What if AI didn’t need electricity to think? Scientists are turning light into a computer

A computer usually handles information by moving electrical signals through circuits. This approach powers modern artificial intelligence, but it also produces heat and adds to the energy needed to run increasingly demanding systems. Researchers are investigating whether some computing work can instead use the behavior of light.

A team led by the University of Tokyo , working with NTT researchers, has demonstrated a way to train AI on light-based hardware. Published November 5 in Nature Computational Science , the work combines a new training technique with a simplified neural network design. The experiment used an optical fiber loop. It is an early demonstration, not evidence that AI can operate without electricity.

Researchers are teaching light to do the work of electronic computers

Photonic computing uses light to process information rather than relying solely on electrical signals propagating through transistors. Light can carry information very quickly. It can also carry multiple signals simultaneously, which gives optical systems the potential for speed and parallelism. They may also generate less heat, which in turn would lead to lower power consumption for both computing and cooling.

Adapting AI training to physical hardware presents a challenge. Many established methods measure and store what happens inside a system as information passes through it. The training process then uses these records to work backward from an error and adjust the network. Collecting and retaining this information adds to the computational burden.

The Tokyo team combined two changes to address this problem. Its state-skipping direct feedback alignment method, or SS-DFA, changes how the system learns. A simpler design called the single-map framework changes the structure of the network.

Researchers say the combined approach matched more complex neural networks and outperformed them in some tests while requiring substantially less computation. They evaluated this in simulations and optical hardware experiments.

A new learning method makes optical AI possible

SS-DFA avoids recording most intermediate states during training. Instead of preserving a detailed account of each stage and retracing it, the method uses the final error and information about the physical system to guide adjustments. The researchers say this reduces measurement, data storage, and computation.

The single-map framework simplifies the network itself. Rather than depending on many complicated internal connections, its units follow general shared rules. Combining the two approaches is intended to make AI learning better suited to physical systems, where the hardware's natural behavior contributes to the computation.

For the hardware demonstration, information was converted into an optical signal and sent through a 200-meter fiber loop. The light-based system processed the signal before it was converted back into electricity for analysis. The setup also used a small number of electronic components. It was therefore not an electricity-free computer.

The study shows that light can perform part of an AI computation, not that electrical power has been removed from the machine. The researchers are exploring other ways to handle processing, with potential efficiency benefits if the approach can be developed further.

A 200-meter optical fiber could point toward more efficient AI

The team of researchers has not tested the framework on today's largest language models or multimodal AI systems. It has incorporated the approach into several relatively basic modern neural network architectures, where the researchers report improved performance alongside reduced computation. Testing larger models and more complex tasks is future work.

Practical use also depends on solving hardware problems. The optical setup must be miniaturized, made stable, and manufactured at scale. The researchers suggest that integrated versions could eventually process information inside phones, cameras, or autonomous vehicles, but these applications remain possibilities rather than demonstrated products.

The study does not replace electronic computers or show that AI can operate without power. If the method can be scaled and the hardware made practical, it could help reduce the resources required for some AI workloads.

Sources: EurekAlert , Tech Xplore

Affiliate links on Cult of Mac may earn us and our partners a commission.