Researcher at Naver Labs Europe

Description

Diane Larlus is a researcher in computer vision at Naver Labs Europe. Her research interests include global and semantic understanding of visual scenes and the learning of generic visual representations. She is the scientific head of a chair on the continuous learning of representations at the Grenoble institute of IA MIAI.

Presentation title

Learning generic and transferable visual representations with weak supervision 

Abstract

Computer vision is at the heart of a growing number of applications, thanks in particular to the recent success of so-called deep learning methods. One of the reasons for this success is related to the development of large, deep architectures capable of producing generic visual representations that can be applied directly - or easily transferred - to a wide variety of tasks. However, learning the parameters of such architectures requires large-scale image collections, rich manual annotations, and high computing power.
In order to simplify the learning of transferable visual descriptors, some recent approaches have looked for ways to use fewer annotations, make do with noisy annotations, or even do without annotations altogether. This presentation covers two of our recent contributions to this research direction. They are part of the MIAI Chair on the continuous learning of representations.