Increasing AI Efficiency with Lasers

image provided by pixabay

This post is also available in: עברית (Hebrew)

Scientists from various institutions demonstrated a new optical neural network training method that could revolutionize electronic microprocessors, and systems like ChatGPT could soon be trained with an immense increase in energy efficiency. This means that the most advanced AI models could be trained with a hundred times less energy, taking up much less space at the same speed.

According to Cybernews, artificial neural networks imitate the way biological brains process information. These AI systems are built to learn, combine, and summarize information from large data sets, and are reshaping the field of information processing.

Current AI models can reach hundreds of billions of artificial neurons and present a challenge to current hardware capabilities. The paper showed that an optical neural network (ONN) approach could overcome the current limitations, stating: “Our technique opens an avenue to large-scale optoelectronic processors to accelerate machine learning tasks from data centers to decentralized edge devices.”

The ONN approach holds great promise to alleviate the bottlenecks of traditional processors such as transistor count, energy consumption in data movement, and semiconductor size. ONNs use light, which can carry a lot of information at once thanks to large optical bandwidth and low-loss data transmission. To move the light around for calculations, the researchers used many laser beams, and believe that the demonstrated system is scalable through mature wafer-scale fabrication processes and photonic integration.

Dirk Englund, an associate professor in MIT’s Department of Electrical Engineering and Computer Science and leader of the work, explains that models such as ChatGPT are limited in their size by the power of today’s supercomputers, so it is not economically viable to train larger models. “Our new technology could make it possible to leapfrog to machine-learning models that otherwise would not be reachable in the near future,” he concludes.

This information was provided by Cybernews.