Photonic Delay-Based Reservoir Computers as Deep Neural Network Preprocessors

Photonic reservoir computing (RC) has been effectively used for solving various complex problems. RC systems are straightforward to train when compared to deep neural networks (DNNs), and doing RC in the photonics domain offers the advantage of high-speed performance, low-energy consumption and the possibility of high inherent parallelism [1], [2]. We propose and investigate to use the output of such a reservoir to preprocess the input data before this data is send though a DNN. The main idea here is to use such a photonic reservoir to transform the input data into a higher dimensional state-space, which could allow the DNN to process the data with increased performance. Based on numerical simulations of delay-based reservoirs using a single-mode semiconductor laser [3]–[6], we show that using such a preprocessing reservoir results in an improved performance of DNNs, and that we do not need to carefully fine-tune the parameters of the preprocessing reservoir.