Project Grant N000142112589

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...

The Trustees of the University of Pennsylvania received a $275,000 project grant from the National Science Foundation's Mathematical and Physical Sciences program (CFDA 47.049) to conduct research on geometrization approaches toward understanding deep learning. Specifically, the three-year award funds research projects analyzing symmetries in trained deep neural networks, examining dynamics of deep learning training, and investigating how deep learning separates data across neural network...

This $298,450 National Science Foundation project grant supports research to quantify the error landscape of deep neural networks. Funded under the Computer and Information Science and Engineering program (CFDA 47.070), the awardee New York University will employ statistical mechanics methods to characterize the basins of attraction in high-dimensional parameter spaces of deep learning models. The university will measure basin volume distributions and flatness as a function of network parameters...

This four-year $800,000 Project Grant from the National Science Foundation's (NSF) Computer and Information Science and Engineering program (CFDA 47.070) will support research at the University of California, Los Angeles (UCLA) to develop theoretical tools for understanding deep neural networks (DNNs). The research aims to identify common themes in how artificial and biological systems like the human brain learn. It will investigate the hypothesis that DNNs succeed when learning tasks exhibit...

This four-year, $622,992 project grant from the National Science Foundation's Computer and Information Science and Engineering program aims to develop methods for making machine learning models more interpretable and reliable. Specifically, researchers at the University of Virginia will investigate the mathematical foundations of deep neural networks, with a focus on geometry and topology, to better understand internal representations. Computational tools will be designed based on these...

The National Science Foundation Division of Computing and Communication Foundations awarded a $500,000 Project Grant to Princeton University from January 1, 2022 to December 31, 2024 to support collaborative research on probabilistic, geometric, and topological analysis of neural networks. This award falls under the Mathematical and Physical Sciences program (CFDA 47.049), which aims to strengthen the Nation's scientific enterprise through advancing knowledge and understanding of major problems....

This three-year, $600,000 project grant from the National Science Foundation's Computer and Information Science and Engineering program aims to advance understanding of convolutional neural networks. Specifically, the University of Maryland, College Park will investigate how multiscale representations and shift invariance affect the dynamics of neural network training. Researchers will analyze kernel properties of overparameterized networks to illuminate relationships between architecture and...

The National Science Foundation (NSF) awarded a $225,000 Project Grant under the Mathematical and Physical Sciences (CFDA 47.049) program to The Leland Stanford Junior University. The 3-year grant, effective July 1, 2024, aims to gain a deeper theoretical understanding of the statistical properties of neural networks, which have revolutionized science and engineering. Key research directions include studying the distinguishing features of deep neural networks compared to classical statistical...

This two-year, $250,000 Project Grant from the National Science Foundation's Computer and Information Science and Engineering program aims to develop techniques for improving the interpretability and robustness of deep neural networks. Specifically, the University of California, Santa Barbara will apply ideas from communication theory and neuroscience to actively shape the features extracted by individual layers of neural networks in addition to end-to-end training. By learning "matched...

This three-year Project Grant from the National Science Foundation's Division of Computing and Communication Foundations will fund the development of analysis and design techniques to provide formally guaranteed error bounds for deep neural networks used in autonomous cyber-physical systems. Totaling $999,996, the award will support research from June 2022 to May 2025. Specifically, the University of California, Los Angeles will investigate two approaches - "correctness-by-training"...