Deep Learning With Coherent Nanophotonic Circuits

Light processor recognizes vowels

Nanophotonic module forms the basis for artificial neural networks with extreme computing power and low energy requirements

Supercomputers are approaching the enormous computing power of up to 200 petaflops, ie 200 million billions of operations per second. Nevertheless, they lag far behind the efficiency of human brains, mainly because of their high energy requirements.

A processor based on nanophotonic modules now provides the basis for extremely fast and economical artificial neural networks. As the American developers reported in the magazine "Nature Photonics", their prototype was able to carry out computing operations at a rate of more than 100 gigahertz with light pulses alone.

"We have created the essential building block for an optical neural network, but not yet a complete system," says Yichen Shen, from the Massachusetts Institute of Technology, Cambridge. The nanophotonic processor developed by Shen, together with his colleagues, consists of 56 interferometers, in which light waves interact and form interfering patterns after mutual interference.

These modules are suitable for measuring the phase of a light wave between the wave peak and the wave trough, but can also be used for a targeted change of this phase. In the prototype processor, these interferometers, which in principle correspond, in principle, to a neuron in a neural network, were arranged in a cascade.

After the researchers simulated their concept in advance with elaborate models, they also practically tested it with an algorithm for recognizing vowels. The principle of the photonic processor: A spoken vowel unknown to the system is assigned to a light signal of a laser with a specific wavelength and amplitude. When fed into the interferometer cascade, this light signal interacts with further additionally fed laser pulses and different interference patterns are produced in each interferometer.

To conclude these extremely fast processes, the resulting light signal is detected with a sensitive photodetector and is again assigned to a vowel via an analysis program. This assignment showed that the purely optical system could correctly identify the sound in 138 of 180 test runs. For comparison, the researchers also carried out the recognition with a conventional electronic computer, which achieved a slightly higher hit rate.

This system is still a long way from a photonic light computer, which can perform extremely fast speech recognition or solve even more complex problems. But Shen and colleagues believe it is possible to build artificial neural networks with about 1000 neurons from their nanophotonic building blocks.

In contrast to electronic circuits of conventional computers, the energy requirement is to be reduced by up to two orders of magnitude. This approach is one of the most promising in the future to compete with the viability of living brains.