On Generative Adversarial Networks and Density Estimation - Fernando Perez-Cruz скачать в хорошем качестве

On Generative Adversarial Networks and Density Estimation - Fernando Perez-Cruz

DALI 2017 Workshop - Theory of Generative Adversarial Networks http://dalimeeting.org/dali2017/gener... Title: On Generative Adversarial Networks and Density Estimation Speaker: Fernando Perez-Cruz (Stevens Institute of Technology), http://www.tsc.uc3m.es/~fernando/ Abstract: Generative adversarial networks (GANs) propose an alternative approach for density estimation that relies on a discriminative network's output to lead a generativenetwork's density model. They seem to be able to tackle density estimation in high dimensional spaces with limited number of training samples. In this paper, we analyze the accuracy of GANs in estimating the underlying density of the training samples. We focus on both the asymptotically analysis and an empirical study with a state-of-the-art GAN implementation. For the former, we show a limitation in the proof of convergence in (Goodfellow et al., 2014) that might prevent the GAN density estimation from converging to the true density as the number of samples increases. For the latter, we advocate for a two-sample test as the proper test for measuring GANs performance, since it is a simple test would tell us if two set of samples come from the same distribution.