An Experimentally Validated, Universal Memristor Model Enabling Temporal Neuromorphic Computation

Abstract

The memristor has been identified as a potential solution for achieving non-von Neumann computation due to its ability to perform computation-in-memory, therefore bypassing the memory transfer bottleneck. Many memristive technologies have emerged with various mechanisms ranging from binary resistive switching to fully analog intercalation-based memristors. However, application of memristors to novel computing architectures have been mostly limited to memory arrays and multiply-and-accumulate functions, for example in convolutional neural networks (CNN) [1]. However, future recurrent neural networks (RNN) must implement complex temporal dynamics. Hence, a memristor model is proposed with the versatility to emulate various memristive technologies, including those with a true or virtual dependence on flux-linkage, as well as deploy temporal computation for the investigation of new computing architectures.