“We have something extremely simple here,” he said Tianwei Wu, lead author of the study. “We can reprogram it by changing the laser patterns on the fly.” The researchers used the system to design a neural network that effectively distinguishes vowel sounds. Most photonic systems must be trained before they are built, because training necessarily involves reconfiguring connections. However, because this system can be easily reconfigured, the researchers trained the model after it was installed on the semiconductor. They now plan to augment the size of the chip and encode more information in different colors of featherlight, which should augment the amount of data it can handle.
Even Psaltis, the creator of facial recognition in the 1990s, finds this progress impressive. “Our wildest dreams 40 years ago were very modest compared to what actually happened.”
The first rays of featherlight
While optical computing has made rapid progress over the past few years, it is still a long way from displacing electronic chips that run neural networks outside the lab. Articles claim that photonic systems perform better than electronic systems, but they generally employ petite models, using ancient network designs and petite loads. Much of the reported data on photonic supremacy does not tell the whole story, said Bhavin Shastri of Queen’s University in Ontario. “It’s very difficult to compare apples to apples with electronics,” he said. “For example, when they use lasers, they don’t really talk about the energy needed to power the lasers.”
Laboratory systems must be scaled before they can demonstrate competitive advantage. “How much does it take to win?” McMahon asked. Answer: extremely enormous. That’s why no one can compare to a chip from Nvidia, whose chips power many of the most advanced AI systems today. Along the way, there is a huge list of engineering puzzles to solve – problems that electronics have been solving for decades. “Electronics is starting with a big advantage,” McMahon said.
Some researchers believe that ONN-based AI systems will first succeed in specialized applications where they provide unique benefits. Shastri said a promising application is countering interference between various wireless transmissions, such as 5G cell towers and radar altimeters that facilitate planes navigate. Earlier this year, Shastri and some colleagues created ONN that can sort through various transmissions and select the signal of interest in real time and with a processing delay of less than 15 picoseconds (15 trillionths of a second) – less than one thousandth of the time it would take an electronic system, while using less than 1/70th the power.
But McMahon said the gigantic vision is worth pursuing – an optical neural network that can outperform general-purpose electronic systems. Last year his group conducted simulations showing that within a decade, a sufficiently enormous optical system could make some AI models more than 1,000 times more capable than future electronic systems. “Many companies are now trying to get 1.5 times the benefit. A thousandfold benefit, that would be amazing,” he said. “This could be a 10-year project if successful.”
Original story reprinted with permission Quanta Magazine, editorially independent publication titled Simons Foundation whose mission is to augment society’s understanding of science by incorporating research developments and trends in mathematics and the physical and life sciences.