Optics makes machine learning much faster

We demonstrate a novel implementation of an artificial neural network (ANN) using integrated photonics technology, which was developed for telecom applications. The photonic ANN has a potential to make machine learning much faster as light can emulate a large-scale neuro response with ultrafast propagation speed owing to its parallelism. Among ANN concepts, we focus on reservoir computing (RC), a neuromorphic system that mimics the human cerebellum, because it is highly suitable for photonic implementation. Here we present the first prototype of the photonic RC. The device successfully emulates neuro responses with ultrafast processing speed (subnano seconds), which is 3 or 4 digits faster than CPU processing. In addition, since the light has complex-valued amplitude, it processes the two dimensional inputs in the same optics. By upgrading the parallel nature of light using wavelength and space division multiplexing, the performance will exceed that of state-of-the-art exa-scale computers in future.

Please click the icon to open the full-size PDF file.