Recent research on computational cognitive models (e.g., ACT-R, 4CAPS)
has begun to develop computational models of cognitive processes that
directly predict features of fMRI activity in different brain regions.
We will examine ways in which machine learning methods might help in
developing more accurate and precise models of this type, by
automatically discovering the spatial-temporal patterns of fMRI
activation associated with specific cognitive subprocesses.  We first
describe our recent successes training statistical machine learning
classifiers to distinguish cognitive subprocesses such as reading a
sentence versus viewing a picture, or reading words about tools versus
words about buildings, based on the observed fMRI brain activation of
the subject.  We then describe our recent research to learn more
complex models describing sequences of multiple cognitive processes
operating in parallel.  The goal of this research is to build models
of interacting cognitive subprocesses that overlap in time and space.