Seminar on Computational Learning and Adaptation

  Bayesian Segmentation: A New Way to Find Structures in Time Series Data and Images

Jeffrey Scargle
NASA Ames Research Center
Mountain View, CA
jeffrey@cloud9.arc.nasa.gov

The Bayesian Blocks algorithm finds the most probable piecewise constant ("blocky") representation for time series data. In ApJ 504, 405, 1998 (xxx.lanl.gov/abs/astro-ph/9711233) the number of blocks was determined in an ad hoc iterative procedure. Another approach maximizes the posterior -- after marginalizing all parameters except the number of blocks -- computed with Markov Chain Monte Carlo methods.

A new algorithm starts with the Voronoi tessellation of the individual events in an arbitrary dimensioned data space, and merges cells based on a Bayesian model comparison criterion. This method addresses problems such as the identification of structures in images and detection of clusters in high dimensional parameter spaces.

I will demonstrate these algorithms on time series data from the most interesting of astronomical objects -- the gamma-ray bursts -- and describe some ideas on the design of a Bayesian Data Processing Automaton based on these ideas.

Date: Thurs., Feb 8

Time: 4:15-5:30PM

Place: Cordura 100

Return to the seminar schedule