Abstract
In the noisy intermediate-scale quantum (NISQ) era, the capabilities of variational quantum algorithms are greatly constrained due to a limited number of qubits and the shallow depth of quantum circuits. We may view these variational quantum algorithms as weak learners in supervised learning. Ensemble methods are general approaches to combining weak learners to construct a strong one in machine learning. In this paper, by focusing on classification, we theoretically establish and numerically verify a learning guarantee for quantum adaptive boosting (AdaBoost). The supervised-learning risk bound describes how the prediction error of quantum AdaBoost on binary classification decreases as the number of boosting rounds and sample size increase. We further empirically demonstrate the advantages of quantum AdaBoost by focusing on a 4-class classification. The quantum AdaBoost not only outperforms several other ensemble methods, but in the presence of noise it can also surpass the ideally noiseless but unboosted primitive classifier after only a few boosting rounds. Our work indicates that in the current NISQ era, introducing appropriate ensemble methods is particularly valuable in improving the performance of quantum machine learning algorithms.
Introduction: The NISQ Bottleneck
We are currently in the Noisy Intermediate-Scale Quantum (NISQ) era, where quantum computers are limited by low qubit counts, shallow circuit depths, and significant hardware noise. These constraints often restrict variational quantum classifiers to acting as "weak learners"—models that perform only slightly better than random guessing. Furthermore, training these models is frequently plagued by "barren plateaus" and local minima. This research addresses a critical question: Can we use classical ensemble techniques—specifically AdaBoost—to transform these weak, noisy quantum circuits into a single, robust, and highly accurate classifier?
Methodology: Quantum Adaptive Boosting
The authors adapt the classical Adaptive Boosting (AdaBoost) algorithm for the quantum domain through an iterative three-step process:
- Sequential Training: A series of weak quantum classifiers, implemented as Parameterized Quantum Circuits (PQCs), are trained one after another.
- Adaptive Weighting: In each round, the algorithm increases the weight of data points misclassified in previous rounds, forcing the next quantum circuit to focus on these "hard" examples.
- Weighted Combination: Finally, all weak learners are combined into a strong ensemble classifier, where each learner’s vote is weighted by its individual accuracy.
Theoretical Learning Guarantees
A core contribution of this work is the derivation of a supervised learning risk bound for binary classification. Using Rademacher complexity, the authors prove that the prediction error decreases as the number of boosting rounds ($T$) and the sample size ($n$) increase. They establish that the complexity of the quantum hypothesis set is bounded by $O(\sqrt{K \log K / n})$, where $K$ is the number of trainable gates. This reveals a fundamental tradeoff: increasing the number of gates makes the model more expressive (lowering training error) but also increases the risk of overfitting (raising generalization error).
Key Results and Empirical Validation
The researchers evaluated their approach using physics data (the ANNNI model for identifying quantum phases) and image data (a subset of the MNIST handwritten digits dataset).
Exponential Convergence
For binary classification, the training error was shown to decrease exponentially with the number of boosting rounds. As illustrated in Figure 5, both the training error and its theoretical upper bound drop sharply as more rounds are added.
Surpassing the "Ideal" Classifier
In 4-class MNIST tests, Quantum AdaBoost demonstrated a remarkable ability to mitigate noise. Even when running on a noisy simulator (subject to depolarizing, amplitude damping, and phase damping noise), the ensemble outperformed an ideally noiseless single QCNN classifier after only 3 to 7 rounds of boosting.
As shown in Figure 6, the ensemble accuracy quickly surpasses the benchmark set by the best individual QCNN. Quantum AdaBoost achieved a peak test accuracy of 97.5%, significantly higher than Quantum Bagging (88.4%) and even outperforming classical CNN-based ensembles despite the classical models having six times more parameters.
Conclusion
This work formalizes the idea that variational quantum algorithms should be treated as weak learners and provides a mathematically sound way to bridge the gap between current hardware limitations and high-accuracy requirements. By focusing on samples that previous noisy circuits missed, the ensemble effectively "averages out" hardware noise, acting as a powerful form of error mitigation for the NISQ era.
Summary updated