Shape Completion Using Deep Boltzmann Machine

Algorithm 1

Training a DBM.

Given: A training set consists of samples denoted as . Number of markov particles . Number of iterations .

Output: A trained DBM model with parame ters

() Use to pre-train the DBM, and get the initial parameters of DBM .

() Randomly initialize the markov particles for MCMC.

() For to :

 (a) For each training sample , use mean-field approach to get the variational parameters.

 (b) For each markov particle , use (3) repeatedly to obtain the state .

 (c) Update the parameters of DBM with equations:

    The update of , , is similar.