Time-Frequency Seminar, 03-28-2002

Change of Location:
This talk will be in Room FINE 214 (second floor of Fine Hall) !

Gilad Lerman

Multiscale geometric clustering of data sets

 Abstract:We present a theory and a fast algorithm for partitioning a low-dimensional data
set in Rn into disjoint clusters of different geometric structures. The main tool used is an
L2 variant of Jones' beta numbers, which  measure the scaled deviations of the data set from
approximating d-planes at different scales and locations. The Jones function is formed by
adding the squares of the L2 Jones numbers at different scales and the same location.  A theory
by Peter W. Jones and the speaker shows that a set with low values of the d-dimensional Jones
function is "well-approximated" by a certain d-dimensional surface with small d-volume.

We suggest a fast algorithm for computing a numerical approximation of the Jones function and
use it to cluster the given set according to the computed values of this function. Here we apply
this procedure to blocks of pixels taken from images (the highet ambient dimension is 25) and to
a six-dimensional set of currency exchange rates.