Evaluation of memristor performance in neural networks using an AHaH framework

Abstract

Memristor-based neural networks show significant potential for advancing neuromorphic computing by mimicking synaptic behavior. However, their performance can be compromised by various operational conditions, including noise, degradation, and sudden resistance changes.

In this paper, we propose a refined simulation method and a novel device evaluation framework, leveraging the AHaH Framework, to enhance the performance and reliability of memristor-based neural networks. The improved simulation approach is designed to incorporate realistic features, such as linear and non-linear decay, periodic and aperiodic fluctuations, and customizable behaviors, allowing for a more accurate depiction of memristor dynamics. Through this evaluation, critical impacts on neural network accuracy and efficiency are uncovered, particularly under complex noise patterns and degradation scenarios.

The device evaluation framework illustrates how devices, despite exhibiting similar classification accuracy, can display distinct dynamic properties through the monitoring of midpoint voltage variations. These findings provide a basis for robust neuromorphic circuit development.