MIT Creates Neural Network Processor Based on Light

A substantial and growing segment of the world’s computing power is being dedicated to machine learning, but not all processing hardware is up to the task. Some AI models are so large they require entire server farms equipped with powerful chips that guzzle energy. Scientists at MIT may have overcome some key limitations with an alternative type of computing hardware, known as photonics. If the early results are borne out, this architecture could vastly accelerate neural networks.

Traditional electronic computers have gotten faster every year since their invention, but there are hints that progress will slow as transistors reach the limits of atomic scale. Researchers have continuously fiddled with optical computing because photons can theoretically provide much higher effective bandwidth, but implementing this at the chip level is challenging.

In 2017, MIT researchers succeeded in creating a chip that can run neural network calculations, but there were some notable limitations. Most of the heavy lifting in machine learning comes in the form of matrix multiplication, in which linear algebra is used to transform data as it passes from one node in the network to the next. However, complex learning functions require nonlinear operations, which could not be run with photonics. Instead, the team translated those functions into electronic signals and processed them on external hardware. While that worked, it squandered the advantages of photonic processing.

In the new work, led by visiting scientist Saumil Bandyopadhyay, the team devised a new component called new nonlinear optical function units (NOFUs). This component combines photonics with traditional computing to keep all neural network processing on the chip. At the beginning of each calculation, MIT’s experimental system encodes data into light using an array of programmable beamsplitters, just like the 2017 version. NOFUs implement nonlinear functions as needed by siphoning a small amount of light to photodiodes that convert it into an electrical signal. This processing happen as needed on the chip, eliminating the need for an external amplifier.

“We stay in the optical domain the whole time, until the end when we want to read out the answer. This enables us to achieve ultra-low latency,” said Bandyopadhyay.

This approach to low-latency photonics allowed the team to build and run neural networks with surprising efficiency. The network achieved more than 96% accuracy in training, which is close to traditional computing. When running machine learning workloads, the photonic system achieved 92% accuracy. Again, this is close to traditional hardware, but the photonic system runs key calculations in about half a nanosecond, which is considerably faster than electronic computers.

The team notes the photonic chip was fabricated using standard tools found in semiconductor foundries today. This could allow photonics to be manufactured at scale with very little change to existing infrastructure. However, transistors are cheap and well-understood. Photonics, as potentially revolutionary as they might be, may have to wait for advances in electronic computing to peter out.