Neuromorphing computing using photorefractive materials

Neuromorphic computing using photonic systems has progressed significantly over recent years, in particular thanks to the development of innovative materials and systems that work as photonic neurons, photonic synapses and nonlinear optical transformations. We discuss here a proposal that makes use of a photorefractive crystal as a so-called reservoir computer (RC). Light is phase-encoded with optical input data and its transformation after passing the crystal is recorded and enables the training of the analog neural network weights. We systematically measure the nonlinearity of the resulting light-matter interaction and correlate that nonlinearity to the performance of the reservoir computer for different benchmark tasks. State-of-the-art performances are achieved for example for multi-step prediction of time-series generated by nonlinear systems.