Quantum Feature Extraction with 156 Qubits Enhances Machine Learning Performance on Complex Data

Quantum computing promises revolutionary advances in machine learning, and a team led by Anton Simen and Carlos Flores-Garrigós, with contributions from Murilo Henrique De Oliveira et al., now demonstrates a novel method for extracting complex features from data using quantum dynamics. The researchers developed a technique that translates classical data into the language of quantum mechanics, specifically by embedding information within the interactions of quantum bits, or qubits. By evolving these quantum states on a digital quantum processor, they effectively map data into a higher-dimensional space, revealing subtle statistical relationships often hidden from traditional machine learning algorithms. This approach, tested on challenging real-world problems such as predicting molecular toxicity and identifying images, yields features that not only complement existing classical methods, but in some cases significantly outperform them, suggesting a promising pathway for utilising near-term quantum devices in practical data analysis.

Researchers introduce a Hamiltonian-based quantum feature extraction method that enhances machine learning performance by generating complex features through the dynamics of many-body spin Hamiltonians. Classical feature vectors are embedded into spin-glass Hamiltonians, where both individual feature contributions and higher-order correlations are represented through interactions between multiple quantum bits. By evolving the system on IBM digital quantum processors with 156 qubits, the data are mapped into a higher-dimensional feature space via measurements of various quantum properties, effectively transforming the input data into a more complex and informative representation.

Hamiltonian Feature Extraction via Hypergraph Embedding

Scientists developed a novel Hamiltonian-based feature extraction method to improve machine learning performance by generating complex features via the dynamics of many-body spin Hamiltonians. The study pioneers a technique where classical feature vectors are embedded into spin-glass Hamiltonians, representing both individual features and their statistical dependencies through interactions between quantum bits. Researchers meticulously constructed a Hamiltonian that encodes both individual features and their relationships, estimating coefficients directly from the dataset’s statistics. This Hamiltonian utilizes hypergraphs to efficiently map interactions onto the architecture of a quantum processor.

The team then evolved the quantum system using a controlled approximate counter-adiabatic dynamics, specifically in a fast impulse regime, on IBM’s 156-qubit quantum processor, ibm_kingston. This carefully designed evolution partially suppresses unwanted transitions, enriching the resulting quantum state and allowing for non-linear mixing of input correlations. Subsequently, scientists measured the expectation values of various quantum properties from the final quantum state, constructing feature maps that capture complex statistical dependencies difficult to access with classical methods. To assess the approach, researchers applied it to high-dimensional, real-world datasets, including molecular toxicity prediction and breast tumor detection using image data. Experiments demonstrate that quantum-derived features consistently enhance the performance of classical models when combined with standard preprocessing techniques, suggesting a scalable and effective route to expressive feature representations on near-term quantum devices. This innovative method provides a pathway for integrating Hamiltonian-based encoding with quantum dynamics, offering a significant advancement in quantum machine learning.

Quantum Feature Extraction with Spin-Glass Hamiltonians

Scientists have achieved a breakthrough in machine learning through the development of a novel quantum feature extraction method. This work centers on encoding classical data into complex, many-body spin-glass Hamiltonians, leveraging the dynamics of quantum systems to generate richer feature representations. The team successfully mapped classical feature vectors onto a 156-qubit digital quantum processor, ibm_kingston, embedding both individual features and their higher-order correlations within the Hamiltonian. This encoding scheme captures statistical dependencies difficult to access with standard classical methods.

Experiments revealed that by driving the quantum system through a controlled, approximate counterdiabatic dynamics, the resulting quantum state is enriched, enhancing the fidelity of subsequent measurements. Specifically, the researchers utilized a fast, impulse-regime dynamics, simulating the evolution with a single step, which restricts the quantum state to lower energy levels. Measurements of various quantum properties from the resulting state then constructed feature maps that capture complex statistical dependencies within the data. The team evaluated this framework on two real-world problems: molecular toxicity prediction and breast tumor detection using image-based datasets.

Results demonstrate that quantum-derived features consistently enhance the performance of classical machine learning models when combined with standard preprocessing techniques. The approach utilizes both two-body and three-body Hamiltonians, extracting up to three-body correlations, and the team analytically computed the adiabatic gauge potential for arbitrary spin-glass problems. These findings suggest that integrating Hamiltonian-based encoding with quantum dynamics provides a scalable and effective route to obtain expressive feature representations on near-term quantum devices, opening new possibilities for data-driven applications.

Quantum Feature Extraction Boosts Classical Machine Learning

This work demonstrates a scalable quantum feature extraction framework that enhances classical machine learning performance by embedding data into spin-glass Hamiltonians and evolving them using quantum dynamics. The team successfully mapped both molecular toxicity and medical image classification tasks onto IBM quantum devices, utilizing up to 156 qubits, and achieved statistically significant performance gains despite hardware limitations. By generating expressive, nonlinearly correlated features, the method improves upon purely classical approaches, demonstrating a practical pathway for building hybrid classical-quantum pipelines. Results indicate a clear advantage for quantum-derived features, with a 121% improvement in precision for molecular toxicity classification and a 5.

5% increase in the mean area under the curve for breast tumor detection. Notably, on the Breast MedMNIST benchmark, the hybrid approach achieved an area under the curve of 0. 937, surpassing deep learning models like ResNet18 and ResNet50. Analysis using SHAP values revealed that quantum-derived features contributed dominantly to model decisions, with up to 86% of the total importance originating from these features, suggesting the capture of unique and complementary information. The authors acknowledge the constraints imposed by hardware connectivity, which currently limits the scale of these quantum computations. Future research will likely focus on mitigating these limitations and exploring the potential of this framework with more advanced quantum hardware. Nevertheless, this work establishes a hardware-efficient approach to quantum feature extraction, demonstrating the ability of quantum dynamics to generate rich, nonlinear features that improve generalization in classical machine learning models.