131. PDG Talk: Wasserstein Distance as a Loss Function
Dezember 12 @ 7:00 pm
!! Erst um 19 Uhr, da wir vorher keinen Raum haben !!
Track: Alternative Loss Functions
Talk: Wasserstein Distance as a Loss Function
Abstract: I’ll discuss the role of loss functions within deep learning using the Cross Entropy Loss to provide the right math framework. Then I’ll introduce the Wasserstein Distance, its properties and its limitations. To be concise, the wasserstein distance is a metric based on the optimal transport theory that allows us to compare two distributions within an embedded distance metric. We can then customize the embedded distance metric to provide a specific evaluation that better matches our problem. On the other hand, the Wasserstein distance is terribly expensive to calculate, and its mathematical properties are not trivial, given to many problems in its understanding and application (i.e., Wasserstein GANs). We will discuss two seminal papers: „Sinkhorn Distances“, by Cuturi, showed us how to calculate the Wasserstein distance fast enough for practical purposes. „Learning with a Wasserstein Loss“, by Frogner, was the first to use Cuturi’s work to train a machine learning.Then we will go through many applications of the Wasserstein loss in several fields: semantic classification, GANs, and semantic embeddings.
Wir treffen uns im Informatik-Gebäude des KIT (50.34), Raum -120.