Using a Multi-Dimensional Scaling Aucarithm to Draw a Time Series

In a wasersten gan, you have basically a loss that's used to train the generator. That loss comes from an underlying notion of a distance that needs to be selected. And so all we do is end up using g t w as this notion of distance, together with an antripe regularization term. It sort of just smooths everything out a little bit. Once you have such a notion and you know how to differentiate through it, you can use it to build a generative model. But a typical application area for methods like this is actually imitation learning. So in imitation learning, there is some kind of expert data collection system. The goal of the system is to