Classical Ensembles of Single-Qubit Quantum Variational Circuits for Classification

TL;DR

This paper explores classical ensemble methods applied to single-qubit quantum classifiers, demonstrating improved accuracy across multiple datasets and highlighting their potential for NISQ device applications.

Contribution

It introduces and evaluates classical bagging and boosting ensembles of single-qubit quantum classifiers, showing their effectiveness beyond problem-specific cases.

Findings

Ensembles improve classification accuracy on pulsar star data.

Boosting achieves high accuracy with limited training.

Bagging yields higher accuracy with more training time.

Abstract

The quantum asymptotically universal multi-feature (QAUM) encoding architecture was recently introduced and showed improved expressivity and performance in classifying pulsar stars. The circuit uses generalized trainable layers of parameterized single-qubit rotation gates and single-qubit feature encoding gates. Although the improvement in classification accuracy is promising, the single-qubit nature of this architecture, combined with the circuit depth required for accuracy, limits its applications on NISQ devices due to their low coherence times. This work reports on the design, implementation, and evaluation of ensembles of single-qubit QAUM classifiers using classical bagging and boosting techniques. We demonstrate an improvement in validation accuracy for pulsar star classification. We find that this improvement is not problem-specific as we observe consistent improvements for the MNIST Digits and Wisconsin Cancer datasets. We also observe that the boosting ensemble achieves an acceptable level of accuracy with only a small amount of training, while the bagging ensemble achieves higher overall accuracy with ample training time. This shows that classical ensembles of single-qubit circuits present a new approach for certain classification problems.