“This is incredibly simple,” says Tianwei Wu, lead author of the study. “You can reprogram it to change the laser pattern on the fly.” Using the system, the researchers designed a neural network that successfully identifies vowel sounds. Most photonic systems need to be trained before they are built, because training necessarily involves reconfiguring connections. But this system is easily reconfigurable, so the researchers trained the model after installing it on the semiconductor. They now plan to increase the size of the chip to encode more information in different colors of light, which should increase the amount of data it can process.
Even Psaltis, who developed facial recognition systems in the '90s, is impressed by the progress: “Our wildest dreams from 40 years ago were very modest compared to what has actually happened.”
The first ray
Optical computing has advanced rapidly in the past few years, but it's still a long way from replacing electronic chips that run neural networks outside of the lab. Papers have shown optical systems outperforming electronic systems, but they're typically running small models using older network designs and small workloads. And many of the reported numbers on optical superiority don't tell the whole story, said Bhavin Shastri of Queen's University in Ontario. “It's very hard to fully equate with electronics,” he said. “For example, when you use lasers, it doesn't say much about the energy that's powering the laser.”
The lab's systems need to scale before they can provide a competitive advantage. “How big do you need to be to win?” McMahon asked. The answer is: extraordinary. That's why no one can match the Nvidia chips that power many of the most advanced AI systems today. Along the way, there will be a ton of hard engineering problems to solve. These are problems that the electronics side has been solving for decades. “Electronics is starting out with a huge advantage,” McMahon said.
Some researchers believe ONN-based AI systems will first find success in specialized applications where they offer unique advantages. Shastri says one promising use is canceling out interference between different wireless transmissions, such as 5G cell towers or radar altimeters that help airplanes navigate. Earlier this year, Shastri and a few colleagues created an ONN that can sort through different transmissions and pick out signals of interest in real time, with processing latency of less than 15 picoseconds (15 trillionths of a second)—less than a thousandth of the time it would take for electronic systems, and consuming 70 times less power.
But McMahon says the grand vision of optical neural networks outperforming electronic systems in general use is worth pursuing. Last year, his group ran simulations showing that within a decade, large enough optical systems could make some AI models more than 1,000 times more efficient than future electronic systems. “Right now a lot of companies are trying really hard to get 1.5x returns. If we could get 1,000x returns, that would be great,” he says. “It's probably a 10-year project, if it's successful.”
original work Reprinted with permission. Quanta Magazine, Editorially independent publication Simons Foundation Its mission is to enhance public understanding of science through reporting on research developments and trends in mathematics, physical sciences, and life sciences.
