Simulations of a Laser-Based Reservoir Computer

Abstract

Reservoir computing provides a robust energy efficient way to do machine learning by minimizing training time and optimizing power efficiency. Physical reservoir computing relies on the dynamic properties of physical systems to achieve Turing-complete computation. We examine the efficacy of one such system using lasers with saturable absorbers to generate a particular subset of physical reservoir computer known as a liquid state machine which relies on spike pulse trains to perform computation. The computational capability of such systems is explored as the number of nodes is scaled through both investigations into the dynamic responses of the networks and the accuracy of machine learning prediction tasks. Two different network topologies are proposed and contrasted in two benchmarks that test the accuracy of the nodes when classifying time independent data and time dependent data. From these benchmarks, it is shown that computational power scales with the number of nodes in the network while training time does not significantly impact performance beyond a threshold value unique to each problem. Similarly, it is shown that neither topology is computationally superior over the other and the trade-offs of each are discussed.