Hybrid Helmholtz machines: a gate-based quantum circuit implementation

Quantum machine learning has the potential to overcome problems that current classical machine learning algorithms face, such as large data requirements or long learning times. Sampling is one of the aspects of classical machine learning that might benefit from quantum machine learning, as quantum computers intrinsically excel at sampling. Current hybrid quantum-classical implementations provide ways to already use near-term quantum computers for practical applications. By expanding the horizon on hybrid quantum-classical approaches, this work proposes the first implementation of a gated quantum-classical hybrid Helmholtz machine, a gate-based quantum circuit approximation of a neural network for unsupervised tasks. Our approach focuses on parameterized shallow quantum circuits and effectively implements an approximate Bayesian network, overcoming the exponential complexity of exact networks. In addition, a new balanced cost function is introduced, preventing the need of millions of training samples. Using a bars and stripes data set, the model, implemented on the Quantum Inspire platform, is shown to outperform classical Helmholtz machines in terms of the Kullback–Leibler divergence.