Faculty of Engineering and Natural Sciences
Competitive Signal Processing
Süleyman Serdar Kozat
Abstract: In this talk, we define a competitive framework for general adaptive signal processing applications where we focus on achieving the best possible performance for any given signal with respect to a class of competitive algorithms. We show that by permitting a combination of candidate models, we can obtain algorithms that can compete well with respect to all candidate algorithms from a given class such that they sequentially achieve the performance of the best algorithm in the class without any assumption on the underlying signal.
To achieve this we use ideas from universal source coding and prediction. We first use a context tree to efficiently combine a super exponential number of algorithms in order to achieve the performance of the best piecewise linear model from a class of piecewise linear models. We then extend the results of linear and nonlinear regression to a framework where the underlying competition class includes the ability to switch among the various algorithms in time. For this, we introduce a transition diagram to represent exponential number of possible transition paths with linear complexity (in data length) in order to achieve the performance of the best competitor that can select the best algorithm for each transition independently. Finally, we extend these results to universal portfolio selection problem where we compete against the best portfolio selection algorithm that can choose the best switching path and the best constant rebalanced portfolios in each region independently.
Biography: Suleyman Serdar Kozat was born in Ankara, Turkey. He received the B.S. degree in Electrical Engineering from Bilkent University, Ankara, Turkey, in 1998. He received full-time scholarship from Bilkent University during his undergraduate studies.
Until 2004, he has been in the graduate program of the Electrical and Computer Engineering Department at the University of Illinois, Urbana Champaign in the Signal Processing group, at Coordinated Science Laboratory under the supervision of Dr. Andrew C. Singer. He received the M.S. and Ph.D. degrees in 2001 and 2004 respectively. Since 2004, he is with IBM Research, Pervasive Speech Technologies group in TJ Watson Research Center, Yorktown, NY, as a full time research staff member. His research interests include machine learning, signal processing, speech processing and statistical signal processing.
He has served as a reviewer for IEEE Transactions on Signal Processing, IEEE Transactions on Image Processing, IEEE Transactions on Information Theory as well as several conferences, such as IEEE International Conference on Signal Processing (ICASSP), IEEE International Conference on Image Processing (ICIP), International Conference on Spoken Language Processing (ICSLP), IEEE International Workshop on Machine Learning for Signal Processing. He is a member of the IEEE, the Signal Processing Society, the IEEE Information Theory Society.
November 1, 2006, 13:40, G032