Hi, For Friday's presentation, I will go through Dellaert's paper (http://www.cc.gatech.edu/~dellaert/em-paper.pdf). As I read some papers on EM, I saw different approaches to maximize the likelihood of parameters given incomplete data or data with hidden parameters all of which are essentially the same. Dellaert uses the lowerbound maximization approach. I chose that paper, since it nicely presents the idea underlying EM and gives you the intuition that EM really works :) Attachment contains the first draft of the presentation. Probably I will add a couple of slides for HMM-Baum Welch and K-means sections. Also a nice discussion: Minka says EM is a primal-dual algorithm. Intuitively, it seems to me very reasonable, but I'm still not clear about how is an EM problem converted to Restricted Primal of the Dual. So, if anybody has an idea or could find out, I think it would be nice to discuss. -f.gelgi
Attachment:
ExpectationMaximization_v2.ppt
Description: MS-Powerpoint presentation