Answers are grounded in this paper and cite their sources.
Neuromorphic hardware for spatiotemporal information processing requires both transient dynamics and programmable weight storage, whereas existing implementations typically rely either on heterogeneous integration of volatile dynamic units and nonvolatile memory devices or on compliance-current regulation to reconfigure volatile/nonvolatile operation within a single device, increasing device diversity or peripheral control complexity. This work reports a stimulus-reconfigurable Ag/ZnS/FTO memristor in which the stability of Ag conductive pathways is regulated by the strength and temporal scale of external stimuli, enabling volatile short-term dynamics and nonvolatile conductance modulation within a single device. Under weak or short stimuli, the device exhibits multiscale short-term plasticity governed by pulse amplitude, width, interval, and number, enabling Morse-code temporal encoding. Under strong or prolonged stimuli, it realizes long-term potentiation/depression and multilevel conductance modulation. Leveraging this dual-mode behavior, a volatile-nonvolatile reservoir computing architecture is constructed, where volatile responses act as reservoir states for dynamic feature expansion and nonvolatile conductance stores readout weights. For spatial image recognition, the architecture achieves 92.94 and 82.27% accuracies on MNIST and Fashion-MNIST, respectively, with an estimated 84.5% fewer devices than a purely nonvolatile multilayer perceptron. The same architecture further achieves 96.13% accuracy and a 95.14% Macro F1 score in ECG heartbeat classification, highlighting a unified device strategy for hardware-efficient spatiotemporal neuromorphic computing.
No takes yet. Share an insight, caveat, or question.
Lu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: