Investigating the Performance of the Quantum Neuron Born Machine on Near-Term Hardware

Unsupervised generative models are a class of machine learning algorithms designed to learn underlying patterns in data for the purpose of generating new samples. While these models have demonstrated powerful performance on challenging tasks ranging from drug discovery to image generation, improving their capabilities is still an active area of research. Parameterized quantum circuits can model probability distributions with classical training and therefore are a natural fit for tackling these data-driven problems. The Quantum Neuron Born Machine (QNBM) is a quantum generative model that employs non-linear activations through repeat-until-success circuits, containing mid-circuit measurements as well as classical control, which enables it to learn complex probability distributions. It has been shown to perform well on an ideal simulator but has yet to be realized and benchmarked on quantum hardware. To benchmark the QNBM on hardware, we look to IBM’s superconducting device as well as Quantinuum’s ion-trap technology. We conduct a further investigation into the model’s resource requirements such as the number of circuit executions to enable good performance of the QNBM on near-term devices while minimizing hardware execution costs. We train the model with various noise models, shot variations, and initial training parameters to provide these estimates, and then determine the essential requirements for the QNBM on the respective devices. We then compare the resource requirements for two versions of the model—one utilizing post-selection executed on IBM’s devices and one employing classical control on Quantinuum’s technology. Then, utilizing our resource estimates, we run the first small-scale hardware realization of the QNBM on both types of technologies and assess its ability to model a cardinality-constrained probability distribution.