The University of California, Berkeley has announced grants totaling $10 million from the Simons Foundation and National Science Foundation in support of a five-year effort aimed at developing a better theoretical understanding of deep learning, an AI approach for teaching computers to learn from data.
Although the approach — part of a broader family of machine learning methods based on artificial neural networks that digest large amounts of raw data inputs and train artificial intelligence systems with limited human supervision — is widely used, its theoretical underpinnings are poorly understood. The Division of Mathematics and Physical Sciences at the Simons Foundation and NSF are partnering to address that gap by funding research that advances understanding of the mathematical and scientific foundations of deep learning and enables researchers to address its limitations, including its vulnerability to data manipulation.
In addition to UC Berkeley, institutions participating in the effort include Stanford University, the Massachusetts Institute of Technology, UC Irvine, UC San Diego, the Toyota Technological Institute at Chicago, the Hebrew University of Jerusalem, and EPFL in Lausanne, Switzerland.
The Data Science Education Program in the Division of Computing, Data Science, and Society at Berkeley seeks to give students in the data sciences the tools needed to consider data confidently, critically, and ethically in whatever career they choose to pursue. To help advance that mission, the newly funded project will organize workshops open to the public, train postdocs in the theoretical basis of the approach, and sponsor an annual summer school for graduate students, postdocs, and faculty.
"Our excitement over receiving this award is that we will be digging into the theoretical foundations of deep learning," said Peter Bartlett, associate director of the Simons Institute for the Theory of Computing, who will lead the project. "The recent success in machine learning has been driven by a spirit of craftsmanship by people who find ways to make this technology successful. But much of this work contradicts a lot of our classical understanding of statistical methodology, and there are many things we don’t understand about how and why these systems work."