H.B. Keller Colloquium

This talk will introduce the concept of active learning in the context of
similarity graphs.  Similarity graphs are constructs in which nodes
represent pieces of data and edges represent similarity between the on the
nodes spanned by the edge.  It is easy to construct a semi-supervised
learning problem using the Graph Laplacian.  A modern problem, not easily
addressed by neural networks alone, is classification of data with very
little training data.  This is often the case when the ground truth can be
hard to establish due to the remoteness of the data, such as in remote
sensing applications.  I will discuss a body of work at UCLA that
addresses this problem for such applications as hyperspectral imaging,
sythetic aperture radar, and multispectral imaging for environmental
problems such as surface water detection.  Some of the new work is joint
with Los Alamos National Laboratory.