Researchers from the Massachusetts Institute of Technology (MIT) have developed a “programmable nanophotonic processor,” a microchip that employs light instead of electrical impulses to implement a neural network. The use of light allows the device to perform neural network operations at a much faster speed while using a fraction of the energy compared to conventional approaches using graphics processing units or central processing units. And to test the processor’s acumen, the researchers had it perform a type of speech recognition listening to vowel sounds.
The programmable nanophotonic processor recognized four vowel sounds at a 77 percent accuracy rate. While this is lower than the 90 percent rate of current systems, there are “no substantial obstacles” to scaling up the processor to attain better accuracy, according to MIT professor Marin Soljačić.
MIT said the programmable nanophotonic processor conducts computations using multiple beams of light that interact with each other. The processor guides the light through a series of coupled photonic waveguides. The waveguides are interconnected in a fashion that allows them to be modified to perform different calculations.
The processor specifically accelerates matrix multiplications, which are particularly taxing on electron-based devices.
The chip accelerates a specific kind of operation used in neural networks, known as matrix multiplications, which are extremely demanding on traditional GPU and CPU architectures.
The researchers estimated that the processor uses less than one-thousandth the amount of electricity consumed by electronic devices when conducting these operations. Performance also is nearly instantaneous compared to electronic processors.
“This chip, once you tune it, can carry out matrix multiplication with, in principle, zero energy, almost instantly,” Soljačić says.
MIT said it has demonstrated the essential building blocks for the nanophotonic processor. However, it still doesn’t have a full system in place. It will likely take a lot of time before it can produce such a complete system.
The processor could be particularly useful in drones, self-driving cars or other applications that have a high demand for fast neural-network computations.
However, the neural network technology has a wide range of other uses, including natural language processing, machine translation and face- and voice-recognition software.
MIT is not the only university applying photonics to the challenge of high-performance neural networks. Princeton researchers in November announced what they billed as the world’s first integrated silicon photonic neuromorphic chip that could be suitable for neural networks. The Princeton team showed the chip could complete a mathematical differential equation nearly 2,000 times faster than a central processing unit chip.
Tyler Schulze is vice president, strategy & development at Veritone. He serves as general manager for developer partnerships, cognitive engine ecosystem, and media ingestion for the Veritone platform. Learn more about our platform and join the Veritone developer ecosystem today.