Project Grant 2239151

This five-year, $224,413 National Science Foundation project grant supports research and educational activities to advance the theoretical understanding of overparameterization in deep neural networks. Funded under the NSF's Computer and Information Science and Engineering program, the grant recipient will examine the implications of model overparameterization on deep network performance and robustness. Specific aims include characterizing test set regression and classification performance, as well as robustness to adversarial perturbations and shifts in the data distribution for both linear and nonlinear overparameterized models. The recipient will also design principles for robust data-driven model selection in modern machine learning. Broader goals are to establish foundational mathematical principles to explain deep learning success and failure modes, enabling efficient and principled solutions. Educational resources on signal processing, machine learning, and data science underlying the research will also be created and disseminated at the high school and undergraduate levels. The Georgia Tech Research Corporation will administer the award on behalf of the principal investigator located in Atlanta, Georgia.