Photonic reservoir computing with nonlinear dynamical systems

The field of reservoir computing is a fast-developing subfield of machine learning and various systems are explored with respect to their potential for fast and energy efficient computations. Optical reservoir computing systems, e.g. coupled semiconductor lasers, passive cavities, or ring lasers, are promising, because data injection as well as readout can be realized in the optical domain. They are especially suited for tasks that require memory in time, as e.g., time series prediction of a chaotic time evolution. In order to reach good performance on this type of tasks, the internal timescales of the system and the control induced bifurcation structure need to be known as they crucially change the system response. For real-world applications, this tunability can be exploited to optimize the machine learning capabilities. Using a semiconductor laser with optical self-feedback as an example for a photonic reservoir computer, we will numerically investigate the impact of charge carrier and photon lifetimes and explore the role of different data injection schemes. Further, we discuss the effect of adding external input and output delay lines to the setup for mitigating the usually tedious hyperparameter optimization in reservoir computing systems.