Quantum breakthrough enhances machine learning with spin-glass dynamics
A research team led by Anton Simen and Carlos Flores-Garrigos has developed a new quantum method for improving machine learning. Their approach uses quantum dynamics to extract complex features from data, boosting the performance of classical models. Tests on real-world tasks, such as predicting molecular toxicity and image recognition, show promising results. The method begins by embedding classical data into quantum systems. Feature vectors are mapped onto spin-glass Hamiltonians, where interactions between qubits represent both individual features and their statistical dependencies. This process captures relationships that classical techniques often miss.
To generate richer features, the team employs a Hamiltonian-based technique on IBM's ibm_kingston quantum processor. They use hypergraph embeddings to represent high-order data as vertices and hyperedges within a quantum state. A parameterised Hamiltonian, combined with approximate counter-diabatic driving, suppresses unwanted excitations and speeds up adiabatic evolution. This produces entangled low-energy states that encode nonlinear feature maps.
The system evolves using controlled dynamics on a 156-qubit processor, extracting up to three-body correlations from both two-body and three-body Hamiltonians. The researchers also computed the adiabatic gauge potential for arbitrary spin-glass problems, ensuring precise control over the quantum state. By evolving these states, the method maps data into a higher-dimensional space, revealing subtle statistical patterns.
Experiments confirm that quantum-derived features enhance classical models when paired with standard preprocessing. The approach has already been tested on practical challenges, delivering features that match or surpass existing classical methods. The new technique bridges quantum mechanics and machine learning by translating classical data into quantum interactions. Its success on tasks like toxicity prediction and image classification suggests broader applications. With further development, this method could expand the capabilities of both quantum and classical computing systems.
