A1095
Title: A Hyvarinen score-based approach to quickest change detection in unnormalized statistical models
Authors: Taposh Banerjee - University of Pittsburgh (United States) [presenting]
Vahid Tarokh - Duke University (United States)
Sean Moushegian - Duke University (United States)
Abstract: Score-based methods have become increasingly popular for modeling and generation. Scores or the gradient of log density are used to develop methods for the quickest detection of changes in unnormalized and score-based models. These methods can be applied to detect changes in high-dimensional data. Analytical bounds on the false alarm and delay performances are provided. A robust method is then discussed for performing detection in such models. Next, a diffusion score-based method is introduced that can be used to approach the performance of optimal likelihood ratio-based tests. Finally, the effectiveness of the methods are shown by applying them to real and simulated high-dimensional data.
