Scholarlink

<p>With the integration of quantum computing technology and classical machine learning (ML), a variety of quantum convolutional neural networks (QCNNs) have been proposed. However, they did not fully exploit the properties of Hilbert space when dealing with classical classification problems. In this work, we propose to encode classical real data into complex vector data based on the fact that Hilbert space is a complex inner product vector space. On this basis, we design two complex-encoded models for the two current mainstream pure QCNNs. The first model CE-QCNN is a fully quantum parameterized convolutional neural network model, and the other model CE-QC-CNN uses quantum operations to implement classical convolutional and pooling operations. Compared to the previous QCNN models, both of our two models use fewer qubits. In addition, we experimentally demonstrate the potential of two models for classification purposes on the MNIST dataset. To further illustrate the universality, we also use CE-QCNN to implement the malicious code recognition task on the CIC-MalDroid2020 dataset.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Complex-encoded quantum convolutional neural networks

  • 期刊论文

  • 发表日期：

    2025年8月8日

查看全文
  • Zhipeng Xiao,
  • Xiaoqing Tan,
  • Daipengwei Bao,
  • Rui Huang

摘要

With the integration of quantum computing technology and classical machine learning (ML), a variety of quantum convolutional neural networks (QCNNs) have been proposed. However, they did not fully exploit the properties of Hilbert space when dealing with classical classification problems. In this work, we propose to encode classical real data into complex vector data based on the fact that Hilbert space is a complex inner product vector space. On this basis, we design two complex-encoded models for the two current mainstream pure QCNNs. The first model CE-QCNN is a fully quantum parameterized convolutional neural network model, and the other model CE-QC-CNN uses quantum operations to implement classical convolutional and pooling operations. Compared to the previous QCNN models, both of our two models use fewer qubits. In addition, we experimentally demonstrate the potential of two models for classification purposes on the MNIST dataset. To further illustrate the universality, we also use CE-QCNN to implement the malicious code recognition task on the CIC-MalDroid2020 dataset.