Thesis Defense: Hadi Salman. Title: Towards ML Models That We Can Deploy Confidently

Abstract: As machine learning (ML) systems are increasingly deployed in the
real world, concerns about their reliability and trustworthiness have become
more pronounced. This thesis aims to address these concerns via two major
thrusts: leveraging the perceived weakness of ML---adversarial
perturbations---to make ML models more trustworthy, and understanding the
underpinnings of reliable ML deployment.
More precisely, the first thrust advances a number of aspects of ML, from the
development of adversarially robust models, to creation of objects that are
easier for ML models to recognize, to safeguarding images from unwanted AI
alterations, and to improving transfer learning.
The second thrust revolves around ML model interpretability and debugging to
ensure safety, equitability, and unbiased decision-making, including diagnosing
the failure modes of such models, and identifying unexpected ways in which
data might introduce biases into them.