University of Tokyo proposes single-mapping optical AI framework

A team led by Kohei Nakajima at the University of Tokyo, together with NTT researchers, proposed a "single-mapping framework" architecture and, combined with the SS-DFA training method, demonstrated AI computing through computer simulations and experiments on an optical computing system consisting of a 200-meter optical fiber and a small number of electronic components. The method does not need to record most intermediate states and uses only the final error to guide learning, thereby reducing measurement, storage, and computation; the team says the simplified design matches or even exceeds more complex neural network architectures in some tests. The team says that for this approach to become practical, further research is still needed in miniaturization, stability, and large-scale manufacturing, and it has not yet been validated on today's largest language models and multimodal models; future plans are to scale up and explore chip-based integration.

AI overview based on source reports · Updated 19 hr ago