Image Recognition by Quantum Annealing Using Multi-bit Spin Variables

Abstract

This paper presents an image-recognition technique using quantum annealing. We used four handwritten numbers of 0, 1, 2, and 3 from the MNIST (Modified National Institute of Standards and Technology) digit classification dataset as images to be recognized. We used two models in this study which were basically the same as a simple neural network, one had one fully connected layer and the other had 3x3 Sobel filters in front of that layer. The images were first resized to 1/2 the original in both directions, and pixel data were converted to 196 one-dimension data then output to four outputs through a fully coupled neural network. The optimum values of 196 x 4 weights and 4 bias values of the fully connected layer were obtained by quantum annealing in which the loss function is expressed as Hamiltonian of the Ising model. The weights and biases are multi-valued variables consisting of multiple spin variables in the Ising model. Quantum annealing was simulated using Pyqubo to create Ising models from flexible mathematical expressions. Prediction accuracy when using the Sobel filters reached about 95% and Sobel filter improved it by about 5%.