MIT News  December 2, 2024
Optical systems can perform linear matrix operations at an exceptionally high rate and efficiency. However, demonstrating coherent, ultralow-latency optical processing of deep neural networks has remained an outstanding challenge. A team of researchers in the US (MIT, industries) realized such a system in a scalable photonic integrated circuit that monolithically integrated multiple coherent optical processor units for matrix algebra and nonlinear activation functions into a single chip and experimentally demonstrated fully integrated coherent optical neural network architecture for a deep neural network with six neurons and three layers that optically computed both linear and nonlinear functions, unlocking new applications that required ultrafast, direct processing of optical signals. They implemented backpropagation-free in situ training on this system on a six-class vowel classification task. According to the researchers their work provides experimental evidence to theoretical proposals for in situ training, enabling orders of magnitude improvements in the throughput of training data, and the fully integrated coherent optical neural network opens the path to inference at nanosecond latency and femtojoule per operation energy efficiency… read more. TECHNICAL ARTICLE
Backpropagation-free in situ training. Credit: Nature Photonics, 2 December, 2024