Quantum Split Neural Network Learning using Cross-Channel Pooling
This work addresses privacy and efficiency challenges in quantum machine learning for researchers and practitioners, though it appears incremental as an extension of classical split learning to quantum settings.
The paper tackles the problem of improving quantum federated learning by proposing quantum split learning with cross-channel pooling, achieving a 1.64% higher top-1 accuracy on MNIST classification compared to quantum federated learning while maintaining robust privacy.
In recent years, the field of quantum science has attracted significant interest across various disciplines, including quantum machine learning, quantum communication, and quantum computing. Among these emerging areas, quantum federated learning (QFL) has gained particular attention due to the integration of quantum neural networks (QNNs) with traditional federated learning (FL) techniques. In this study, a novel approach entitled quantum split learning (QSL) is presented, which represents an advanced extension of classical split learning. Previous research in classical computing has demonstrated numerous advantages of split learning, such as accelerated convergence, reduced communication costs, and enhanced privacy protection. To maximize the potential of QSL, cross-channel pooling is introduced, a technique that capitalizes on the distinctive properties of quantum state tomography facilitated by QNNs. Through rigorous numerical analysis, evidence is provided that QSL not only achieves a 1.64\% higher top-1 accuracy compared to QFL but also demonstrates robust privacy preservation in the context of the MNIST classification task.