Adversarial Flows
Title: Adversarial Flows
SNIC Project: Berzelius-2021-70
Project Type: LiU Berzelius
Principal Investigator: Emir Konuk <ekonuk@kth.se>
Affiliation: Kungliga Tekniska högskolan
Duration: 2021-12-02 – 2022-02-01
Classification: 10207
Keywords:

Abstract

The goal of the project is to develop a novel generative model for the natural images domain. To this end, we train a generative adversarial network (GAN) with a normalizing flow (NF) network as the discriminator. The main benefit of the approach would be to have a model that has the flexibility of a GAN combined with the exact likelihood estimation capability of a NF. We hypothesize that adversarial training will improve the likelihood estimation of the NF, yielding e.g. better outlier detection or robustness vs. adversarial (generated or real) samples. We also hypothesize that the invertibility property of the flow model would help stabilize GAN training since a vanilla deep network tends to overfit when employed as a discriminator.