Magnet seminars are usually held in room B21 on Thursdays, 11am. Check below for upcoming seminars and potential changes of schedule/location. You may also import the Magnet seminars public feed into your favorite calendar app. For further information, contact Aurélien or Michaël.
Thu, December 17, 2020
Where? Virtual (see Zoom link)
Online learning is a dominant theoretical framework present in many sectors of our world such as product retailing, autonomous driving, market survey, robotics, etc.. The main limitation of the state-of-the-art online learning strategies is data-inefficiency: the majority of online learning algorithms are data hungry. Therefore they are highly suboptimal when acting in non-stationary and high-dimensional environments, requiring a large number of interactions with the environment before learning the optimal policy. In this talk, we claim that to learn efficiently in complex domains data, one must ultimately be able to discover and exploit the structure of the high-dimensional ambient space. Graph signal processing (GSP) provides the right framework to understand and model high-dimensional data, being therefore an ideal tool for advancing policy learning frameworks in large and irregular domains. After an overall discussion on how GSP can improve online learning in general, we will present our current work on graph-based multi-arm bandit problems, where the ambient space of the problem will be represented as a graph, and the low-dimensionality of graph-based representations will be exploited to achieve data-efficient learning. We will provide first a theoretical error bound for graph Laplacian regularized estimator, and then show how this can be applied to bandit problems.
Zoom link: univ-lille-fr.zoom.us/j/97722191935
Thursday, December 17, 2020 - 11:00 to 12:00
Virtual (see Zoom link)