Researchers Develop New Quantum Feature Map for Machine Learning

A generalised two-qubit Hamiltonian-based projective quantum feature map efficiently encodes classical data using both local and pairwise qubit interactions, as presented by Rafael Simões do Carmo and colleagues at São Paulo State University and Hospital Israelita Albert Einstein. The approach increases information density within quantum circuits while accommodating current hardware limitations and has been implemented in the open-source Python library, pqfmlib. Benchmarking across four biomedical datasets, utilising both IBM quantum processors and statevector simulations, demonstrates that this generalised Hamiltonian family consistently outperforms classical baselines, suggesting a key pathway towards achieving practical quantum advantage in the near term.

Enhanced quantum advantage via axis-resolved encoding in biomedical data analysis

Statistical gains over classical baselines reached 156 qubits, with a consistent pattern observed across four biomedical datasets. Achieving statistical quantum advantage previously required fewer qubits and was limited to narrow applications, but this generalised Hamiltonian family expands the scope of potential utility. Efficiently encoding classical data, the generalised two-qubit Hamiltonian-based projected quantum feature map embeds variables along different axes within a single qubit, maximising information density.

Implementation within the open-source library, pqfmlib, enables further research and development in this rapidly evolving field. More data can be processed in shallower circuits via axis-resolved encoding, reducing susceptibility to errors inherent in current quantum hardware. A demonstration revealed that the generalised two-qubit Hamiltonian family consistently outperformed reference projected quantum feature maps across the four biomedical datasets, achieving statistical gains when utilising up to 156 qubits on IBM quantum processors, alongside statevector simulations for comparison.

Further investigation is needed to understand how dataset characteristics and encoding choices influence performance. While these findings suggest a promising path towards near-term quantum utility, practical, broadly applicable quantum advantage is not yet definitively established. The observed performance remains heavily dependent on the specific dataset and hardware limitations, highlighting the need for continued optimisation and exploration of different encoding strategies.

Encoding classical data for enhanced quantum machine learning performance

A new approach to preparing quantum computers for machine learning has been engineered by researchers at São Paulo State University, focusing on efficiently translating complex data into a quantum format. This promises to increase the density of information within quantum circuits, an important step towards overcoming limitations imposed by current hardware. Success hinges on careful optimisation, as the performance of these quantum features is demonstrably sensitive to the specific dataset used and the choices made during the encoding process.

The team at São Paulo State University has demonstrated a flexible method for translating information into a format quantum computers can utilise, a key step given current hardware limitations. This approach, implemented in the publicly available ‘pqfmlib’ software, offers a pathway to explore how quantum processors might enhance machine learning tasks, even with imperfect systems. Embedding multiple classical variables within a single qubit builds upon this ability, increasing information density and potentially reducing the demands on error-prone quantum circuits.

Representing a step forward in preparing quantum computers for practical machine learning tasks, the team’s generalised two-qubit Hamiltonian-based projected quantum feature map moves beyond previously limited approaches. Offering a more flexible method for translating classical data into a quantum format suitable for processing, this advance allows for the exploration of different encoding strategies and their impact on machine learning outcomes. The technique provides a valuable tool for optimising quantum machine learning workflows.

Projected Quantum Feature Maps (PQFMs) represent a hybrid quantum-classical strategy for machine learning, leveraging the potential of quantum processors as feature generators. Classical machine learning algorithms often require careful feature engineering to achieve optimal performance. PQFMs offer a way to automate this process by utilising quantum circuits to transform classical data into a quantum feature space. The efficacy of a PQFM relies on its ability to create a feature space where data points are more easily separable, thereby improving the performance of subsequent classical classifiers. Previous PQFM designs, such as those based on counterdiabatic Ising-glass and one-dimensional Heisenberg models, have demonstrated initial promise but often lack the flexibility to efficiently encode diverse datasets.

The generalised two-qubit Hamiltonian-based PQFM introduced by Simões do Carmo et al. addresses this limitation by providing a unified framework for encoding classical features. This framework utilizes both local Pauli fields, which act on individual qubits, and pairwise two-qubit Pauli interactions. Crucially, distinct classical variables can be embedded along different Pauli axes (X, Y, and Z), effectively tripling the information density per qubit compared to encoding schemes that rely on a single axis. This axis-resolved encoding is a key innovation, allowing for a more compact and potentially more expressive feature map. The Hamiltonian itself is constructed to facilitate this encoding, with parameters adjusted to reflect the values of the classical variables being embedded.

The implementation of this PQFM within the pqfmlib library is significant as it provides a readily accessible platform for researchers to explore and build upon this work. The library facilitates the creation, training, and evaluation of quantum machine learning models using this specific feature map. The benchmarking process involved four biomedical datasets, chosen to represent a range of data characteristics and complexities. These datasets were used to compare the performance of the generalised Hamiltonian family against both classical machine learning baselines and other existing PQFM designs. The use of both IBM quantum processors and statevector simulations allowed for a comprehensive evaluation, accounting for the effects of quantum noise and hardware limitations. The achievement of statistical gains with up to 156 qubits suggests a substantial improvement in performance, although the precise nature of this advantage requires further investigation.

The observed performance gains are not merely academic; they have potential implications for various biomedical applications. These include disease diagnosis, drug discovery, and personalised medicine, where accurate classification and prediction are paramount. By efficiently encoding complex biomedical data, this PQFM could enable the development of more powerful and accurate machine learning models for these critical tasks. However, it is important to acknowledge that the performance of the PQFM is sensitive to both the dataset and the specific encoding choices made. Factors such as the dimensionality of the data, the presence of noise, and the choice of Pauli axes all influence the resulting feature space and the performance of the classifier. Therefore, careful optimisation and adaptation are necessary to achieve optimal results for any given application.

Future research will focus on exploring the interplay between dataset characteristics, encoding strategies, and quantum hardware limitations. Investigating the robustness of this approach to noise and the potential for error mitigation techniques will be crucial for realising practical quantum advantage. Furthermore, extending this generalised Hamiltonian family to incorporate more qubits and more complex interactions could lead to even more expressive feature maps and improved machine learning performance. While the path to broadly applicable quantum advantage remains challenging, this work represents a significant step forward in harnessing the power of quantum computing for machine learning.

The researchers demonstrated that a generalised two-qubit Hamiltonian-based projected quantum feature map consistently outperformed classical baselines in biomedical classification tasks. This is significant because it provides a more efficient method for encoding classical data onto quantum processors, potentially improving the performance of machine learning models. Using up to 156 qubits on IBM quantum processors and statevector simulations, the study showed statistically supported gains across four datasets. The authors intend to further investigate how dataset characteristics and hardware limitations influence performance, alongside exploring error mitigation techniques.