Physical reservoir computing (RC) extracts temporal features by utilizing the inherent physical characteristics of materials. Although memristor-based RC systems process time series effectively, they have issues with explainability and compatibility for edge intelligence. In order to go beyond single-mode current/conductance sampling in conventional physical RC, this work makes use of the coexistence of nonlinear event detection and linear exact mapping in phase-change materials to construct a novel spiking RC system. Device-mapped features are processed using an innovatively proposed sliding-window spike sampling architecture and a parameter optimization approach that creates a high-dimensional mapping reservoir by combining a simulated annealing algorithm with a generative adversarial network. This system achieves a low error rate of 0.075 in Mackey-Glass time series prediction and 96.67% accuracy in Iris data set classification. Additionally, this work not only introduces a novel material system into physical RC but also establishes a three-dimensional collaborative mapping mechanism to improve explainability by including a weight-quantification-based explainability analysis method. This method is adaptable to broader material platforms for advancing physical RC development.
Li et al. (Mon,) studied this question.