Hybrid Quantum Downsampling Networks

Yifeng Peng, Xinyi Li, Zhiding Liang, Ying WangarXiv 2024 · DOI · arXiv

Abstract

Classical max pooling plays a crucial role in reducing data dimensionality among various well-known deep learning models, yet it often leads to the loss of vital information. We proposed a novel hybrid quantum downsampling module (HQD), which is a noise-resilient algorithm. By integrating a substantial number of quantum bits (qubits), our approach ensures the key characteristics of the original image are maximally preserved within the local receptive field. Moreover, HQD provides unique advantages in the context of the noisy intermediate-scale quantum (NISQ) era. We introduce a unique quantum variational circuit in our design, utilizing rotating gates including RX, RY, RZ gates, and the controlled-NOT (CNOT) gate to explore nonlinear characteristics. The results indicate that the network architectures incorporating the HQD module significantly outperform the classical structures with max pooling in CIFAR-10 and CIFAR-100 datasets. The accuracy of all tested models improved by an average of approximately 3%, with a maximum fluctuation of only 0.4% under various quantum noise conditions.

Introduction

In classical computer vision, max pooling is the standard technique for spatial dimension reduction. While computationally efficient, it is inherently lossy, discarding all but the maximum value in a local window and potentially deleting vital structural information. As we transition into the Noisy Intermediate-Scale Quantum (NISQ) era, there is a growing interest in hybrid quantum-classical architectures. The challenge lies in designing quantum components that offer performance advantages while remaining robust against the high error rates characteristic of current quantum hardware. This paper introduces the Hybrid Quantum Downsampling (HQD) module as a drop-in replacement for classical pooling layers.

Methodology

The HQD module leverages a Quantum Variational Circuit (QVC) to process local image features. The process involves four primary stages:

  1. Encoding: Classical pixel values from a local receptive field (e.g., 3x3) are mapped to rotation angles for quantum gates (RX and RZ), embedding the data into a high-dimensional quantum state.
  2. Processing: A series of trainable RY gates and CNOT gates create entanglement between qubits, allowing the network to capture complex, nonlinear relationships between pixels.
  3. Measurement: A Pauli-Z measurement collapses the quantum state, yielding a classical value for the downsampled output.
  4. Hybrid Strategy: A mixing parameter ($\ell$) balances the contribution of classical max pooling and quantum downsampling, optimizing the trade-off between efficiency and feature richness.
Architecture of the Hybrid Quantum Downsampling (HQD) module.

The quantum parameters are optimized end-to-end with classical weights using the parameter-shift rule for gradient estimation.

Key Insights and Results

The HQD module was evaluated using standard backbones like ResNet and Xception on CIFAR-10 and CIFAR-100 datasets.

Performance and Efficiency

The integration of HQD led to an average top-1 accuracy improvement of approximately 3%. For instance, a ResNet-18 backbone on CIFAR-10 improved from 75.50% to 78.36%. Remarkably, this gain is achieved with minimal overhead; a 3x3 HQD kernel adds only 9 additional trainable parameters.

Information Preservation

Qualitative analysis reveals that HQD preserves significantly more structural detail than classical max pooling. While max pooling tends to flatten textures and edges, HQD retains these critical features, as highlighted in comparative visualizations where red boxes indicate preserved structural elements.

Qualitative comparison between original images, HQD features, and classical max pooling.

Noise Resilience

A standout feature of the HQD module is its inherent resilience to quantum noise. Under simulated conditions involving bit-flips and phase-flips, model accuracy fluctuated by only 0.4%. This suggests that the hybrid training process allows the network to adapt to and "train through" the noise inherent in NISQ devices.

Conclusion

The Hybrid Quantum Downsampling module demonstrates a clear performance advantage over classical max pooling by better preserving image features and offering robustness to quantum noise. While current limitations include the computational overhead of classical simulation and memory constraints, HQD represents a significant step toward practical hybrid quantum-classical deep learning. Future research will focus on scaling these modules for high-resolution imagery and testing on physical quantum processors.

Summary updated

Abstract

Classical max pooling plays a crucial role in reducing data dimensionality among various well-known deep learning models, yet it often leads to the loss of vital information. We proposed a novel hybrid quantum downsampling module (HQD), which is a noise-resilient algorithm. By integrating a substantial number of quantum bits (qubits), our approach ensures the key characteristics of the original image are maximally preserved within the local receptive field. Moreover, HQD provides unique advantages in the context of the noisy intermediate-scale quantum (NISQ) era. We introduce a unique quantum variational circuit in our design, utilizing rotating gates including RX, RY, RZ gates, and the controlled-NOT (CNOT) gate to explore nonlinear characteristics. The results indicate that the network architectures incorporating the HQD module significantly outperform the classical structures with max pooling in CIFAR-10 and CIFAR-100 datasets. The accuracy of all tested models improved by an average of approximately 3%, with a maximum fluctuation of only 0.4% under various quantum noise conditions.
BibTeX
@misc{peng2024hybrid,
  author = {Yifeng Peng and Xinyi Li and Zhiding Liang and Ying Wang},
  title = {{Hybrid Quantum Downsampling Networks}},
  year = {2024},
  eprint = {2405.16375},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2405.16375},
}
Figure