Resource-efficient Variational Quantum Classifier Enables Near-Deterministic Predictions with Reduced Circuit Executions

Quantum computing holds immense promise for transforming machine learning, particularly in complex classification tasks, but realising this potential demands solutions to inherent challenges in quantum systems. Petr Ptáček, Paulina Lewandowska, and Ryszard Kukulski, from VSB-Technical University of Ostrava and IT4Innovations, address a critical limitation arising from the probabilistic nature of quantum predictions, which typically requires numerous repeated calculations. Their research introduces a new measurement strategy for quantum classifiers that delivers near-certain predictions, markedly reducing the computational resources needed while maintaining strong classification accuracy even in the presence of noise. This achievement represents a significant step towards practical, resource-efficient quantum machine learning, offering a favourable balance between performance and computational cost.

Reducing Measurements in Variational Quantum Classifiers

Scientists have developed a new method for improving the efficiency and reliability of variational quantum classifiers, a promising area of quantum computing. The research focuses on reducing the number of measurements, or executions of the quantum circuit, needed to make a prediction. By comparing their approach to standard methods, the team demonstrates that comparable or better performance can be achieved with significantly fewer circuit executions. Performance was evaluated across three datasets, revealing variations in effectiveness depending on the data characteristics. The experiments also explored the impact of noise, demonstrating the method’s resilience to imperfections in the quantum circuit. This work addresses a key challenge in quantum machine learning, where the computational cost of repeated measurements can limit practical implementation. The team’s approach offers a pathway towards more efficient and scalable quantum classifiers, potentially enabling their use on near-term quantum hardware.

Deterministic Classification with an Abstain Option

Scientists have engineered a novel measurement strategy to overcome limitations in variational quantum classification. The study introduces an unambiguous quantum classifier for binary classification, designed to achieve near-deterministic predictions with reduced computational overhead. Recognizing that traditional quantum classifiers require repeated circuit executions due to the probabilistic nature of quantum measurements, researchers developed a system where the classifier yields one of three outcomes: class “0”, class “1”, or “I don’t know”. If the initial measurement results in “I don’t know”, the experiment is repeated, incrementing the number of measurement shots, until a definitive classification of 0 or 1 is achieved.

This strategy enables near-deterministic predictions while maintaining adequate classification accuracy, even in noisy environments, and significantly reduces the number of quantum circuit executions needed. Experiments employed data re-uploading techniques and careful selection of cost functions to optimize the performance of the unambiguous quantum classifier. Results demonstrate a slight decrease in average accuracy, representing a favorable trade-off for improved resource efficiency, indicating the proposed method is viable for use on real quantum hardware where noise is a significant challenge. This work establishes a pathway towards practical, resource-efficient quantum classification algorithms.

Near Deterministic Classification With Fewer Measurements

Scientists have developed a new measurement strategy for variational quantum classification that delivers near-deterministic predictions while significantly reducing the number of quantum circuit executions required. This breakthrough addresses a key limitation of quantum classifiers, which traditionally require repeated measurements due to the inherent probabilistic nature of quantum mechanics. The research team successfully defined an unambiguous quantum classifier, enabling clearer, more reliable classification outcomes. The work focuses on improving the decision stage of quantum classification, where repeated circuit executions are typically needed to achieve deterministic predictions.

Researchers demonstrated that their new strategy allows for near-deterministic classification, meaning the classifier consistently predicts the correct label with high confidence. This achievement represents a substantial step towards practical quantum machine learning, as it minimizes the computational overhead associated with the measurement process. Experiments reveal that this new measurement strategy maintains competitive classification accuracy even in noisy environments. While acknowledging that achieving truly single-shot, trainable quantum learning is challenging, the team’s work demonstrates a favorable trade-off between performance and resource efficiency. This research paves the way for more efficient and practical quantum machine learning algorithms.

Unambiguous Classification With Reduced Circuit Depth

This work introduces the unambiguous quantum classifier, a novel measurement strategy designed to address the computational overhead inherent in quantum classification tasks. By enforcing an acceptance threshold on measurement outcomes, the researchers achieve near-deterministic predictions while significantly reducing the number of required quantum circuit executions. Experimental results, obtained across three publicly available datasets, demonstrate an average accuracy of 76% in a noiseless, three-qubit setup, representing a modest performance difference of 6. 25% compared to existing state-of-the-art methods.

Notably, the unambiguous classifier achieves this performance with approximately 222times fewer circuit executions. Even under noisy conditions, the performance gap remains around 7%, with a comparable reduction in required executions, roughly 277times fewer. The team attributes this efficiency to the use of expectation-value-based measurements, which facilitate smoother cost functions and efficient gradient-based optimization, and to the incorporation of data re-uploading, which accelerates convergence during training. Comparisons with recent studies indicate that the developed models consistently achieve superior performance, demonstrating the potential of this approach for near-term quantum hardware where minimizing circuit execution cost is paramount. This research offers a promising solution to the challenges of resource-efficient quantum classification.