1/27/04: Rich Zemel, University of Toronto
<br>
<br>

Learning Parts
<br>
<br>

                                                                                                    
Many collections of data exhibit a common underlying structure: they
consist of a number of parts or factors, each with a range of
possible states.  For example, in a collection of facial images,
every image contains eyes, a nose, and a mouth, each of which has a
number of possible appearances.  I will present a new method for the
unsupervised learning of parts-based representations of
high-dimensional data.  Our technique automates the segmentation of
data dimensions into parts, while simultanously learning a model of
part appearances.  Inference and learning are carried out
efficiently via variational algorithms.  I will describe
applications of this approach to problems in image decomposition,
text modeling, and collaborative filtering.
<BR>                                                                                                         
Time permitting, I'll briefly describe other current work on
learning multiscale random fields for image labeling.