Accuracy of a tensor network-based image classifier saturates with bond dimension

Abstract

In the field of physics, tensor networks have been successfully applied to solve for problems requiring high-dimensionality. As data required in data science are also high-dimensional, tensor methods have also gained appreciation as a tool in machine learning. In this study, we have utilized an existing tensor network-based machine learning program to investigate on the role of bond dimension in the learning rate and the learning accuracy in an image classification problem. In our simulation, the increase in bond dimension leads to improved accuracy but easily reached a saturation point where the gains are overshadowed by the computational cost.