[Paper Review] Scalable Quantum Convolutional Neural Networks

Hankyul Baek, Won Joon Yun+1 more|arXiv (Cornell University)|Sep 26, 2022

Quantum Computing Algorithms and Architecture4 citations

TL;DR

This paper proposes a scalable quantum convolutional neural network (sQCNN) that overcomes barren plateaus in quantum neural networks by using multiple finite-sized filters with a novel reverse fidelity training (RF-Train) algorithm. By leveraging quantum state fidelity to regularize feature diversity, sQCNN achieves 16% higher top-1 accuracy than vanilla-trained QCNN on MNIST and FMNIST, demonstrating improved scalability and performance with increasing filters.

ABSTRACT

With the beginning of the noisy intermediate-scale quantum (NISQ) era, quantum neural network (QNN) has recently emerged as a solution for the problems that classical neural networks cannot solve. Moreover, QCNN is attracting attention as the next generation of QNN because it can process high-dimensional vector input. However, due to the nature of quantum computing, it is difficult for the classical QCNN to extract a sufficient number of features. Motivated by this, we propose a new version of QCNN, named scalable quantum convolutional neural network (sQCNN). In addition, using the fidelity of QC, we propose an sQCNN training algorithm named reverse fidelity training (RF-Train) that maximizes the performance of sQCNN.

Motivation & Objective

• To address the barren plateau problem in quantum convolutional neural networks (QCNNs) when scaling qubit count per filter.
• To enable scalable feature extraction in QCNNs without increasing qubit count per filter by using multiple finite-sized filters.
• To develop a training algorithm that enhances feature diversity and model performance using quantum fidelity as a regularizer.
• To empirically validate the scalability and performance superiority of sQCNN over conventional QCNNs on standard vision datasets.

Proposed method

• Proposes sQCNN, a scalable QCNN architecture using multiple finite-sized quantum filters instead of increasing qubit count per filter to avoid barren plateaus.
• Designs a reverse fidelity (RF) regularizer based on the fidelity between output states of different filters to maximize feature diversity.
• Integrates the RF regularizer into the total loss function as $\mathcal{L}_{total} = \mathcal{L}_{CE} + \lambda \mathcal{L}_{RF}$, where $\mathcal{L}_{RF}$ minimizes similarity between filter outputs.
• Employs cross-entropy loss $\mathcal{L}_{CE}$ for classification and combines it with the RF regularizer to jointly optimize accuracy and feature diversity.
• Uses a minibatch-based training procedure with adjustable $\lambda$ to control the strength of the fidelity regularization.
• Applies the training algorithm to MNIST and FMNIST datasets to evaluate performance and scalability.

Experimental results

Research questions

• RQ1 Can a QCNN architecture be made scalable without increasing the number of qubits per filter, thereby avoiding barren plateaus?
• RQ2 Does using quantum state fidelity as a regularizer improve feature diversity and model performance in quantum convolutional networks?
• RQ3 How does the performance of sQCNN compare to vanilla-trained QCNNs in terms of top-1 accuracy on standard vision benchmarks?
• RQ4 Does increasing the number of filters in sQCNN lead to consistent performance gains, unlike in conventional QCNNs?

Key findings

• sQCNN with RF-Train achieves 16% higher top-1 accuracy than sQCNN without RF-Train ($\lambda = 0$) on the MNIST dataset.
• The RF-Train method increases the Euclidean distance between extracted features, confirming enhanced feature diversity.
• sQCNN maintains stable performance as the number of filters increases, demonstrating scalability.
• In contrast to QCNN, which suffers a 30% performance drop when qubit count per filter increases from 8 to 16, sQCNN shows consistent improvement with more filters.
• The RF regularizer effectively reduces similarity between filter outputs, as evidenced by decreasing fidelity between output states.
• sQCNN outperforms vanilla-trained QCNN across both MNIST and FMNIST datasets, especially at higher $\lambda$ values.