Toward Learning Distributions of Distributions

Moritz Wohlstein, Ulf Brefeld

Keywords: MMD, distribution embedding, hypernetwork, kernel embedding, GMMN

Abstract: We propose a novel generative deep learning architecture based on generative moment matching networks. The objective of our model is to learn a distribution over distributions and generate new sample distributions following the (possibly complex) distribution of training data. We derive a custom loss function for our model based on the maximum mean discrepancy test. Our model is evaluated on different datasets where we investigate the influence of hyperparameters on performance.

Submission Number: 39