Double-Layer Reservoir Computer with Three Nonlinear Nodes Based on Polarization-Multiplexed Optoelectronic Feedback Loops

Multi-layer neural networks have enabled a breakthrough compared with single-layer neural networks in sequence prediction, nonlinear channel equalization, and pattern recognition, especially challenging tasks with large database. Time-delayed reservoir computer, which is known as competitive tool to deal with time-dependent information, has a significant advantage of compact hardware compared with other neural networks. To balance the compact hardware and multi-layer's capacity, a double-layer reservoir computer with three nonlinear nodes based on polarization-multiplexed optoelectronic feedback loops is proposed in this work, by using a dual-polarization Mach-Zehnder modulator. A numerical simulation and a referred proof-of-principle experiment are performed. The proposed double-layer reservoir computer has an outstanding performance in Texas Institution and Massachusetts Institute of Technology phoneme recognition task, nonlinear autoregressive moving average 10 and memory capacities evaluation. As the system has a simpler hardware implementation compared with previous schemes, a great potential of optoelectronic hardware implementation of multi-layer reservoir computer is confirmed.