Multichannel Quantum Data Preprocessing for Image Classification

The feasibility of applications of hybrid quantum–classical neural networks is investigated for image classification problems. The research is focused on the hybridization technique where quantum transformations serve as data augmentation operations before the input to the neural network. The described approach allows us to split the quantum and classical parts of the neural network (NN) into two separate steps of the process, which can be beneficial in certain scenarios. Several variations of a simple hybrid quantum–classical convolutional neural networks (HNNs) were built, where the quantum device is used as the first convolutional layer of a network. The corresponding classical convolutional neural networks (CNN) were used as reference models for comparing the results and performing the feasibility study on the CIFAR100 dataset. The best results were obtained for deep HNN with additional convolutional layers which outperformed less-deep CNN and significantly outperformed flat NN in terms of both accuracy and loss metrics. Also, some experiments indicated that the HNN training process requires fewer epochs than the training of classical reference models to reach higher accuracy values.