Light-powered AI: Tokyo researchers cut computation with a new way to train optical hardware
Researchers in Japan have demonstrated a way to run AI on hardware that computes with light instead of electricity, using a simplified training method they say cuts the amount of computation needed.
The work, by a team at the University of Tokyo with researchers from the telecom company NTT, was published Nov. 5 in Nature Computational Science.
AI’s energy problem
AI systems are getting bigger and more power-hungry, and training large models in particular uses huge amounts of energy. One alternative is “physical” AI, which lets the natural behaviour of a physical system do part of the computation. Optical systems are attractive because light carries information very quickly, can carry many signals at once, and produces little heat.
The catch is that standard AI training methods are hard to run on such hardware. They need the system’s intermediate states to be measured and stored at every step, so that learning can work backward to figure out what to change.
Skipping the hard parts
The team combined a training method called state-skipping direct feedback alignment (SS-DFA) with a simpler network design they call a “single-map” architecture. The training method uses mainly the final error, plus some information about the system, to guide learning. That avoids most of the measurement and data storage. The simpler architecture has many units following the same general rules instead of relying on complex internal connections.
In computer simulations and in experiments on an optical computing system, the simplified design matched, and in some tests outperformed, more complex neural network designs while needing far less computation, the researchers say.
Not ready for ChatGPT-scale models yet
“We have not yet tested it on today’s largest AI models, particularly language and so-called multimodal models, and this is an important direction for future research,” said Associate Professor Kohei Nakajima, who led the team. The framework is modular, he said, so it could eventually be slotted into existing networks. If the system can be shrunk onto a chip, it could run AI directly in phones, cameras or self-driving cars.
The researchers say the approach could eventually reduce the energy, cooling and water use of AI data centres. But miniaturization, stability and large-scale manufacturing all still need work.
Source: Zhang, Y., Nakajima, M., Inoue, K., Hashimoto, T., Kuniyoshi, Y., & Nakajima, K., “Scalable physical deep learning using optical dynamics with state-skipping training and single-map architecture,” Nature Computational Science, Nov. 5, 2026. doi:10.1038/s43588-026-01075-w; University of Tokyo news release.
