Researchers at the University of Tokyo have developed a new artificial intelligence framework that uses physical systems to process information more efficiently. Demonstrated on optical computing hardware, the approach uses light instead of conventional electrical signals to perform calculations, potentially reducing the energy and computational resources required to run AI applications.
The framework combines a simplified neural network architecture with a redesigned training method. According to the researchers, it achieved performance comparable to, and in some tests better than, more complex network designs while requiring substantially less computation. The findings could contribute to the development of faster and more energy-efficient AI hardware.
Optical Computing Offers an Alternative to Conventional AI Hardware
The growing capabilities of artificial intelligence have increased demand for computing power, particularly during model training. Conventional computer chips rely on electrical signals passing through large numbers of transistors, a process that consumes energy and generates heat.
Optical computing offers a potential alternative by using light to carry and process information. Light can travel rapidly and transmit multiple signals simultaneously, making it attractive for applications that require high-speed processing and parallel computation.
Associate Professor Kohei Nakajima from the University of Tokyo’s Laboratory for Intelligent Systems and Informatics led the research alongside his team and researchers from NTT. The team explored how AI learning methods could be redesigned to work more naturally with physical computing systems.
“Today’s AI systems have achieved remarkable capabilities, but they also require enormous computational resources. Instead of continuing to build larger and more complicated models, we wanted to rethink the learning process itself so that physical systems can perform computation much more naturally,” said graduate student Yongbo Zhang.
The researchers focused on making AI systems compatible with optical hardware without requiring the extensive measurements, data storage and control associated with conventional training approaches.
Single-Map Framework Simplifies AI Training and Architecture
The new approach combines a training technique called SS-DFA with a simplified neural network architecture known as the single-map framework.
Traditional training methods often require information about a system’s intermediate states to be measured and stored as it processes information. The learning process then uses this data to determine how the system should be adjusted to produce the desired output.
The researchers’ method reduces this requirement by using the final error, alongside information about the system, to guide the learning process. This limits the need to access and record intermediate states, reducing measurement, storage and computational demands.
“We developed a new overall approach combining a training method called SS-DFA with a simpler network architecture called the single-map framework,” said Nakajima. “Conventional training requires information about the system to be measured and recorded as it evolves, before the learning process can then backtrack to determine what needs to change to achieve a desired outcome. Our method avoids the need to access most of these intermediate states, using the final error together with information about the system to guide learning. This reduces the amount of measurement, data storage and computation required. We also designed a simpler architecture in which many units follow the general rules, rather than relying on a set of highly complex internal connections. We demonstrated the combined approach through computer simulations and experiments using an optical computing system.”
To test the framework, the team used an optical computing system featuring a 200-metre fibre loop and a small number of electronic components. Information was converted into optical signals, processed as it travelled through the fibre loop, and then converted back into electrical signals for analysis.
The results showed that the simplified architecture could match or outperform more complex neural network designs in certain tests while using considerably fewer computational resources.
Potential Applications in Energy-Efficient AI Systems
The findings build on earlier research by the team, which showed that complex patterns of activity can naturally emerge within deep neural networks, including large language models. The latest experiments demonstrate that similar principles can be reproduced using optical computing hardware.
The researchers believe the framework could eventually help reduce the energy required to operate AI systems and cool the infrastructure supporting them. Lower heat generation could also reduce water consumption associated with cooling AI data centres.
However, the technology has not yet been tested on the largest AI models, particularly large language models and multimodal systems. Further research will examine whether the framework can handle more complex tasks, process different types of data and operate across a wider range of computing platforms.
“At present, we have integrated the proposed architecture into several relatively baseline modern neural network architectures, and the results show that it can improve performance while reducing computation,” said Nakajima. “However, 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. Because our framework is modular and can be inserted as a computational block into existing networks, it provides a possible route towards such scales. And, if our system can be miniaturized and integrated onto a chip, it could process data directly inside phones, cameras or self-driving cars, enabling faster responses with lower energy consumption. Before this becomes practical, however, further work is needed on miniaturization, stability and large-scale manufacturing.”
Read More :