Quantum Measurement Achieves 83% Accuracy Without Entangling Gates

Binghamton University, in collaboration with Graduate School of Informatics and Engineering and University of Electro-Communications, present a new approach to quantum classification that prioritises minimising the use of computationally expensive two-qubit gates. Michael A. Magid and colleagues introduce Optimal Quantum Measurement Decoding (OQMD), a technique which optimises the mapping of quantum results to classical labels by training a readout layer alongside the quantum circuit, without increasing the number of CNOT gates. Experiments on the Iris benchmark reveal that a zero-CNOT configuration utilising OQMD achieves 83.33% accuracy. This result sharply surpasses both the best 18-CNOT controls (56.67%) and the best 18-CNOT configuration with OQMD (66.67%). The research highlights that circuit complexity, as measured solely by CNOT count, may not always correlate with performance, and suggests a pathway towards more efficient near-term quantum classifiers.

Optimal decoding boosts quantum classification beyond complex circuit thresholds

Scientists at Binghamton University, in collaboration with Graduate School of Informatics and Engineering and University of Electro-Communications, achieved a sharp leap in quantum classification accuracy. The best 0-CNOT configuration, utilising Optimal Quantum Measurement Decoding (OQMD), reached 83.33% accuracy, exceeding previous benchmarks. This result is particularly notable as it surpasses the 56.67% accuracy of the best 18-CNOT controls and the 66.67% achieved by the best 18-CNOT configuration with OQMD, all under a standardised protocol. The Iris dataset, commonly used for benchmarking machine learning algorithms, presents a classification challenge involving 150 samples with four features, representing measurements of sepal length, sepal width, petal length, and petal width for three different species of iris. Achieving high accuracy on this dataset demonstrates the potential of the OQMD technique for more complex classification tasks.

OQMD optimises the interpretation of quantum measurement outcomes as classical labels by training a readout layer jointly with the quantum circuit, without increasing the number of computationally expensive CNOT gates. This challenges the assumption that more complex circuits are always superior, suggesting that intelligent decoding can be a powerful alternative to circuit expansion. A configuration utilising nine CNOT gates reached 83.33% accuracy on the Iris dataset, a benchmark for quantum classification, according to further performance metrics detailed by the team. The conventional approach to improving quantum classifier performance often involves increasing the ‘entangling depth’ of the circuit, that is, adding more layers of CNOT gates. However, two-qubit gates are prone to errors and introduce latency in near-term quantum devices, making this approach increasingly problematic. OQMD offers a potential solution by focusing on optimising the final measurement and decoding process, rather than solely on circuit depth.

Trainable triple single-qubit rotations were employed to implement OQMD, optimising the translation of quantum results into understandable labels. These rotations act as parameters within the readout layer, allowing the model to learn the optimal mapping from quantum states to classical predictions. The training process involves adjusting these rotation angles to minimise the classification error on the training data. Statistical analysis using Mann-Whitney U tests revealed a small pooled mean shift of approximately 0.03, indicating consistent gains across different quantum architectures, although these gains were not uniformly large in practical effect. This suggests that OQMD’s benefits are relatively robust and not heavily dependent on the specific hardware implementation. Despite these improved accuracy levels, the discrete and non-Gaussian nature of the data means that practical application to more complex, real-world datasets remains a considerable challenge. The Iris dataset, while useful for initial validation, is a relatively simple dataset, and the performance of OQMD on more complex and high-dimensional data requires further investigation.

Optimal data decoding boosts quantum classification accuracy without circuit expansion

Researchers from Binghamton University, in collaboration with Graduate School of Informatics and Engineering and University of Electro-Communications have demonstrated a pathway to improved accuracy in quantum classification without necessarily increasing the complexity of the quantum circuit itself. Intelligent interpretation of quantum results can be equally effective as adding more two-qubit gates to enhance performance. The team’s findings are significant because they challenge a prevailing assumption within quantum computing: adding more complex connections between quantum bits does not automatically guarantee better results, even with increased computational cost. This approach centres on refining the mapping of quantum results to classical labels via a trained ‘readout layer’, effectively calibrating the system to better understand its outputs. The readout layer functions as a trainable post-processing step, learning to compensate for imperfections in the quantum circuit and to extract the most relevant information from the measurement outcomes.

The implications of this work extend beyond simply achieving higher accuracy on benchmark datasets. Near-term quantum devices are limited by the number of qubits and the fidelity of quantum gates. Reducing the reliance on CNOT gates, which are typically the most error-prone operations, is crucial for building practical quantum classifiers. OQMD offers a promising strategy for mitigating the impact of gate errors and for maximising the performance of available quantum hardware. Furthermore, the technique is not limited to specific quantum circuit architectures. It can be applied to a wide range of variational quantum circuits, making it a versatile tool for quantum machine learning. The potential applications of improved quantum classification algorithms are broad, ranging from image recognition and natural language processing to financial modelling and drug discovery.

The methodology employed in this research involved training the readout layer using a standard gradient descent optimisation algorithm. The parameters of the single-qubit rotations were adjusted iteratively to minimise a loss function that quantifies the classification error. The performance of different configurations, including 0-CNOT, 9-CNOT, and 18-CNOT circuits, was compared using the Iris dataset. The use of the Mann-Whitney U test allowed the researchers to statistically assess the significance of the observed performance differences. Future work will focus on exploring the scalability of OQMD to larger datasets and more complex quantum circuits, as well as investigating the robustness of the technique to noise and imperfections in the quantum hardware. Understanding the interplay between circuit design and decoding strategies is essential for realising the full potential of near-term quantum computing.

The research demonstrated that optimising how quantum outcomes are interpreted, termed Optimal Quantum Measurement Decoding, achieved 83.33% accuracy on the Iris benchmark using circuits with zero CNOT gates. This is significant because it suggests that improved accuracy in quantum classification does not necessarily require deeper, more complex circuits reliant on potentially error-prone CNOT operations. By training a readout layer alongside the quantum circuit, researchers surpassed the performance of both 18-CNOT configurations with and without OQMD. The authors intend to explore how this technique scales to larger datasets and more complex circuits.