Robustness to shifts in input distributions is crucial for reliable deployment of deep neural networks. Unfortunately, neural nets are extremely sensitive to distributional shifts, making them undesirable in safety-critical applications. For instance, perception system of a self-driving car trained on sunny weather conditions fails to perform well on snow. In this talk, I will present several algorithms for robust learning of deep neural networks against input distributional shifts.

First, I will present some results on likelihood computation using generative models, and how these likelihood estimates can be used for quantifying distributional shifts. Then, I will discuss robust learning algorithms for two broad classes of distributional shifts - naturally occuring covariate shifts, and artificially constructed adversarial shifts. For adapting to covariate shifts, I will present techniques using Generative Adversarial Networks (GANs) and regularization strategies. For adversarial shifts, I will discuss why current robust training algorithms have poor generalizing effect, and propose a technique for improving generalization.


Examining Committee: 

 

                          Chair:               Dr. Rama Chellappa
                          Dept rep:         Dr.  Soheil Feizi
                          Members:        Dr. Abhinav Shrivistava