[Paper Review] Quantum-Enhanced Machine Learning for Covid-19 and Anderson Insulator Predictions

Paul-Aymeric McRae, Michael Hilke|arXiv (Cornell University)|Dec 7, 2020

Quantum Computing Algorithms and Architecture4 citations

TL;DR

This paper evaluates quantum-enhanced machine learning using a 16-qubit IBM quantum computer, applying Quantum Support Vector Machines (QSVM) and Variational Quantum Classifiers (VQC) to classify COVID-19 spread data and Anderson insulator transitions. Despite theoretical promise, quantum enhancement fails to outperform classical methods on complex datasets, with quantum models often underperforming or exhibiting artificial accuracy due to class imbalance and limited hyperparameter tuning.

ABSTRACT

Quantum Machine Learning (QML) algorithms to solve classifications problems have been made available thanks to recent advancements in quantum computation. While the number of qubits are still relatively small, they have been used for "quantum enhancement" of machine learning. An important question is related to the efficacy of such protocols. We evaluate this efficacy using common baseline data sets, in addition to recent coronavirus spread data as well as the quantum metal-insulator transition in three dimensions. For the computation, we used the 16 qubit IBM quantum computer. We find that the "quantum enhancement" is not generic and fails for more complex machine learning tasks.

Motivation & Objective

• To assess the practical efficacy of quantum-enhanced machine learning for real-world classification tasks.
• To compare quantum-classical performance on complex, non-baseline datasets like COVID-19 spread and 3D Anderson localization.
• To investigate whether 'quantum enhancement' in QSVM and VQC algorithms delivers measurable advantages over classical counterparts.
• To identify implementation challenges such as artificial accuracy from class imbalance and limited hyperparameter control in quantum frameworks.
• To evaluate the scalability and computational complexity of quantum machine learning pipelines on near-term quantum hardware.

Proposed method

• Employs Quantum Support Vector Machines (QSVM) and Variational Quantum Classifiers (VQC) via IBM QISKit on a 16-qubit quantum processor.
• Uses classical SVMs (linear and RBF kernels) as baselines for comparison on the same datasets.
• Applies feature mapping and kernel functions in QSVM, with quantum circuits handling the kernel computation.
• Utilizes variational quantum circuits with parameterized gates, iteratively optimized via classical feedback loops.
• Employs simulated quantum circuits and real IBM quantum hardware to evaluate performance across data size and feature dimension.
• Measures classification accuracy, computation time, and scalability trends across training sample count and feature dimensions.

Experimental results

Research questions

• RQ1 Does quantum-enhanced machine learning outperform classical machine learning on complex, real-world datasets like COVID-19 and Anderson insulator data?
• RQ2 To what extent do QSVM and VQC algorithms exhibit 'quantum enhancement' in practice compared to classical SVMs?
• RQ3 What are the primary performance bottlenecks in current quantum machine learning pipelines, particularly regarding scalability and hyperparameter tuning?
• RQ4 Why do some quantum models achieve artificially high accuracy, especially in imbalanced datasets?
• RQ5 How do computation times scale with training data size and feature dimension in quantum vs. classical classification?

Key findings

• Quantum-enhanced methods did not outperform classical SVMs on the Breast Cancer and Wine benchmark datasets, with classical models consistently achieving higher accuracy.
• On the more complex COVID-19 and Anderson insulator datasets, quantum models generally underperformed classical counterparts, with no consistent quantum advantage observed.
• In rare cases, quantum models showed slightly better accuracy, but only when data included randomness, suggesting no robust improvement.
• Several quantum models exhibited catastrophic performance by predicting only one class, especially in imbalanced datasets, due to limited hyperparameter tuning in QISKit.
• QSVM computation time scaled quadratically with data size, while VQC showed linear dependence on data size but exponential scaling with feature dimension.
• The observed performance gap between theory and practice is attributed to simulation overhead and implementation differences, not just theoretical complexity.