Automating Variational Inference for Statistics and Data Mining
Invited talk at the 74th Annual Meeting of the Psychometric Society (IMPS 2009)
I will describe Infer.NET, a free software package from Microsoft that automatically applies variational Bayesian inference to a statistical model of your choosing. Unlike sampling methods, variational methods approximate the posterior distribution as a point estimate plus uncertainty, making them well suited to large-scale time-varying datasets. Infer.NET is structured as a compiler: it takes a model specification as input and produces a specialized inference program as output. This automated process makes it easy to experiment with different models and get an efficient program for each one. I will demonstrate Infer.NET with models from the psychometric literature.