1: \begin{abstract}

2: Fast convergence speed is a desired property for training latent Dirichlet allocation (LDA),

3: especially in online and parallel topic modeling for massive data sets.

4: This paper presents a novel residual belief propagation (RBP) algorithm

5: to accelerate the convergence speed for training LDA.

6: The proposed RBP uses an informed scheduling scheme for asynchronous message passing,

7: which passes fast-convergent messages with a higher priority to influence those slow-convergent messages at each learning iteration.

8: Extensive empirical studies confirm that RBP significantly reduces the training time until convergence while achieves a much lower predictive perplexity

9: than other state-of-the-art training algorithms for LDA,

10: including variational Bayes (VB),

11: collapsed Gibbs sampling (GS),

12: loopy belief propagation (BP),

13: and residual VB (RVB).

14: \end{abstract}