Kernels, tensors and structured data.

The aim of this project is to develop and analyze new machine learning algorithms for structured data by combining elements from tensor-based methods, kernels and sparsity. Objectives involve, in particular: the development of a systematic kernel-based framework for the analysis of data that can be represented by means of tensors. This includes: the development of new tensorial kernels for structured data in particular dynamical systems and graph. The in-depth analysis of subspace-based kernels and, in particular, the study of the induced similarity measure. B) Existing tensor methods are mostly unsupervised techniques. A second objective relies on convex optimization and aims at combining tensors and kernels on a broad range of learning tasks. This includes in particular the extension of nuclear norm regularization from affine tensor-based models to non-parametric models based on product kernels (such as the Gaussian RBF kernel). The use of novel regularization schemes for affine tensor-based models beyond the recently proposed approach that uses nuclear norms. The application on a number of actively studied real-life problems.

Keywords:Kernel methodes, Dynamische systemen, Tensoren, Ijlheid inducerende schatters, Kernel methods, Dynamical systems, Tensors, Sparsity inducing estimators

Disciplines:Modelling, Multimedia processing