Neural networks form the basis of a new class of applications and services based on artificial intelligence. Natural language processing, machine translation, face and voice recognition software — all use artificial neural networks. The artificial neural networks are highly demanding, even for the best hardware. Researchers from the Massachusetts Institute of Technology (MIT) have developed a way to accelerate the computation with neural networks based on light instead of electricity. The results of the research has been published in Nature Photonics.

The computation is carried out in the way multiple beams of light interact with each other, a more complicated version of the process by which a regular lens transforms a beam of light. The researchers are referring to the device as a “programmable nanophotonic processor“. 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.

Optical chips using the newly developed architecture can accelerate the calculations performed in artificial intelligence algorithms, as well as consume about one thousandth as much energy as conventional electronics based processors, per operation. The device was put through its paces, where the researchers tested a neural network for recognizing four vowel sounds. The system worked at a 77 percent accuracy level compared to 90 percent of conventional systems. According to the researchers, there are no substantial obstacles for increasing the accuracy of the system.

It will take a “lot of time” for the system to be made useful. However, once the processor is ready, it can be used in data centers and security systems. The nanophotonic processor could be particularly useful in drones or self driving cars, or any other application which is very demanding in terms of computation, but there is not too much power or time available for the calculations.

Publish date: June 13, 2017 7:37 am| Modified date: June 13, 2017 7:37 am

Tags: , , , ,