Wasserstein distance and proxies, application and statistics - Guillaume Mahey скачать в хорошем качестве

Wasserstein distance and proxies, application and statistics - Guillaume Mahey

Wasserstein distance and proxies, application and statistics Abstract: The Wasserstein distance (WD) has been proven useful in many application of data science. However, as we will see, the WD suffers from two problems: its computational cost and the behaviour of its empirical approximation. In order to fixed this two problems a trend of work consists at using proxies of the WD, namely quantities that behave similarly but that have favorable computation and Statistics properties. We will introduce the most famous proxy called Sinkhorn divergence and recalled some of its strengths. Finally, we will introduce a new proxy we proposed, study its topological and statistical properties.