Séminaire de R. Willett

Rebecca WILLETT  (Université du Wisconsin-Madison) invitée par Joseph Salmon, donnera un séminaire le Mercredi 18 Juin à 13h30, salle DB003 (site Dareau)

Titre : Efficient tracking in dynamic environments

Résumé : Modern sensors are collecting data at unprecedented rates, often from platforms with limited processing power and bandwidth for data transmission. To cope with this data deluge, we must develop robust methods for efficiently extracting information from large-scale streaming data. This task is most tractable when the information of  interest exhibits low-dimensional spatio-temporal structure. However,  in practical scenarios ranging from motion imagery to network analysis, the environment is nonstationary, resulting in both poor empirical performance and weak theoretical guarantees. I will describe a novel “dynamic mirror descent” method which addresses this challenge by learning and tracking low-dimensional models underlying the data.The associated regret bounds, in contrast to competing online learning regret bounds, reflect this underlying spatio-temporal structure. These concepts are demonstrated empirically in the context of sequential compressive observations of a dynamic scene and tracking  influences within a dynamic network.

Comments are closed.