Robust Machine Learning: Detection, Evaluation and Adaptation Under Distribution Shift

Deep learning, despite its broad applicability, grapples with robustness chal lenges in real-world applications, especially when training and test distributions  differ. Reasons for the discrepancy between training and test distributions  include gradual changes in human behavior or differences in the demographics  of the environment where the service is being used. While obtaining labeled  data for anticipated distribution shifts can be daunting, unlabeled samples are  relatively cheap and abundantly available My research leverages unlabeled data from the target domain to identify  structural relationships between the target and source domains, and then use  them to adapt and evaluate models. The work discussed in thesis involves  understanding the behavior of deep models, both theoretically and empirically,  and using those insights to develop robust methods. In particular, this thesis  surveys my work on the following three questions:  

Q1: How to adapt models in the face of distribution shifts? Absent assump tions on the nature of the distribution shifts, this task is impossible. My research  in this direction is focused on formulating assumptions on distribution shift  scenarios appearing in the wild and developing procedures that improve and  adapt deep models under those shifts by leveraging unlabeled data. Part I and  II of this thesis delve into this research.  

Q2: How can we evaluate models’ performance without access to labeled data?  Deep learning models fail silently, i.e., they cannot flag uncertain decisions. To  build reliable machine learning systems, obtaining certificates for accuracy is as  important as robustifying these systems. Part III discusses my research in this  direction and presents techniques that leverage unlabeled data to predict model  accuracy.  

Q3: How can we leverage foundation models to address distribution shift  challenges? Foundation models, such as vision language models, have demon strated impressive performance on a wide range of tasks. However, these models  also lack robustness due to spurious correlations, poor image-text alignment,  etc. Moreover, they also get outdated as the internet data evolves presenting  novel challenges in keeping them up-to-date. Part IV of my thesis discusses  my work on understanding the behavior of foundation models and developing  techniques to improve their robustness under distribution shifts.  

Overall, this thesis expands the frontier of robust machine learning by  developing techniques that leverage unlabeled data to adapt and evaluate  models under distribution shifts. The work presented here is a step towards  developing a comprehensive toolkit for robust machine learning in the face of  distribution shifts.