Benchmarking a tensor network-based machine learning algorithm for supervised image classification

Abstract

Real-world data is often scarce and the careful setting of hyperparameters to efficiently train machine learning algorithms is necessary. In this study the role of bond dimension on the performance of an existing tensor network-based machine learning program was benchmarked on the MNIST and fashion-MNIST data sets. In particular, the effects of bond dimension on the accuracy and execution times of the algorithm were investigated. Results show that a bond dimension of less than 100 is sufficient to reach an acceptable classification accuracy, and increasing the bond dimension further extends the execution time and consumes prohibitive amounts of memory without providing a reasonable improvement in accuracy.