LAUSR

Reservoir computing (RC) has attracted a lot of attention in the field of machine learning because of its promising performance in a broad range of applications. However, it is difficult to implement standard RC in hardware. Reservoir computers with a single nonlinear neuron subject to delayed feedback (delay-based RC) allow efficient hardware implementation with similar performance to standard RC. We propose and study two different ways to build ensembles of delay-based RC with several delayed neurons (time-delay reservoirs): one using decoupled neurons and the other using coupled neurons through the feedback lines. In both cases, the outputs of the different neurons are linearly combined to solve some benchmark tasks. Simulation results show that these schemes achieve better performance than the single-neuron case. Moreover, the proposed architectures boost the RC processing speed with respect to the single-neuron case. Both schemes are found to be robust against small mismatches between delayed neuron parameters.