Speaker: Dana Wilkinson
Given a collection of data points annotated with action labels
(obtained, for example, from a robot mounted with a web camera or
a laser range finder) we wish to learn a representation of the
data which can be used as a map (for planning, etc.). We argue
that creating maps involves a high degree of compression and that
the actions must correspond to simple operations on the map, in
other words the problem will be cast as a variation of a common
machine learning task---dimensionality reduction---which somehow
respects the action labels.
Here, a solution is proposed where the actions correspond to
distance-preserving transformations in a representation learned
with a semidefinite program. Additionally, the transformations in
the learned representation which correspond to the actions can be
recovered and used. Some resulting maps (and their uses) will be
demonstrated in a robot-inspired image domain.