Physical Reservoir Computingの数理

抄録

Reservoir computing (RC) was first proposed as a framework to train recurrent neural networks. In this framework, a low-dimensional input is projected to high-dimensional dynamical systems, which are typi-cally referred to as a reservoir. If the dynamics of the reservoir involve adequate nonlinearity and memory, em-ulating nonlinear dynamical systems only requires adding a linear, static readout from the high-dimensional state space of the reservoir. Because of its generic nature, RC is not limited to digital simulations of neural net-works, and any high-dimensional dynamical system can serve as a reservoir if it has the appropriate properties. The approach using a physical entity rather than abstract computational units as a reservoir is called physical reservoir computing (PRC). In this presentation, several novel platforms based on PRC are introduced using physical substrates. These platforms include soft materials (e.g., silicone-based soft robotic arm) and faraday waves generated on the water surface (which we call, “physical liquid state machines”), and they illustrate the potentials of the framework through a number of experiments. The focus will particularly be on how dynamical system aspects can provide a novel view of the PRC framework, including the relevance of noise-induced phe-nomena or random dynamical systems.