On-line calibration scheme for coaching restricted Boltzmann machines with quantum annealing
Authors: Takeru Goto, Masayuki Ohzeki
Summary: We suggest a scheme for calibrating the D-Wave quantum annealer’s inner parameters to acquire well-approximated samples to coach a restricted Boltzmann machine (RBM). Empirically, samples from the quantum annealer obey the Boltzmann distribution, making them appropriate for RBM coaching. Nevertheless, it’s exhausting to acquire applicable samples with out compensation. Current analysis typically estimates inner parameters, such because the inverse temperature, for compensation. Our scheme makes use of samples for RBM coaching to estimate the inner parameters, enabling it to coach a mannequin concurrently. Moreover, we take into account extra parameters past inverse temperature and show that they contribute to enhancing pattern high quality. We consider the efficiency of our scheme by evaluating the Kullback-Leibler divergence of the obtained samples with classical Gibbs sampling. Our outcomes point out that our proposed scheme demonstrates efficiency on par with Gibbs sampling. As well as, the coaching outcomes with our estimation scheme are higher than these of the Contrastive Divergence algorithm, often known as a normal coaching algorithm for RBM
