Modeling of variability-aware memristive neural networks

Abstract

In recent years, memristive neuromorphic systems have gained much attention. In this work, we developed a physics-based framework to model transport in valence change memory (VCM) memristors, implemented in Verilog-A. This has enabled us to scale up and simulate the performance of these devices in a crossbar array/neural network for pattern classification, for instance. The system's performance is analyzed based on classification accuracy in different conditions. We anticipate that this will provide useful insights into the design of these systems by analyzing their performance, based on our model.