Beatrice Acciaio (London School of Economics)

In this talk I will present a new method to train generative models, based on non-anticipative optimal transport in conjunction with Sinkhorn divergences. The generator is trained to produce real-looking evolutions of processes, given long time-series of data. To improve its training, a discriminator is set against it, with the task of evaluating the distance between real and fake data. We employ recurrent and convolutional neural network architectures to account for the dynamic nature of the problem. A conditional modification of our model leads to prediction of sequential data. Several applications will be discussed.