DATA Talk : `Variational Approaches to Continual Online Learning''

Alessandro Betti (Université Côte d’Azur) -

Date: -
Location: Eurecom

Abstract: Continual lifelong learning has recently attracted much attention in the Machine Learning literature. Online Continual learning, in particular, is of particular interest for all those learning problems in which data is available as coherent stream of information which have its proper dynamics and temporal scales; one emblematic example is that of visual information. In contrast to classical statistical learning approaches which work with huge collections of data using "static" optimization techniques a crucial component of continual approaches is the "dynamical" aspects the learner should possess in order to adapt to the flow of information. To this end we argue that formulating a learning theory in terms of evolution laws instead shifts the attention to the dynamical behaviour of the learner. We discuss how variational approaches can be used to frame lifelong learning tasks and the main difficulties that this entails. Finally we show how these variational approaches could be used to extract a particular kind of visual features together with their optical flow. Bio: Alessandro Betti received the M.S. Degree in Theoretical Physics in 2016 form the University of Pisa, Italy and the PhD degree in Computer Science (Smart Computing) awarded jointly from the the Universities of Florence, Pisa, and Siena in 2020. He is currently a postdoctoral researcher at Université Côte d’Azur in the Maasai team. His main research interests are in Machine Learning and specifically in the formulation of a class of learning problems that posses a natural temporal embedding using the formalism of Calculus of Variations, with applications to Online Learning, Continuous Learning from Video Streams and Computer Vision. EURECOM Data Science Seminar