Hierarchical Bayesian Modeling of Text and Image Statistics

• Activity Number: 316
• Type: Invited
• Date/Time: Wednesday, August 6, 2003 : 8:30 AM to 10:20 AM
• Sponsor: General Methodology
• Author(s): David M. Blei*+ and Michael I. Jordan+
• Companies: University of California, Berkeley and University of California, Berkeley
• Keywords: graphical models ; empirical Bayes ; hierarchical models ; information retrieval ; image processing

Abstract

We present a simple hierarchical Bayesian model for collections of multi-media documents. The model posits that documents or images are generated by choosing a random set of multinomial probabilities for a set of possible "topics," and then repeatedly generating words by sampling from the topic mixture. This model is intractable for exact probabilistic inference, but approximate posterior probabilities and marginal likelihoods can be obtained via fast variational methods. We also consider the problem of modeling annotated data, i.e., data with multiple types where the instance of one type (such as a caption) serves as a description of the other type (such as an image). We describe three hierarchical mixture models that are aimed at such problems, culminating in the "Corr-LDA model," a latent variable model that is effective at both joint clustering and automatic annotation. We conduct experiments testing these models using the Corel database of images and captions.