A quantum deep convolutional neural network for image recognition

TLDR

Deep learning has achieved unprecedented success across many fields, yet its high memory and time demands remain a challenge, while quantum computing’s superposition and entanglement offer a promising avenue to address these limitations. This study investigates a quantum deep convolutional neural network (QDCNN) based on parameterized quantum circuits for image recognition. The QDCNN consists of successive quantum convolutional layers followed by a quantum classification layer, trained via a quantum–classical hybrid scheme inspired by variational quantum algorithms. Analysis shows the QDCNN achieves exponential speed‑up over classical counterparts, and simulations on MNIST and GTSRB confirm its feasibility and validity.

Abstract

Deep learning achieves unprecedented success involves many fields, whereas the high requirement of memory and time efficiency tolerance have been the intractable challenges for a long time. On the other hand, quantum computing shows its superiorities in some computation problems owing to its intrinsic properties of superposition and entanglement, which may provide a new path to settle these issues. In this paper, a quantum deep convolutional neural network (QDCNN) model based on the quantum parameterized circuit for image recognition is investigated. In analogy to the classical deep convolutional neural network (DCNN), the architecture that a sequence of quantum convolutional layers followed by a quantum classified layer is illustrated. Inspired by the variational quantum algorithms, a quantum–classical hybrid training scheme is demonstrated for the parameter updating in the QDCNN. The network complexity analysis indicates the proposed model provides the exponential acceleration comparing with the classical counterpart. Furthermore, the MNIST and GTSRB datasets are employed to numerical simulation and the quantitative experimental results verify the feasibility and validity.