moteur de recherche

Groupe de travail d'apprentissage et d'optimisation - Alessandro Rudi and David Picard
Groupe de travail d'apprentissage et d'optimisation - Alessandro Rudi and David Picard
13-Oct-2020 10:00
Age: 99 days

10h-11h Alessandro Rudi (INRIA Paris),

Introduction on reproducing kernel Hilbert spaces, with application to machine learning

In this talk, I will introduce the machinery of reproducing kernel Hilbert spaces (RKHS), which is a key tool when we need to represent a rich space of functions over an arbitrary domain and have at the same time good approximation properties. A key example where RKHS are used is in supervised learning. It will be developed during the talk starting from the definition of the supervised learning problem and its solution via empirical risk minimization over suitable RKHSs.

11h15-12h15 David Picard (ENPC, LIGM)

Image Similarity: from Matching Kernels to Deep Metric Learning

Being able to define a similarity between images is a key task in computer vision as it is a necessary step to solve many popular problems such as image retrieval, automatic labeling, segmentation, etc. A historic approach inspired by stereo-analysis consists in counting the number of nearly identical regions of interest between two images. However, this approach is not compatible with machine learning tools and has to be adapted using what is known as matching kernels. In this talk, we show examples of these matching kernels and how their linearization leads to powerful representation learning techniques. We draw a parallel between these approaches and recent deep metric learning developments and we show that both are trying to solve a similar problem of distribution matching. We finally propose to solve this distribution matching problem in deep metric learning by introducing a high order moment based regularization criterion.

<- Back to: Accueil