How photonic reservoirs improve the performance of deep neural networks

Large deep neural networks (DNNs) have been successfully used for solving complex problems. However, they are often very energy-intensive and time-consuming to train. In this work, we propose to use a photonic reservoir to preprocess the input instead of directly injecting it into the DNN. The advantage of using such a photonic reservoir is that it consists of a network of many randomly connected nodes which do not need to be trained. We numerically show that using such a photonic reservoir as preprocessor results in an improved performance for deep neural networks on a time-series prediction task. We observe that using a stand-alone DNN results in a high test error. If we combine this DNN with a photonic reservoir as preprocessor, we achieve a better performance, shown by a lower test error. Furthermore, it is able to outperform the long short-term memory (LSTM) network with a lower test error and lower total training time. Photonic reservoirs are ideal candidates as physical preprocessors to deep neural networks due to their fast computation times and low-energy consumption.