Packard fellow McMahon rethinks neural-network computing

Artificial intelligence and machine learning have become crucial tools for medical imaging, autonomous vehicles and many other applications, but engineers warn that the exponential increase in computer power needed to run such neural network software is unsustainable without dramatically new types of hardware.

Using a Packard Fellowship for Science and Engineering from the David and Lucile Packard Foundation, Peter McMahon, assistant professor of applied and engineering physics in the College of Engineering, aims to harness the power of photonics to build processors for neural networks that are over 1,000 times more energy efficient than today’s most advanced digital processors. These new processors will enable larger networks that will transform the capabilities of artificial intelligence.

Peter McMahon

McMahon is among 20 early-career researchers in the 2021 class of Packard fellows, announced Oct. 14. The fellowship comes with $875,000 over five years to explore research with almost no restrictions.

“There is a groundswell of renewed interest in unconventional approaches to computing and it’s a very exciting time to be in the field,” said McMahon, noting the fellowship will help him explore concepts in photonic computing that are more speculative than what standard funding sources would typically risk supporting.

The interest in exploiting photonics – the science and technological application of photons, which are particles of light – for neural networks comes at a time when ever-larger numbers of transistors are crammed onto computer chips, consuming more and more power. As a result, datacenters are growing at a rate that is neither economically nor environmentally sustainable.

“The largest planned datacenters will consume 1 gigawatt of power, more than the entire city of San Francisco,” McMahon said, “so they are already nearing the edge of what is feasible.”

Compared to digital electronics, analog photonics has the potential for processors to harness far larger bandwidths, with greater energy efficiency. There are a number of scientific challenges that need to be overcome for photonic processing to become competitive with electronic processors, but McMahon previously demonstrated the use of photonics in an unconventional processor that solved optimization problems with the same performance as state-of-the-art electronic processors.

With the Packard fellowship, McMahon will design, prototype and study photonic processors for neural networks. His lab will engineer free-space optical apparatus and photonic chips that take advantage of both linear and nonlinear optics to realize low-power neural networks useful for a variety of tasks, including image recognition, object tracking and optical sensing.

McMahon described receiving the fellowship as a “team effort,” and acknowledged Cornell colleagues who have influenced his research direction, including postdoctoral researchers Tatsuhiro Onodera, Tianyu Wang and Logan Wright. McMahon also credited Iwijn De Vlaminck, associate professor of biomedical engineering, and several previous Packard fellowship awardees at Cornell for providing feedback on his proposal.

Recent Cornell winners include Kirstin Petersen, assistant professor of electrical and computer engineering (2019); Ilana Brito, assistant professor of biomedical engineering (2017); and Lena Kourkoutis, associate professor of applied and engineering physics (2014).

Syl Kacapyr is public relations and content manager for the College of Engineering.