Utilizing Differences in the Sampling Efficiency of the Quantum Annealer and the Classical Sampling for Mitigating Catastrophic Forgetting in Training Undirected Probabilistic Graphical Models.

The D-Wave Quantum Annealer (QA) showed an early promise of efficient sampling to train undirected probabilistic graphical models. In our previous work, important differences in the D-Wave and the Gibbs sampling techniques have been revealed when sampling from classically trained restricted Boltzmann machines (RBMs).

In this work, we will demonstrate how the previously observed trends explain the failure of many earlier investigations to achieve substantial improvement in RBM training when using D-Wave-based sampling. Further, we will explore possibilities for improvement by using a classical-quantum sampling approach. We will show that the potential for improvement is limited by the fact that the QA and the classical techniques are less complementary to each other than could be expected. Next, we will demonstrate a strong promise for another relevant application of the QA for RBMs, which benefits from the D-Wave’s ability to efficiently generate a large variety of potentially important samples, including states with low probability. We will show how the D-Wave can be used to generate samples belonging to desirable classes, to be subsequently used as memories for a replay-based mitigation of catastrophic forgetting (CF) during RBM training. High efficiency of the D-Wave as a generative model allows nearly complete mitigation of the CF under certain applications-relevant constraints. Most promising possibilities of overcoming those constraints will be also investigated.