Time and Location:   April 7, 2004, 2-3:00pm, Room 1314.

The Intrinsic Geometry of Natural Image Data and Learning by Diffusion

Speaker:  Ann Lee, Yale University


An important question in vision research is how to best model and code
natural images. We will first explore the empirical statistics of large
sets of patches from images in a normal visual environment and ask what is
the intrinsic geometry of natural image data. A new picture of natural
image statistics seems to emerge from this study, where geometric
primitives (such as edges, blobs and bars) generate a complex hierarchy of
non-linear manifolds where the data is densely concentrated in state
space. In the second part of the talk, we will then describe a general
methodology for organizing high-dimensional data sets by embedding them in
Euclidean space via a non-linear diffusion map.  Preliminary results will
be shown for both generic image patches and hyperspectral data of
pathology slices.