This $160,497 project grant awarded by the National Science Foundation (NSF) under the Computer and Information Science and Engineering (CFDA 47.070) program will fund research by the University of Michigan to analyze the computational landscape of nonconvex machine learning models and develop reliable, efficient algorithms for training these models. The project aims to demonstrate that the local solutions of many practical nonconvex models can be more tractable and generalize better than their global optima. Key activities include conducting a systematic analysis of the optimization landscape around true solutions and designing algorithms to solve nonconvex ML problems at meaningful scales. The project will integrate educational programs for K-12, undergraduate, and graduate students, with a focus on outreach to underserved communities. Funds will support this 5-year research and education initiative at the University of Michigan, with the goal of transmuting nonconvexity from a curse to a blessing in machine learning.
Generated 8/13/24, 5:34 AM