Artificial intelligence (AI) is becoming increasingly powerful, but at the same time very "greedy": training large models requires enormous computing resources and energy. A group of researchers, including staff from the University of Tokyo and the company NTT, has proposed an approach that simplifies both the training and the design of so-called physical AI systems. In these, the computations are performed not by an ordinary electronic chip but by the physical medium itself - for example, light. The results were published in the journal Nature Computational Science.
Conventional processors compute using electrical signals passing through many transistors. This requires energy and generates heat. Light, by contrast, can carry information very quickly and over several channels at once. Optical computing has long been considered a promising platform for AI, but transferring familiar algorithms to such hardware is fraught with difficulties: much must be measured, controlled and stored as intermediate states of the system.
To get around this, the team led by Associate Professor Kohei Nakajima combined a new training method, SS-DFA (simplified: training using the final error), with a simpler network architecture called the single-map framework. In conventional training, one has to track how the system changes step by step and then "rewind" the process to understand what to correct. The new method relies mainly on the final error and general information about the system, so it requires less measurement, data storage and computation.
In addition, the researchers simplified the structure: instead of many complex internal connections, many elements operate by common rules. The approach was tested in computer simulations and experiments on an optical computing system. Unexpectedly, the simplified design did not lag behind more complex neural networks and in some tests outperformed them, using noticeably less computation.
The authors emphasize that so far this is a laboratory demonstration, not a finished product. They did not test the method on the largest modern AI models, especially language and multimodal ones (those that work with text, images and other types of data at once). Ahead lies verification of scaling, more complex tasks and other computing platforms. Because the development is modular, it can be inserted into existing networks as a separate computing block.
If the system can be shrunk and built into a chip, it could process data directly in phones, cameras or self-driving cars - faster and with lower energy consumption. However, this still requires work on miniaturization, stability and large-scale production. The potential gain is a reduction in energy costs not only for computing but also for cooling future data centers, as well as lower water consumption.