Bayesian Learning in Dual-Wing Harmoniums Applied to Information Retrieval and Genomics

Dual-wing harmoniums is a promising technique for modeling the relationship between heterogeneous data sources, like associated text and images or genetic variations and observed traits. Unsatisfied with contrastive divergence (CD) and maximum likelihood learning, we implemented Bayesian learning in DWH using brief Langevin MCMC approach. We proposed three different types of priors for both information retrieval and quatitative trait loci (QTL) mapping and examined the results. Bayesian learning and different priors are evaluated by classification and retrieval task for TRECVID’03 video clips and cross validation of predicting morphological shapes of two Drosophila species. Experiments shows that Bayesian learning gives close to CD accuracy and average precision in information retrieval, but fails on QTL prediction. We discussed possible causes and future directions.