from static images. In my talk, I will ask how we can make profit of these deep learning tools to let a developmental agent build its own representations of objects, its own object-oriented goals, etc. I will explain why combining both components is not straightforward and I will propose a general framework for the discovery of objects in dynamical scenes that may help making this integration easier.