Quantum Networks Simplify Data Clustering Without Complex Measurements

A new unsupervised clustering framework, Qlustering, utilises steady-state quantum transport in quantum networks. Shmuel Lorber and Yonatan Dubi at Ben Gurion University of the Negev present a method for extracting data structure from unlabeled datasets via accessible transport observables, avoiding the need for full state tomography. The framework offers a hybrid classical-quantum workflow, performing data preparation and training classically, but using quantum dynamics for clustering, and provides a potentially scalable, tomography-free mechanism for unsupervised learning in open quantum networks. Benchmarking on synthetic, localization, QM9 and Iris datasets reveals competitive performance and strong stability across varying levels of dephasing.

Unsupervised quantum clustering via steady-state currents bypasses full state tomography

Performance across benchmark datasets improved by up to 15 per cent, demonstrating Qlustering’s ability to extract data structure from unlabeled datasets. Previous methods often struggled with task-specific requirements or inaccessible observables, but this advance unlocks unsupervised learning in open quantum networks. By utilising only steady-state output currents for cluster assignments, the need for full state tomography, a complex and resource-intensive process, is bypassed. Full state tomography requires measuring the density matrix of a quantum system, a task that scales exponentially with the number of qubits, rendering it impractical for even moderately sized systems. Qlustering circumvents this limitation by focusing on collective, measurable properties of the quantum network, specifically the steady-state currents flowing through it.

The Qlustering framework, developed through algorithm-hardware co-design, establishes a hybrid classical-quantum workflow, performing data preparation and training classically while using quantum dynamics for clustering itself. This co-design approach is crucial, ensuring the algorithm’s requirements align with the capabilities of current and near-future quantum hardware. Data preparation classically involves feature scaling and normalisation, transforming the raw data into a format suitable for encoding into the quantum system. The quantum stage then leverages the principles of quantum transport to map data points onto distinct clusters based on their inherent similarities. Synthetic datasets, alongside localization, QM9 and Iris, confirm stability across a broad range of dephasing strengths, a common source of error in quantum systems. Dephasing, caused by interactions with the environment, degrades quantum coherence and can significantly impact algorithm performance. The observed robustness of Qlustering to dephasing suggests a degree of resilience, potentially attributable to the use of steady-state measurements which average out some of the noise. Qlustering’s ability to extract data structure from steady-state output currents avoids full state tomography, relying instead on accessible transport observables. Measurement of steady-state currents offers an advantage over methods requiring more complex measurements, reducing experimental demands. The steady-state currents are determined by solving the GKSL (Gorini-Kossakowski-Sudarshan-Lindblad) master equation, which describes the time evolution of the quantum system under the influence of both coherent dynamics and environmental interactions. Unlabeled data structure can be extracted directly from transport observables, but scalability to larger datasets or more complex quantum systems remains a key challenge for practical application. The current implementation relies on relatively small quantum networks, and extending this to larger, more intricate systems will require further algorithmic optimisation and potentially novel hardware architectures.

Circumventing full quantum state measurement for scalable machine learning

Dr. Alessandro Silva and Dr. Patrick Rebentrost-Saclay are building machine learning tools using the principles of quantum mechanics, aiming to bypass limitations inherent in classical computing. Classical machine learning algorithms often struggle with high-dimensional data and complex relationships, while quantum algorithms offer the potential for exponential speedups in certain tasks. Qlustering, this new approach, offers a potentially flexible method for finding patterns in data without needing to measure every quantum detail, a notoriously difficult task. The framework’s unsupervised nature is particularly valuable, as it does not require labelled training data, which can be expensive and time-consuming to obtain. Initial demonstrations rely on a limited selection of datasets, however, raising questions about handling messy, high-dimensional data encountered in real-world applications. The synthetic dataset allows for controlled testing of the algorithm’s core principles, while the localization dataset tests its ability to identify patterns in spatially correlated data. The QM9 dataset, comprising molecular energies and structures, provides a benchmark for assessing performance on a chemically relevant problem. Finally, the Iris dataset, a classic machine learning benchmark, tests the algorithm’s ability to classify data points into different categories.

Although initial tests used relatively simple datasets, establishing a principle for quantum machine learning that sidesteps demanding measurement requirements is significant. Full state tomography is exceptionally challenging, and Qlustering’s reliance on measurable electrical currents represents a practical advantage. This advantage is particularly important for experimental implementations, where the cost and complexity of measurement can be a major bottleneck. This focus on accessible data opens avenues for applying quantum techniques to complex, real-world datasets currently beyond the reach of many quantum algorithms. Analysing steady-state output currents in engineered quantum networks reveals patterns within unlabeled data, and combining classical data preparation with quantum dynamics for clustering offers a hybrid approach suitable for implementation on existing and emerging quantum platforms, marking a progression in quantum machine learning beyond reliance on detailed quantum state measurements. Future work will focus on exploring different quantum network architectures, optimising the data encoding scheme, and developing methods for scaling the algorithm to larger datasets. The potential applications of Qlustering extend beyond traditional machine learning tasks, potentially finding use in areas such as materials discovery, drug design, and financial modelling.

The research demonstrated a new unsupervised clustering framework, Qlustering, which extracts data structure from steady-state quantum transport in quantum networks. This method avoids the need for full state tomography by instead measuring electrical currents, offering a more practical approach to quantum machine learning. Qlustering was benchmarked on synthetic, localization, QM9 and Iris datasets, achieving competitive performance and stability across varying conditions. The authors intend to explore network architectures and data encoding to scale the algorithm to larger datasets, representing an advance in utilising accessible quantum observables for data analysis.