Recent work at the confluence of machine learning and statistics has
taken motivation from the fact that many problems are very high
dimensional, with the number of variables often much larger than the
number of samples. In this "large p small n" regime, even classical
questions take on new interest, and become challenging. We present an
overview of recent results on the problem of estimating sparse graphs
in high dimensional data. The graphs are usually associated with a
probabilistic graphical model, such as a Gaussian random field or
Ising model. In the high dimensional regime, it is of interest to
characterize the scaling laws in the number of model parameters and
sample size under which the graph can be accurately estimated --
either with or without regard for computational efficiency. We
describe such results for Ising models, Gaussian graphical models,
time series models, and certain nonparametric extensions of these
models. This is joint work with Larry Wasserman, Martin Wainwright,
Pradeep Ravikumar, and Han Liu.
Meet the speaker in Room 212 Cockins Hall at 4:30
p.m. Refreshments will be served.