CATVI: conditional and adaptively truncated variational inference for hierarchical Bayesian nonparametric models

Current variational inference methods for hierarchical Bayesian nonparametric models can neither characterize the correlation struc- ture among latent variables due to the mean- eld setting, nor infer the true posterior dimension because of the universal trunca- tion. To overcome these limitations, we pro- pose the conditional and adaptively trun- cated variational inference method (CATVI) by maximizing the nonparametric evidence lower bound and integrating Monte Carlo into the variational inference framework. CATVI enjoys several advantages over tra- ditional methods, including a smaller diver- gence between variational and true posteri- ors, reduced risk of undertting or overt- ting, and improved prediction accuracy. Em- pirical studies on three large datasets re- veal that CATVI applied in Bayesian non- parametric topic models substantially out- performs competing models, providing lower perplexity and clearer topic-words clustering.