March 13th, 2009, 1:21 pm
QuoteThis kind of modeling is nice in principle, but doing inference is slow. Being nonparametric means that you have a much bigger space to explore for inference. True speed is a problem, but some recent developments in improving computational efficiency shows promise. With nonparametric modelling we might assume, for example, that the data are sampled from a completely unknown distribution, F, and the goal is to make inferences about functions, or even the pdf, of F. We could think of F as belonging to the class of all continuous distributions on the real line for example. So while the space is large in nonparametric modelling, once we combine data information through the posterior we reduced this. It's more complex than this in real life but I will spare you the details SG