Generative model benchmarks for superconducting qubits

In this work we demonstrate experimentally how generative model training can be used as a benchmark for small (fewer than five qubits) quantum devices. Performance is quantified using three data analytic metrics: the Kullback-Leibler divergence and two adaptations of the F₁ score. Using the 2×2 bars and stripes data set, we train several different circuit constructions for generative modeling with superconducting qubits. By taking hardware connectivity constraints into consideration, we show that sparsely connected shallow circuits outperform denser counterparts on noisy hardware.