Defense Department to fund Brown faculty work on neural networks for AI

George Karniadakis, a professor of applied math and engineering, was one of nine faculty scientists and engineers from across the U.S. to receive a Vannevar Bush Faculty Fellowship.

PROVIDENCE, R.I. [Brown University] — The U.S. Department of Defense has selected George Karniadakis, a Brown University professor of applied math and engineering, as a 2022 Vannevar Bush Faculty Fellow.

Karniadakis is one of nine scientists and engineers from across the nation to receive the fellowship — the most prestigious research grant award from the Department of Defense, named for Vannevar Bush, who directed the U.S. Office of Scientific Research and Development after World War II. In line with Bush’s vision, the fellowship aims to advance transformative, university-based fundamental research. Each fellow receives up to $3 million over a five-year term to pursue cutting-edge fundamental research.

Karniadakis was at a scientific conference in Oslo, Norway, when he learned that he had been selected for the fellowship.

“The director of basic research at the U.S. Department of Defense, Dr. Bindu Nair, called me via Zoom and asked me if I had enough bandwidth to take on this project,” Karniadakis said. “I told her that of course, this is a major priority for me. Dr. Nair responded, ‘Well then, congratulations!’ I was very happy.”

A focus of Karniadakis’ research is machine learning for scientific computing, and his team at Brown has been developing physics-informed neural networks (PINNs) that can be used in artificial intelligence applications. He said that all mobile platforms, including robots, airplanes and self-driving cars, have sensors that accumulate data. The processing of this data by supercomputers based on the current class of artificial neurons, or artificial neural networks, is not only wildly expensive in terms of financial cost but also in terms of energy.

His new objective is to accelerate PINNs by a factor of 1,000 using spiking neural networks, which function more like human neurons and consume much less energy. Large companies like Intel and IBM have recently built computers that can be used for this new generation of neural networks.

The biologically based neural networks that we hope to develop, which have many applications in artificial intelligence, will be vastly more efficient and use less electricity than the current generation — making the process far less taxing on the environment.

George E. Karniadakis Charles Pitts Robinson and John Palmer Barstow Professor of Applied Mathematics and Engineering
 
George Karniadakis

For example, Karniadakis said, one of the biggest neural networks in the world, known as Generative Pre-trained Transformer 3 (GPT-3), consumes 20 megawatts of energy to perform a computing task that human brain does with only 20 watts of energy — which can be supplied by eating a small chocolate bar.

The next generation of biologically based neural networks, Karniadakis said, will be 1,000 times faster, and therefore consume 1,000 times less energy, than the current generation.

“The biologically based neural networks that we hope to develop, which have many applications in artificial intelligence, will be vastly more efficient and use less electricity than the current generation — making the process far less taxing on the environment,” Karniadakis said.

This will allow mobile platforms to have on-board computers that can make decisions much faster, and a huge, central computer will no longer be needed.

“One of the best parts of the Vannevar Bush fellowship is that you are free to experiment and think out of the box, and that’s extremely exciting when setting out to develop something as risky and ambitious as biologically plausible neural networks,” Karniadakis said.

The Vannevar Bush Faculty Fellowship is sponsored by the Basic Research Office within the Office of the Under Secretary of Defense for Research and Engineering. The grants are managed by the Office of Naval Research. For the 2022 competition, the department received more than 300 white papers.

The 2022 recipients join a cadre of approximately 50 current fellows who conduct basic research in areas of importance to the Department of Defense, including materials science, cognitive neuroscience, quantum information sciences and applied mathematics, among other disciplines. In addition to their research, fellows can also collaborate with Department of Defense laboratories and share insights with leadership and the broader national security community.