Exploring the features of quanvolutional neural networks for improved image classification

Abstract

Image classification has an important role in many machine learning applications. Numerous classification techniques based on quantum machine learning have been reported recently. In this article, we investigate the features of the quanvolutional neural network—a hybrid quantum-classical image classification technique, which is inspired by the convolutional neural network and has the potential to outperform current image processing techniques. We improve the training strategy and evaluate the classification tasks on three traditional public datasets in terms of different topologies, sizes, and depths of filters. Finally, we propose four efficient configurations for the quanvolutional neural network to improve image classification.

Appendix

As mentioned in Section 2.2, the quantum compiling technique will consume large resources when dealing with a very large dataset. Thus, we can convert patches of each $\{X_j\}$ in dataset $\mathcal {X}$ based on filter size f, store it, and then use it later:

$$\begin{aligned} \{|\psi (x_{i,j})\rangle \}=\{U_{\text {QC}, f}(x_{i,j}, \varvec{\theta }^{QC}_{i,j})|\varvec{0}\rangle \} \end{aligned}$$

where $\varvec{\theta }^{QC}_{i,j}$ is the optimal parameter in the quantum compiling process and the depth of the encoder is the depth $U_{\text {QC}, f}$. Then, this quantum dataset can be used in various problems.