Delay-embedding for recurrence in optical reservoir computing

Time-delay embedding is a prominent method in neural networks to improve the performance of predictions of the dynamics of nonlinear systems. Applied to reservoir computing, a recurrent machine learning model in which the output weights are trained only, an improved prediction performance for a broader range of hyperparameters can be realized. Here, we present an optical implementation of reservoir computing on the basis of a spatial light modulator. In addition to the high parallelism of the optical setup and the fast processing capabilities with scalable numbers of neurons, a recursive cascading effect is generated for successful optical delay embedding. Our laboratory experiment is evaluated for the one-dimensional Kuramoto-Sivashinsky equation and is found to perform better than a classical implementation of reservoir computing.