An Adaptive Training Method for Memristive Next Generation Reservoir Computing Systems

Next-generation reservoir computing, combining reservoir computing with statistical prediction methods, has gained attention for its efficiency in computing time and dataset requirements for learning. This approach has proven effective for predicting the behavior of complex dynamical systems, which is especially relevant in hardware implementations. We recently showed that memristors are promising candidates for encoding synaptic weights in hardware. However, ridge regression, typically used for training artificial neural networks, requires a precise adjustment of the memristors’ resistances. Also, the memristive weights could be dynamically adapted over training, but this capability is not utilized in ridge regression. This work proposes an innovative next generation reservoir computing structure, which, subject to an adaptive learning process, becomes tolerant to the cycle-to-cycle variability inherent to memristive synapses. In particular, the network synaptic weights are conditionally modulated in steps over the learning phase through a gradient descent-based training protocol. This allows to mitigate the detrimental effects on the reservoir time series prediction accuracy due to the unavoidable stochastic errors originating in the process of writing weights into the memristive physical stacks. Additionally, the adaptive learning process may compensate for potential changes in the source data to predict.