Multi-Scale Feature Fusion Quantum Depthwise Convolutional Neural Networks for Text Classification

TL;DR

This paper targets efficient text classification with quantum models by proposing MSFF-QDConv, a QCNN-based architecture that combines quantum depthwise convolution with multi-scale feature fusion. It introduces quantum word embedding and quantum sentence embedding to reduce embedding parameters and uses a quantum fully connected layer for classification, achieving a new RP state-of-the-art accuracy of 96.77% and showing strong parameter efficiency over classical counterparts. The method jointly captures local word-level and global sentence-level information, with ablation studies validating the contributions of depthwise convolution and fusion. The work demonstrates practical potential for quantum NLP on near-term devices and highlights directions for richer quantum designs and hardware experiments.

Abstract

In recent years, with the development of quantum machine learning, quantum neural networks (QNNs) have gained increasing attention in the field of natural language processing (NLP) and have achieved a series of promising results. However, most existing QNN models focus on the architectures of quantum recurrent neural network (QRNN) and self-attention mechanism (QSAM). In this work, we propose a novel QNN model based on quantum convolution. We develop the quantum depthwise convolution that significantly reduces the number of parameters and lowers computational complexity. We also introduce the multi-scale feature fusion mechanism to enhance model performance by integrating word-level and sentence-level features. Additionally, we propose the quantum word embedding and quantum sentence embedding, which provide embedding vectors more efficiently. Through experiments on two benchmark text classification datasets, we demonstrate our model outperforms a wide range of state-of-the-art QNN models. Notably, our model achieves a new state-of-the-art test accuracy of 96.77% on the RP dataset. We also show the advantages of our quantum model over its classical counterparts in its ability to improve test accuracy using fewer parameters. Finally, an ablation test confirms the effectiveness of the multi-scale feature fusion mechanism and quantum depthwise convolution in enhancing model performance.

Figures (9)

• Figure 1: An example of a variational quantum circuit. Initially, the encoder $U_e(x)$ maps the classical input data $x$ onto a quantum state. Subsequently, this encoded quantum state is transformed by the ansatz $U_a(\theta)$ , where $\theta$ represent the adjustable parameters. Lastly, the decoder retrieves classical information from the resulting quantum state by performing quantum measurements.
• Figure 2: Quantum word embedding. $U_{AE}(x)$ represents amplitude encoding where $x$ is the input one-hot encoded vector. $U_{a}(\theta)$ denotes the ansatz circuit where $\theta$ are trainable parameters. Due to amplitude encoding, the size of the output word embedding is exponentially reduced compared to the input one-hot vector.
• Figure 3: Quantum sentence embedding. $U_{AE}(x)$ represents amplitude encoding where $x$ is the input TF-IDF vector. $U_{a}(\theta)$ denotes the ansatz circuit where $\theta$ are trainable parameters. Due to amplitude encoding, the size of the output word embedding is exponentially reduced compared to the input TF-IDF vector.
• Figure 4: An example of 1D quantum convolution with a kernel size of 3. The kernel is a variational quantum circuit. In the process of convolution, the input patches are first encoded onto quantum states through angle encoding. Then, an ansatz is used to perform a unitary transformation on these quantum states. Finally, the measurement results on the final quantum states are summed up as the output feature map.
• Figure 5: Comparison of two types of quantum convolutions. The input and output sequences both contain two channels. (a) is an example of standard quantum convolution. Two sets of quantum kernels are used to extract features. The first set contains VQC1 and VQC2, while the second set contains VQC3 and VQC4. In each set, one VQC is used to extract features from input channel 1, and the other one is used to extract features from input channel 2. The results from each set are then summed to produce one output channel. (b) is an example of quantum depthwise convolution. A single quantum kernel, namely VQC1, is used to separately extract features from input channels 1 and 2, producing output channel 1 and 2.