Speaker
Description
Quantum kernel methods embed classical data into quantum Hilbert spaces via parameterised feature-map circuits, but the design of these circuits remains largely manual, relying on fixed ansätze chosen by convention rather than principled optimisation. We present an evolutionary algorithm (EA) framework that searches directly over the space of quantum circuit architectures using a hierarchical domain-specific language, replacing hand-crafted feature maps with automatically discovered ones. Our EA employs a composite fitness function combining cross-validation accuracy, a complexity penalty, and a scalability bonus, with a monotonicity gate to prevent reward-hacking of the scalability term.
Across benchmark datasets including MNIST digit classification subtasks, we find that EA-discovered architectures are most competitive against classical and fixed-ansatz quantum baselines at small training set sizes, and that discovered structures generalise across qubit counts (nq = 2–10) without re-running the search from scratch, revealing qubit-agnostic structural principles rather than dataset- or size-specific solutions. Notably, we find that kernel-target alignment (KTA), a widely used theoretical proxy for kernel quality, anti-correlates with true classification accuracy when used as a direct optimisation target, suggesting caution in its use as a fitness signal for kernel search more broadly. We additionally introduce a sparse kernel evaluation strategy to reduce the computational cost of kernel matrix construction during search. Together, these results offer both a practical architecture-search tool and a cautionary methodological finding for the design of fitness functions in quantum kernel optimisation.
| Topic | Quantum algorithms for machine learning applications |
|---|