Rajesh Ranganath

Interpretability and Out of Distribution Generalization in Deep Predictive Models

Speaker: Rajesh Ranganath, NYU

Abstract:

Interpretability enriches what can be gleaned from a good predictive model. Techniques that learn-to-explain have arisen because they require only a single evaluation of a model to provide an interpretation. In the first part of this talk, I will discuss a flaw with several methods that learn-to-explain: the optimal explainer makes the prediction rather than highlighting the inputs that are useful for prediction. I will also describe an evaluation technique that can detect when the explainer makes the prediction along with a new method that learns-to-explain without this issue. In the second part of my talk, I will discuss our work on representation learning for out of distribution generalization. I will construct a family of representations that generalize when under changing nuisance-induced spurious correlations and have applications to images and chest X-rays. I will show how nuisance variables can be constructed using limited prior knowledge and augmentations of the input.

Bio: Rajesh Ranganath is an assistant professor at NYU’s Courant Institute of Mathematical Sciences and the Center for Data Science. He is also affiliate faculty at the Department of Population Health at NYUMC. His research focuses on approximate inference, causal inference, probabilistic models, and machine learning for healthcare. Rajesh completed his PhD at Princeton and BS and MS from Stanford University. Rajesh has won several awards including the NDSEG graduate fellowship, the Porter Ogden Jacobus Fellowship, given to the top four doctoral students at Princeton University, and the Savage Award in Theory and Methods.

This meeting was not recorded.