Lean classical-quantum hybrid neural network model for image classification

TL;DR

The paper addresses the scalability challenges of quantum machine learning for image classification by proposing a lean classical-quantum hybrid neural network (LCQHNN) that uses a classical CNN front end for feature extraction and a compact variational quantum circuit for classification. The four-layer VQC, coupled with a small subsequent classical readout, reduces quantum resource demands while maintaining strong performance. Empirical results on MNIST, FashionMNIST, and CIFAR-10 show high accuracies (MNIST: 100%, FashionMNIST: 99.02%, CIFAR-10: 85.55%) and faster convergence than equivalent-parameter classical CNNs, with Grad-CAM visualizations confirming meaningful feature focus during training. The work demonstrates a practical path toward quantum-enhanced image classification under resource constraints, while noting limitations related to dataset scale and reliance on simulated hardware, and outlining future directions for multi-class tasks and real quantum hardware experiments.

Abstract

The integration of algorithms from quantum information with neural networks has enabled unprecedented advancements in various domains. Nonetheless, the application of quantum machine learning algorithms for image classification predominantly relies on traditional architectures such as variational quantum circuits. The performance of these models is closely tied to the scale of their parameters, with the substantial demand for parameters potentially leading to limitations in computational resources and a significant increase in computation time. In this paper, we introduce a Lean Classical-Quantum Hybrid Neural Network (LCQHNN), which achieves efficient classification performance with only four layers of variational circuits, thereby substantially reducing computational costs. Our experiments demonstrate that LCQHNN achieves 100\%, 99.02\%, and 85.55\% classification accuracy on MNIST, FashionMNIST, and CIFAR-10 datasets. Under the same parameter conditions, the convergence speed of this method is also faster than that of traditional models. Furthermore, through visualization studies, it is found that the model effectively captures key data features during training and establishes a clear association between these features and their corresponding categories. This study confirms that the employment of quantum algorithms enhances the model's ability to handle complex classification problems.

Figures (8)

• Figure 1: Lean classical-quantum hybrid neural network structure for studying image classification. Firstly, feature extraction is performed using classical neural networks, and then the extracted features are passed on to VQCs for classification tasks. Finally, the quantum data is converted back to classical data through the measurement layer to obtain the final classification result.
• Figure 2: Variational Quantum Circuit Structure. The encoding layer of this network introduces H gates and U1 gates to each qubit, enabling adjustment and manipulation of individual qubit states by selecting appropriate rotation angles for the U1 gates. The research model integrates CNOT gates, a crucial multi-qubit operator that facilitates non-local state transfer by controlling interactions between the control and target qubits.
• Figure 3: Bloch sphere representation of encoded quantum states. (a) After the input qubit undergoes the encoding circuit with H and U1 gates, it becomes uniformly distributed on the equatorial plane of the Bloch sphere, at this point, the qubit is in a superposition state of $|0\rangle$ and $|1\rangle$. (b) The image displays the distribution of the quantum state after rotation around the y-axis. The parameter of the RY gate determines the final position of the quantum state on the Bloch sphere.
• Figure 4: Experimental results on the MNIST dataset. (a) The comparison of the LCQHNN against traditional CNNs with varying numbers of parameters in terms of accuracy evolution during the training process was conducted following the unified setup described in the paper. (b) The accuracy of different models at the 10th epoch during training.
• Figure 5: Experimental results on the FashionMNIST dataset. (a) Comparison of the accuracy change of LCQHNN training on the FashionMNIST dataset with traditional CNNs with different numbers of parameters. (b) The accuracy of each model on the FashionMNIST dataset after training for 10 epochs.