Researchers at the University of Tokyo, in collaboration with NTT researchers, have developed a new artificial intelligence (AI) framework that could significantly reduce the computational resources required to run AI systems. The approach uses the natural behaviour of physical systems to process information and was demonstrated using an optical computing system that operates with light instead of electrical signals. The findings could contribute to the development of faster and more energy-efficient AI hardware.
The growing size and complexity of AI systems have increased concerns over the energy required to train and operate them. Conventional computer chips rely on electrical signals passing through large numbers of transistors to perform calculations, generating heat and consuming power. Researchers are therefore exploring alternative computing methods that can improve efficiency without continuously increasing the size of AI models. One such approach involves using specialised physical systems to carry out computational tasks.
The research team, led by Associate Professor Kohei Nakajima of the University of Tokyo's Laboratory for Intelligent Systems and Informatics, redesigned the way AI systems learn so that physical hardware could perform computations more naturally. The work was demonstrated through simulations and experiments using optical computing, which processes information using light. Graduate student Yongbo Zhang said the team wanted to rethink the learning process rather than continue developing increasingly complicated models.
"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," Zhang said.
Optical computing offers potential advantages over conventional electronic processing because light can carry information at high speeds and transmit multiple signals simultaneously. These properties make it suitable for parallel processing and potentially reduce energy consumption. However, implementing conventional AI learning methods on optical and other physical systems can be difficult, as they often require extensive measurements, data storage and control of the system during training.
To address this, the researchers combined a training method called state-skipping direct feedback alignment (SS-DFA) with a simplified neural network architecture called the single-map framework. Conventional training methods require information about a system's intermediate states to be measured and recorded as it processes data. The learning algorithm then uses this information to determine how the system should be adjusted to reduce errors.
The new method avoids accessing most of these intermediate states. Instead, it uses the final error and information about the system to guide the learning process. This reduces the requirements for measurement, data storage and computation. The researchers also simplified the network architecture by allowing multiple units to follow general operating rules rather than depending on complicated internal connections.
The simplified framework achieved performance comparable to more complex neural network architectures and outperformed them in some tests, while requiring substantially less computation. The researchers also demonstrated that patterns of activity previously observed in deep neural networks, including large language models, could be reproduced in an optical computing system.
The team has integrated the framework into several existing neural network architectures, with results indicating improved performance alongside reduced computational requirements. However, the approach has not yet been tested on the largest AI models, particularly large language models and multimodal systems that process different types of data.
Nakajima said the framework's modular design could allow it to be incorporated into existing networks as a computational component, providing a possible route towards larger applications.
The researchers plan to investigate whether the framework can be scaled to larger models, handle more complex tasks involving multiple types of data and operate across different computing platforms. If successfully developed, optical systems could also reduce the energy required for cooling AI infrastructure because they generate less heat. This could lower the electricity and water consumption associated with cooling future AI data centres.
The team also sees potential applications in devices such as smartphones, cameras and self-driving cars. Integrating a miniaturised version of the system onto chips could allow data to be processed within these devices, enabling faster responses with lower energy consumption. However, Nakajima said further work was needed to address miniaturisation, system stability and large-scale manufacturing before such applications become practical.