This $1.6 million Project Grant from the National Science Foundation's Computer and Information Science and Engineering program will fund research at the University of Pennsylvania from October 2022 through September 2026. The research focuses on developing theoretical tools to build an understanding of why deep neural networks (DNNs) work and when they can fail. Investigators will seek to identify common themes in how artificial and biological systems like the human brain learn. They will also examine the hypothesis that DNNs succeed when learning tasks exhibit a "sloppy" structure characterized by a Fisher information matrix with exponentially distributed eigenvalues. If validated, this could help explain DNNs' abilities to generalize well despite large model sizes. Additional goals include curriculum development at the intersection of computer science, neuroscience and mathematics, as well as AI education outreach for high school students and mentoring for young Latinx researchers.
Generated 1/7/24, 10:13 AM