A qubit as a Kernel: an efficient quantum-classical network for image classification with Quantum Independent Convolution-like Kernel Layer

doi:10.1007/s11227-026-08360-5

A qubit as a Kernel: an efficient quantum-classical network for image classification with Quantum Independent Convolution-like Kernel Layer

2026-03-08

0

PRE

AI

DOI：10.1007/s11227-026-08360-5

Original

Original request for help

Save

Abstract

En 中文

To address the limitations of the existing quantum convolution algorithm in feature extraction and scalability, this paper proposes a Quantum Independent Convolution-like Kernel Layer (QICKL), which mirrors classical convolution by encoding a kernel’s weighted sum into a qubit rotation angle. We evaluate the effectiveness of QICKL using a hybrid network constructed with QICKL and a classical fully connected layer. Result on the MNIST/FashionMNIST dataset shows that our network outperforms existing quantum-classical convolution networks while using same or fewer qubits, demonstrating its efficiency and scalability. Comparing other embedding methods, QICKL demonstrates advantage in convergence speed and accuracy. This work provides a new design perspective for resource-efficient quantum algorithms.

Keywords:

Quantum machine learning

Hybrid quantum-classical architecture

Quantum convolutional layer

Data embedding

Image classification

A qubit as a Kernel: an efficient quantum-classical network for image classification with Quantum Independent Convolution-like Kernel Layer

Abstract

Journal

Organization