2 July 2021, CMST, 27(3), 99-107

Kurowski, K., Slysz, M., Subocz, M., & Różycki, R.

Applying a Quantum Annealing Based Restricted Boltzmann Machine for MNIST Handwritten Digit Classification

Keywords: machine learning, RBM training, quantum annealing, D-Wave quantum computer, MNIST dataset

Under a Creative Commons license

Open Acces

https://doi.org/10.1016/j.ejor.2023.03.013

Abstract:

As indicated in various recent research, there may still be challenges in achieving acceptable performance using quantum computers for solving practical problems. Nevertheless, we demonstrate promising results by using the recent advent of the D-Wave Advantage quantum annealer to train and test a Restricted Boltzmann Machine for the well studied MNIST dataset. We compare our new model with some tests executed on the previous D-Wave 2000Q system and show an improved image classification process with a better overall quality. In this paper we discuss how to enhance often timeconsuming RBM training processes based on the commonly used Gibbs sampling using an improved version of quantum sampling. In order to prevent overfitting we propose some solutions which help to acquire less probable samples from the distribution by adjusting D-wave control and embedding parameters. Finally, we present various limitations of the existing quantum computing hardware and expected changes on the quantum hardware and software sides which can be adopted for further improvements in the field of machine learning.

Introduction

As indicated in various recent research, there may still be challenges in achieving acceptable performance using quantum computers for solving practical problems. Nevertheless, we demonstrate promising results by using the recent advent of the D-Wave Advantage quantum annealer to train and test a Restricted Boltzmann Machine for the well studied MNIST dataset. We compare our new model with some tests executed on the previous D-Wave 2000Q system and show an improved image classification process with a better overall quality. In this paper we discuss how to enhance often timeconsuming RBM training processes based on the commonly used Gibbs sampling using an improved version of quantum sampling. In order to prevent overfitting we propose some solutions which help to acquire less probable samples from the distribution by adjusting D-wave control and embedding parameters. Finally, we present various limitations of the existing quantum computing hardware and expected changes on the quantum hardware and software sides which can be adopted for further improvements in the field of machine learning.