Streaming Variational Inference for the Indian Buffet Process

Abstract

The Indian Buffet Process (IBP) is a nonparametric prior that provides latent feature models with a Bayesian framework for learning the set of features that co-occur in our observations. For example, images may be composed of several objects, and we may wish to identify which images contain which objects. The IBP would provide a principled prior in situations where the number of hidden features is unknown and potentially unbounded, as it assumes that the observed data only manifests a finite subset of the unbounded amount of latent features. However, contemporary inference procedures for the IBP in general do not scale well to the size of the data set, or might require the number of features in the IBP prior to be truncated in order to perform inference effectively. Hence, we develop an inference procedure for the IBP that can be used in the streaming context, where we receive small batches of data at a time, and also propose a technique for adaptively modifying model complexity on-the-fly. We then evaluate our proposed methods on several data sets and show the improved predictive performance and computational speed over other inference methods for the IBP.