Abstract

Robust estimation under Huber's $\epsilon$-contamination model has become animportant topic in statistics and theoretical computer science. Rate-optimalprocedures such as Tukey's median and other estimators based on statisticaldepth functions are impractical because of their computational intractability.In this paper, we establish an intriguing connection between f-GANs and variousdepth functions through the lens of f-Learning. Similar to the derivation off-GAN, we show that these depth functions that lead to rate-optimal robustestimators can all be viewed as variational lower bounds of the total variationdistance in the framework of f-Learning. This connection opens the door ofcomputing robust estimators using tools developed for training GANs. Inparticular, we show that a JS-GAN that uses a neural network discriminator withat least one hidden layer is able to achieve the minimax rate of robust meanestimation under Huber's $\epsilon$-contamination model. Interestingly, thehidden layers for the neural net structure in the discriminator class is shownto be necessary for robust estimation.

 

Introduction (beta)

None

 

Conclusion (beta)

None