Dynamic Memristor-Based Reservoir Computing for Gesture Recognition Application

In recent years, brain-inspired computing has garnered widespread research interest. Compared to traditional recurrent neural networks (RNNs), reservoir computing (RC) is a low training cost network that efficiently processes temporal signals. A key challenge in hardware implementation is generating a rich reservoir state. In this work, we report a dynamic memristor-based reservoir computing system that achieves diverse reservoir states through a time division multiplexing masking process. By using an AgInSbTe (AIST) layer in electrochemical metallization (ECM) memory, we demonstrate substantial potential for stable RC. This AIST-based RC systems exhibits nonlinearity, cycling stability, and relatively low average power consumption, meeting the requirements of RC systems. The system achieves a gesture recognition accuracy of up to 96%. Our work could pave the way for future efficient memristor-based RC systems capable of handling more complex temporal tasks.