Abstract:
In this talk, I will introduce the newly proposed Stability Region based Expectation-Maximization (EM) algorithm and discuss its advantages and applications. In the context of model-based clustering, the widely used EM algorithm often gets stuck at a local maximum of the log-likelihood surface. To overcome this problem, a novel two phase algorithm that can systematically compute neighborhood local maxima on the log-likelihood surface is proposed. Our algorithm uses the newly developed TRUST-TECH (TRansformation Under STability reTained Equilibria CHaracterization) framework that can explore the dynamic and geometric characteristics of the stability regions of the nonlinear dynamical system corresponding to its nonlinear function. As a case study of the proposed framework, I will discuss the application of our algorithm to the motif finding problem in bioinformatics. Our algorithm has been tested on both synthetic and real datasets and significant improvements in the performance compared to other approaches are demonstrated. This generic framework also provides the flexibility of using different local solvers and global methods that work well for some specific datasets. Other advantages and applications of general TRUST-TECH based algorithms will be presented in the context of machine learning and knowledge discovery. I will also discuss some of my other works on scale-space theory in relation with boosting and mixture modeling. Finally, I will conclude my discussion with the future research directions that I would like to pursue.

Biography:
Chandan Reddy is a PHD candidate in the department of Electrical and Computer Engineering at Cornell University. He obtained his Masters and Bachelors degrees in Computer Science and Engineering from Michigan State University and Pondicherry University respectively. His research interests are interdisciplinary in nature and include the following areas: machine learning, data mining, computational statistics, bioinformatics and biomedical imaging. He is currently working with IBM Research and he also worked with Siemens Corporate Research in the past.