Date: Monday, December 15, 2014
Location: Rockefeller Hall 310
Title: What's ‘best' in 'Inference to the best explanation’? Model selection and Bayesian Occam’s Razor
Most inductive systems balance fit with existing data against ‘parsimony’, or simplicity of generalizations. Using an example from grammar-learning, I will describe how the likelihood function in statistics gives rise to a definition of parsimony in terms of model expressiveness, and illustrate how a Bayesian treatment allows this same sense of parsimony to be used to select among models of varying complexity, without separate mechanisms to protect against overfitting. In the second half of the talk, I will review examples from two projects modeling natural language and visual scene data where this 'Bayesian Occam's Razor' provides a principled way to balance fit and parsimony in problems with latent variables, such as classification and clustering.
Tea at 4:15pm in Rockefeller Hall 305