12h30-13h45 : déjeuner buffet en salle 2B140
13h45-14h15 : informations diverses sur le laboratoire et le suivi de thèse (Olivier Guédon, Sakina Kawami, Claire Lacour)
14h15-14h45 : Aurélie Bigot CLT in some Banach spaces
(Summary : The central limit theorem is one of, if not the, central theorems in statistics. In this talk, we look at the assumptions of the theorem that we seek to weaken. To this aim, we extend the central limit theorem under the Dedecker-Rio condition to adapted stationary and ergodic sequences of random variables taking values in a class of smooth Banach spaces.)
14h45-15h15 : pause café
15h15-15h45 : Valentin Lemarié Heat equation and Besov spaces
(Summary : The heat equation, first introduced by Joseph Fourier in 1807, serves as a reference for the study of parabolic equations. In the case of the whole space R^d, the existence and uniqueness of solutions in the space of temperate distributions is easily obtained by using the Fourier transform. Studying this equation in Sobolev spaces does not give all the information we might expect: the gain in derivatives with respect to the initial condition and the external source is not maximal.
To achieve this optimality, the appropriate Banach spaces are Besov spaces: they were partly created and used to maximize information from a linear point of view on the heat equation. For this talk, I will illustrate this point in an L² setting, where the use of Plancherel's theorem greatly simplifies the computation of a priori estimates.)
15h45-16h15 : Florian Valade Optimizing Deep Learning Networks: An Introduction to Early Exits
(Summary: Deep learning networks are at the core of numerous technological advancements, but their complexity and energy demands pose significant challenges. This introductory talk will begin with a concise overview of the fundamental principles and functioning of deep learning networks. We will then explore the concept of "early exits," a promising technique that reduces computational costs by enabling early decision-making in deep networks. The focus will be on the potential of this approach to make models more efficient while maintaining their performance.)
