Machine-learning system based on light could yield more powerful, efficient large language models
Aug. 23, 2023.
1 min. read Interactions
A 100-fold improvement in energy efficiency and 25-fold improvement in compute density
Physical neural network learns and remembers ‘on the fly’ like neurons for low-energy machine intelligence
An MIT-led team has developed a system based on light that could lead to machine-learning programs several orders of magnitude more powerful than the one behind ChatGPT, and that uses several orders of magnitude less energy than state-of-the-art supercomputers, according to Elizabeth A. Thomson at MIT’s Materials Research Laboratory.
The team reports a greater than 100-fold improvement in energy efficiency and a 25-fold improvement in compute density (a measure of the power of a system) over the state-of-the-art supercomputers behind the machine-learning models of today.
Computations based on light
In the July 17 issue of Nature Photonics, the researchers report the first experimental demonstration of the new system, which performs its computations based on the movement of light, rather than electrons, using hundreds of micron-scale lasers.
In the paper, the team also cites “substantially several more orders of magnitude for future improvement.” As a result, cell phones and other small devices could become capable of running programs that can currently only be computed at large data centers.
In addition, the components of the system can be created using fabrication processes already in use today, so “we expect that it could be scaled for commercial use in a few years,” says Zaijun Chen, first author, an assistant professor at the University of Southern California. “For example, the laser arrays involved are widely used in cell-phone face ID and data communication.”
Citation: Chen, Z., Sludds, A., Davis, R., Christen, I., Bernstein, L., Ateshian, L., Heuser, T., Heermeier, N., Lott, J. A., Reitzenstein, S., Hamerly, R., & Englund, D. (2023). Deep learning with coherent VCSEL neural networks. Nature Photonics, 17(8), 723-730. https://doi.org/10.1038/s41566-023-01233-w