Feature-Selective Preprocessing with Electrically Robust Boron Nitride-Based Dynamic Memristors for Reliable...

Reservoir computing (RC) has gained attention for its ability to improve the efficiency of artificial intelligence (AI) computations through dynamic preprocessing of input data. However, RC has an accuracy limitation due to data loss when mapping original inputs into a high-dimensional computational space. Furthermore, physical reservoirs switch frequently in RC systems, making switching endurance a critical factor that requires further investigation. Here, we report an electrically robust Cu/a-BN/Ti(TiO_X) dynamic memristor with excellent endurance and uniformity, utilizing a built-in series resistance structure to suppress filament overgrowth in the device. The device exhibits volatile and gradual switching with multiple reservoir dynamics. It also maintains identical reservoir dynamics over 1.2 million switching updates. Noisy MNIST images and satellite image classification simulations reveal that the reservoir simplifies images by selecting spatially clustered key features in the input images. Accordingly, the implementation of reservoir computing enhances the network efficiency, achieving over 5% higher accuracy with 50% fewer network parameters. In addition, the advantages of processing wide RC (WRC) using three-dimensional kernels in convolutional neural networks (CNNs) were identified. Based on these findings, a WRC/CNN architecture is proposed that achieves high accuracy while minimizing the increase in energy consumption in WRC applications, thereby enabling lightweight neural networks for edge device AI computing.