“What we have here is something incredibly simple,” he said. Tian Wei Wu, lead author of the study. “We can reprogram it, changing the laser patterns on the fly.” The researchers used the system to design a neural network that successfully discriminated vowel sounds. Most photonic systems need to be trained before being built, since training necessarily involves reconfiguring connections. But since this system is easily reconfigured, the researchers trained the model after installing it on the semiconductor. They now plan to increase the size of the chip and encode more information in different colors of light, which should increase the amount of data it can handle.
It’s progress that even Psaltis, which built the facial recognition system in the ’90s, finds impressive. “Our wildest dreams 40 years ago were very modest compared to what really happened.”
First rays of light
While optical computing has advanced rapidly in recent years, it is still far from displacing electronic chips that run neural networks outside of laboratories. The papers advertise photonic systems that perform better than electronic ones, but they typically run small models using old network designs and small workloads. And many of the reported numbers on photonic supremacy don’t tell the whole story, said Bhavin Shastri of Queen’s University in Ontario. “It’s very difficult to do an apples-to-apples comparison with electronics,” he said. “For example, when they use lasers, they don’t really talk about the energy to power them.”
Laboratory systems must be scaled up before they can show competitive advantages. “How big do you have to get to win?” -McMahon asked. The answer: exceptionally large. That’s why no one can match a chip made by Nvidia, whose chips power many of today’s most advanced AI systems. There’s a huge list of engineering puzzles to solve along the way: questions that electronics have solved for decades. “Electronics is getting off to a great start,” McMahon said.
Some researchers believe that ONN-based AI systems will first succeed in specialized applications where they provide unique advantages. Shastri said one promising use is to counter interference between different wireless transmissions, such as 5G cell towers and radar altimeters that help planes navigate. Earlier this year, Shastri and several colleagues created an ONN which can classify different streams and select a signal of interest in real time and with a processing delay of less than 15 picoseconds (15 billionths of a second), less than one-thousandth of the time it would take an electronic system, while using less than 1 /70 power.
But McMahon said the big vision is worth pursuing: an optical neural network that can outperform general-purpose electronic systems. Last year his group I ran simulations showing that, within a decade, a sufficiently large optical system could make some AI models more than 1,000 times more efficient than future electronic systems. “Many companies are now striving for 1.5 times profit. A thousand times greater benefit, that would be incredible,” he stated. “This is maybe a 10-year project, if it’s successful.”
original story reprinted with permission of Quanta Magazine, an editorially independent publication of the Simons Foundation whose mission is to improve public understanding of science by covering developments and trends in research in mathematics and the physical and biological sciences.