New theoretical and algorithmic advances in machine learning with optimal transport
16th November 2017; 14:00
Abstract: Optimal transportation problem is a powerful technique that has recently found its application in various areas of machine learning including, for instance, computer vision, information retrieval and music unmixing. In this talk, I will introduce the basic concepts of optimal transportation theory as well as some algorithmic ideas that were proposed in machine learning based on it. These algorithmic ideas include a recently proposed unsupervised learning algorithm based on regularized optimal transport and a new method for feature selection applied in the context of domain adaptation for prostate cancer mapping. From that point I will further present some theoretical insights that were obtained based on the optimal transportation theory for domain adaptation in order to justify its use in this context and to analyze the a priori success of adaptation represented by the existence of the joint hypothesis between source and target domains.