In this talk, we explore the application of Generalized Approximate Message
Passing (GAMP) to binary linear classification and feature selection.  We show
that GAMP can easily accommodate standard response models (logistic, probit,
hinge loss) and outlier-corrupted extensions, as well as standard
sparsity-promotion mechanisms for the predictor (ell-1 regularization, sparse
probabilistic priors).  Moreover, the hyperparameters that govern these models
can be automatically tuned using an EM approach.  In addition, a state-evolution
framework characterizes the algorithm's trajectory under certain assumptions
about the feature matrix.  Finally, kernel methods can be easily incorporated if
desired.  Numerical results demonstrate the flexibility, efficiency, and
accuracy of this approach.