Stochastic Systems Group
Home Research Group Members Programs  
Demos Calendar Publications Mission Statement Alumni

SSG Seminar Abstract


Geometric entropy minimization (GEM)

Alfred Hero
University of Michigan, Ann Arbor


We introduce geometric entropy minimization as a framework for non-parametric inference. The framework is based on the asymptotic behavior of k-point minimal graphs that tends to pick out most concentrated regions of a data sample. In this framework we formulate a transductive anomaly detection method that is asymptotically equivalent to minimum volume set estimation but is applicable to high dimensional feature spaces. We also use this framework to derive a spectral clustering method, called dual rooted diffusions, and apply it to semi-supervised classification problems.



Problems with this site should be emailed to jonesb@mit.edu