Influence of Compact Memristors’ Stability on Machine Learning

Abstract

Memristors attract wide attention due to its high integration and parallel computation, having great potential to promote the development of machine learning. As memristors are prone to internal and external variabilities, their variabilities hurt the performance of memristors and, therefore, the performance of memristive neural networks. In this paper, the influence of memristors' stability on machine learning is analyzed. Based on a filamentary memristors' compact circuit model, two typical machine learning methods, a feed-forward network and a data clustering, as the representatives of supervised and unsupervised learnings, are tested, following the model's four variation parameters, the variations of maximum memristances, of conductive filaments' change speeds, of initial conductive filaments' lengths, and of minimum memristances. Results show that in a feed-forward network, the changing speeds of conductive filaments' length play a key role. What is more, the smaller feed-forward network tends to worse performance. In data clustering, the variations of the maximum and minimum memristances have a determinant effect on performance. While the variations of conductive filaments' change, speeds, and initial conductive filaments' lengths show random influence. Moreover, the migration trend of clustering centers does not change with the size of neural networks. We hope the exploration in this paper can deepen the understanding of memristor's role in machine learning and give guidelines for the design and fabrication of memristive neural networks.