Abstract:Deep Learning has shown remarkable capabilities in pattern recognition, feature extraction, and classification decision-making. With the rise of quantum computing, the potential of quantum neural networks (QNNs) in Artificial Intelligence is emerging. However, the intrinsic mechanisms and decision transparency of QNNs remain unclear. In this paper, we propose a classic-quantum hybrid network framework (CQH-Net), which uses traditional machine learning methods for feature extraction and quantizes neural networks for classification tasks. We apply CQH-Net to image classification on public datasets. Experimentally, CQH-Net achieves an average convergence rate improvement of 72.8% compared to classical convolutional networks (CNNs) with identical parameters. On the Fashion MNIST dataset, it reaches a final accuracy of 99.02%, representing a significant increase of 5.07% over CNNs. Furthermore, we explore visual explanations for CQH-Net's decision-making process. Results show that the model effectively captures key data features during training and establishes associations between these features and their corresponding categories. This study demonstrates that quantization enhances the models ability to tackle complex classification problems while providing transparency in its decision-making process further supporting quantum advantages in machine learning.

What problem does this paper attempt to address?

The problems that this paper attempts to solve mainly focus on the following aspects: 1. **Improving the performance of classification tasks**: By proposing a classical - quantum hybrid network framework (CQH - Net), it aims to combine the advantages of traditional machine learning methods in feature extraction and the potential of quantum neural networks (QNNs) in classification tasks, so as to improve the performance of image classification tasks, especially in terms of convergence speed and final classification accuracy, surpassing traditional convolutional neural networks (CNNs). 2. **Enhancing decision - making transparency**: The paper also explores how to explain the decision - making process of CQH - Net through visualization techniques (such as Grad - CAM), thereby providing a deeper understanding of the internal working principles of the model and enhancing the transparency of its decision - making. This is crucial for understanding and optimizing the behavior of the model when dealing with complex data patterns. ### Specific problem description - **Performance improvement**: - **Convergence speed**: The experimental results show that the average convergence speed of CQH - Net on multiple public datasets is 72.8% faster than that of the classical convolutional network with the same parameters. - **Classification accuracy**: On the Fashion MNIST dataset, CQH - Net achieves a final classification accuracy of 99.02%, which is 5.07% higher than that of CNNs. - **Decision - making transparency**: - **Visual explanation**: Heat maps are generated through the Grad - CAM technique to show how CQH - Net gradually learns and identifies key features in the image during the training process. These heat maps not only reveal the changes in the focus of the model in the early and late stages of training, but also verify that the model can effectively capture and utilize the key areas of the input image for accurate prediction. ### Method overview - **Classical - quantum hybrid network structure**: - **Feature extraction module**: Two convolutional layers and two pooling layers are used to extract initial features, and non - linearity is introduced through the ReLU activation function. - **Dimensionality reduction module**: The feature dimension is reduced from 256 to 4 through Dropout and Flatten operations to meet the requirements of quantum computing. - **Quantum encoding module**: The extracted features are converted into quantum states and processed through the quantum encoding layer (including H - gate and U1 - gate). - **Variational quantum circuit (VQC)**: A VQC with 4 qubits and 8 parameters (4 of which are adjustable) is used for classification tasks. - **Prediction module**: The final classification result is generated by measuring quantum data. - **Experimental setup**: - The MNIST and Fashion MNIST datasets are used for experiments to compare the performance of CQH - Net with that of classical CNNs with different numbers of parameters. - The same optimizer (Adam), number of training epochs (50), batch size (64) and learning rate (0.001) are used in the experiments. ### Conclusion By proposing the CQH - Net framework, the paper successfully combines the advantages of classical and quantum computing, achieving significant performance improvement and enhanced decision - making transparency in image classification tasks. These results provide strong support for further exploring the application of quantum computing in deep learning in the future.

A Classic-Quantum Hybrid Network Framework: CQH-Net

Image Classification Using Hybrid Classical-Quantum Neutral Networks

Quantum convolutional neural network for classical data classification

Hybrid quantum neural network structures for image multi-classification

A Quantum Convolutional Neural Network for Image Classification

A quantum deep convolutional neural network for image recognition

A hybrid quantum–classical neural network with deep residual learning

QuClassi: A Hybrid Deep Neural Network Architecture based on Quantum State Fidelity

QFCNN: Quantum Fourier Convolutional Neural Network

Quantum-inspired activation functions in the convolutional neural network

Accurate Image Multi-Class Classification Neural Network Model with Quantum Entanglement Approach

Quantum-inspired activation functions and quantum Chebyshev-polynomial network

Quantum convolutional neural networks for jet images classification

DeepQMLP: A Scalable Quantum-Classical Hybrid DeepNeural Network Architecture for Classification

Parallel Structure of Hybrid Quantum-Classical Neural Networks for Image Classification

The Design of a Novel Hybrid Quantum Deep Neural Network in INEQR Images Classification

Shallow hybrid quantum-classical convolutional neural network model for image classification

Quantum Fourier Convolutional Network

Classical-to-quantum convolutional neural network transfer learning

A Scalable Quantum Non-local Neural Network for Image Classification