Fall 2021 GRASP SFI: Alex Robey, University of Pennsylvania, “Model-Based Robust Deep Learning”

In the last decade, deep learning has resulted in major breakthroughs in many domains.  However, the frameworks commonly used in deep learning remain fragile to a variety of innocuous transformations of data.  And while some progress has been toward improving the robustness of deep learning, the fragility of these methods represents one of the most significant barriers to their deployment in safety-critical applications.  Toward addressing these shortcomings, in this talk I will discuss the theory and practice of robust deep learning.  In particular, I will describe a unifying paradigm for training deep neural networks that are robust to broad classes of transformations of data, resulting in state-of-the-art robust performance in a variety of fields, including adversarial robustness, domain generalization, and robust certification.  I will also highlight myriad directions for future work toward improving the robustness of deep learning methods.