Optimal nonlinear information processing capacity in delay-based reservoir computers

Reservoir computing is a recently introduced brain-inspired machine learning paradigm. We focus on time-delay reservoir (TDR) computers that have been physically implemented us-ing optical and electronic systems and show excellent com-putational performances at unprecedented information pro-cessing speeds. TDRs are known for their easy-to-implement training but also for their problematic sensitivity to architec-ture parameters. We address the reservoir design problem, which remains the biggest challenge at the time of applying the reservoir computing techniques to sophisticated machine learning tasks. More specifically, the information available regarding the optimal operating regimes of a reservoir is used to construct a functional link between its parameters and its performance. This function is then used to explore various properties of the device and to choose the optimal architec-ture, thus replacing tedious and time consuming parameter scannings by well-structured optimization problems. Architecture of the time-delay reservoir (TDR) computer c c