Student Talk- Dan Kaslovsky October 28th, 2010
4:00-5:00 PM in ECOT 831
Topic: Local Manifold Geometry for Processing Large Data Sets

Abstract:
The volume of data created by science and society is growing at an unprecedented rate. As our ability to collect and store information is ever growing, it is essential that our ability to understand and thus process these massively large datasets keeps pace. Many datasets, while presented in high dimension, are actually organized along a lower dimensional manifold. Processing such datasets then becomes a problem of learning the geometry of this manifold. Traditional algorithms proceed by producing a global representation of the data. We are developing a technique for learning a manifold's geometric properties at a local scale and are designing algorithms that exploit this local information to allow for processing (e.g., approximation, denoising, searching) of data to an accuracy limited only by the original sampling.