In this talk, we demonstrate iterative-structured machine learning such as deep learning by using the D-Wave Advantage.
As is well known, the D-Wave quantum annealer outputs many realizations of the spin configurations following the Gibbs-Boltzmann distribution. The D-Wave quantum annealer is thus expected to be used in the Boltzmann machine learning and related models to compute the expectation at each epoch during the learning process. However, due to the bottleneck to newly input the parameters on the quantum processing unit, the efficient implementation is not demonstrated for deep architecture. In particular, the sampling conditioned on the visible variables demands repetition of the sampling by changing the parameters of the machine learning model.
We propose a new algorithm to perform the efficient Boltzmann machine learning including the hidden units and generalization to the iterated-structured machine learning such as deep learning. In addition, we introduce several aspects of our proposal. The first topic is that the architecture can be regarded as the variational circuits consisting of the quantum processing units on the D-Wave quantum annealer. The second one is that, by taking some limitation on several conditions, we find a relationship between our proposed method and several machine learning methods. The third one is that our technique makes it easier to perform the Boltzmann machine learning with the continuous variables available for practical applications.

Please access here for the registration.