Vlad Shahuro: Training generative neural networks using Maximum Mean Discrepancy

Training generative neural networks using Maximum Mean Discrepancy There are several approaches to training generative models based on neural networks. The most popular are variational autoencoder and adversarial networks. In this talk I tell about alternative approach for training generative models. It is based on technique from statistical hypothesis testing known as maximum mean discrepancy (MMD). Such technique leads to a simple loss function that tries to match all orders of statistics between training