Stochastic Computing-based on-chip Training Circuitry for Reservoir Computing Systems

Reservoir Computing (RC) is considered an emerging computational paradigm for the analysis of temporal data, where the training of RC systems is usually implemented through the use of a linear regression. Most RC hardware approaches perform the training stage off-chip at the server, thereby increasing the processing time, the latency and the power dissipation. This work proposes a non-iterative supervised learning method for RC systems which consists of a tropical-algebra-based regression that has been implemented in hardware using Stochastic Computing techniques. Our approach is capable of integrating a non-iterative training together with the inference in a simple and compact circuitry.