Photonic Processor for Artificial Neural Networks

This futuristic drawing shows programmable nanophotonic processors integrated on a printed circuit board and carrying out deep learning computing. (Source: RedCube Inc. / MIT)

“Deep learning” computer systems, based on artificial neural networks that mimic the way the brain learns from an accu­mulation of examples, have become a hot topic in computer science. In addition to enabling techno­logies such as face- and voice-recog­nition software, these systems could scour vast amounts of medical data to find patterns that could be useful diagnos­tically, or scan chemical formulas for possible new pharma­ceuticals. But the compu­tations these systems must carry out are highly complex and demanding, even for the most powerful computers.

Now, a team of researchers at MIT and elsewhere has developed a new approach to such compu­tations, using light instead of elec­tricity, which they say could vastly improve the speed and effi­ciency of certain deep learning compu­tations. Marin Soljačić says that many researchers over the years have made claims about optics-based computers, but that “people drama­tically over-promised, and it backfired.” While many proposed uses of such photonic computers turned out not to be practical, a light-based neural-network system developed by this team “may be applicable for deep-learning for some appli­cations,” he says.

Traditional computer archi­tectures are not very efficient when it comes to the kinds of calcu­lations needed for certain important neural-network tasks. Such tasks typically involve repeated multi­plications of matrices, which can be very compu­tationally intensive in conven­tional CPU or GPU chips. After years of research, the MIT team has come up with a way of performing these opera­tions opti­cally instead. “This chip, once you tune it, can carry out matrix multi­plication with, in principle, zero energy, almost instantly,” Soljačić says. “We’ve demonstrated the crucial building blocks but not yet the full system.”

By way of analogy, Soljačić points out that even an ordinary eyeglass lens carries out Fourier transform on the light waves that pass through it. The way light beams carry out compu­tations in the new photonic chips is far more general but has a similar underlying principle. The new approach uses multiple light beams directed in such a way that their waves interact with each other, producing inter­ference patterns that convey the result of the intended operation. The resulting device is something the researchers call a pro­grammable nano­photonic processor.

The result, Yichen Shen says, is that the optical chips using this archi­tecture could, in principle, carry out calcu­lations performed in typical arti­ficial intelli­gence algo­rithms much faster and using less than one-thousandth as much energy per operation as con­ventional electronic chips. “The natural advantage of using light to do matrix multi­plication plays a big part in the speed up and power savings, because dense matrix multi­plications are the most power hungry and time consuming part in AI algorithms” he says. The new pro­grammable nano­photonic processor uses an array of wave­guides that are inter­connected in a way that can be modified as needed, programming that set of beams for a specific compu­tation.

“You can program in any matrix operation,” Nicholas Harris says. The processor guides light through a series of coupled photonic wave­guides. The team’s full proposal calls for inter­leaved layers of devices that apply an operation called a nonlinear acti­vation function, in analogy with the operation of neurons in the brain. To demonstrate the concept, the team set the pro­grammable nano­photonic processor to implement a neural network that recog­nizes four basic vowel sounds. Even with this rudi­mentary system, they were able to achieve a 77 percent accuracy level, compared to about 90 percent for conven­tional systems. There are “no sub­stantial obstacles” to scaling up the system for greater accuracy, Soljačić says.

Dirk Englund adds that the pro­grammable nano­photonic processor could have other appli­cations as well, including signal processing for data trans­mission. “High-speed analog signal processing is something this could manage” faster than other approaches that first convert the signal to digital form, since light is an in­herently analog medium. “This approach could do processing directly in the analog domain,” he says. The team says it will still take a lot more effort and time to make this system useful; however, once the system is scaled up and fully func­tioning, it can find many user cases, such as data centers or security systems. The system could also be a boon for self-driving cars or drones, says Harris, or “whenever you need to do a lot of compu­tation but you don’t have a lot of power or time.” (Source: MIT)

Reference: Y. Shen et al.: Deep learning with coherent nanophotonic circuits, Nat. Phot., online 

Link: Research Laboratory of Electronics, Massachusetts Institute of Technology, Cambridge, USA

Speak Your Mind