NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:3262
Title:Learning from Bad Data via Generation

The authors propose a new training methodology where the model is learned to minimize the expected loss with respect to the worst case distribution within an epsilon ball of the empirical distribution. The adversarial formulation appears novel and the paper provides interesting theoretical results. The empirical evaluation of the proposed methods is well done and convincing. Hence, we recommend acceptance.