Reservoir computing with continuous variable quantum systems

I will present our very recent theoretical results on reservoir computing based on continuous variable (CV) quantum systems acting as the reservoir. Reservoir computing is a machine learning paradigm suited in particular for online processing of temporal signals. The learning theory is quite general and applicable to cases where a desired input-output map is realized by driving a generic complex system (the reservoir) with the input. The map is achieved by training a simple readout mechanism that implements a function of the reservoir observables to reals. Apart from a few pioneering works with spin systems, the full potential and viability of reservoir computing with quantum systems is still mostly unexplored territory. Our research focuses on a particular, experimentally feasible class of CV quantum systems and states. Despite being in many ways a simple model, we show that it constitutes a universal class of systems for reservoir computing. We analyze the information processing capacity of generic instances of the model and show that the nonlinearity of the reservoir memory is easily tunable, ranging from fully linear to strongly nonlinear. Finally, we verify that working with quantum fluctuations instead of classical quantities of light emission does not lead to a loss in computational power.