EPFL tackles power bottleneck of optics-based AI

08 Aug 2024

Scattered light from low-power laser performs computations using less energy.

A project at Swiss research center EPFL has developed a novel framework for optical neural networks, one that exploits multiple scattering from a low-power laser to carry out its functions.

Published in Nature Photonics, the research "overcomes a key computational bottleneck of optics-based artificial intelligence systems," according to EPFL.

The energy requirements of digital AI systems and the carbon emissions associated with them have become urgent topics as AI becomes more widespread.

Deep neural networks, inspired by the architecture of the human brain, are especially power-hungry due to the millions or billions of connections between multiple layers of neuron-like processors.

Using photons to process data should be inherently more rapid than using electronic circuits. But although optical systems have indeed accelerated the speed of computation considerably, researchers have continued to seek ways of reducing the energy involved, often by altering the properties of the light pulses being used in the processing.

"In order to classify data in a neural network, each node, or 'neuron', must make a decision to fire or not based on weighted input data," said EPFL's Christophe Moser. "This decision leads to a 'nonlinear' transformation of the data, meaning the output is not directly proportional to the input."

Although digital neural networks perform nonlinear transformations easily using transistors, the equivalent behavior in optical systems has so far been created via raw power, and forcing photons to interact indirectly through light energy alone while they pass along an optical material.

The EPFL project worked towards a more energy-efficient method for performing these nonlinear computations optically, developing a framework termed "non-linear processing with only linear optics" or nPOLO.

Energy usage reduced by eight orders of magnitude

Deep neural networks operate by leveraging multiple layers of data processing, but the challenge of this approach in optical networks lies in realizing multiple optical layers without resorting to electronic components.

The nPOLO approach does so by creating multiple planes of modulated light using a spatial light modulator. Multiple sequential instances of scattering can effectively encode the pixels of an image spatially into the altered parameters of a low-power laser beam, carrying the data layer by layer. This synthesizes linear and nonlinear transformations at the same time, according to EPFL, but using a fraction of the energy usually needed to create non-linearity.

The encoding can be carried out two, three or even ten times, increasing both the extent of non-linearity and the precision of the calculation. EPFL estimates that the energy required to optically compute a multiplication using nPOLO is eight orders of magnitude less than that required for an electronic system.

The scalability of this low-energy approach could prove to be a major advantage, as the ultimate goal would be to use hybrid electronic-optical systems to fully mitigate the energy consumption of digital neural networks. This in turn will require a compiler able to translate digital data into code that optical systems can use, a challenge to which the researchers have now turned their attention.

"Our image classification experiments on three different datasets showed that our method is scalable, and a promising platform for realizing optical neural networks," said Demetri Psaltis from EPFL Optics Laboratory.