Bounded Component Analysis: An Algorithmic Framework for Blind Separation of Independent and Dependent Sources
Department of Electrical-Electronics Engineering
Invited by: Bülent Karasözen
Place: IAM S212
Date/Time: 11.04.2017 -15.40
Abstract: In many scientific experiments and engineering applications, measurements can be modeled as linear mixtures of desired source signals, while the mixing system is unknown. The goal in Blind Source Separation (BSS) is to process these measurement sequences in an unsupervised manner to learn the inverse of the mixing system, potentially by exploiting some side information about the sources. Most popular solution to BSS problem is Independent Component Analysis (ICA) which assumes that the sources are mutuallly statistically independent. The fact that strong independence assumption may not hold in various scenarios has led to search for new frameworks that are capable of separating dependent sources.
Bounded Component Analysis (BCA) is a recent algorithmic BSS framework introduced in this direction. BCA can be considered as an extension of the ICA framework where the boundedness of sources is exploited to replace independence assumption with a weaker ``domain separability'' assumption. As a result BCA algorithms can be used to separate dependent as well as independent signals from their linear mixtures. In this talk, I'll introduce a geometric approach for developing instantaneous and convolutive BCA algorithms. Furthermore, I'll illustrate the potential benefits of the corresponding BCA algorithms through different application examples.
This is a joint work with Huseyin A. Inan. and Eren Babatas
Alper T, Erdogan received his B.S.(93) in EE from Middle East Technical University in Turkey, M.S. (95) and Ph.D.(99) in EE from Stanford University. He worked as a principal research engineer in Globespan-Virata (formerly Excess Bandwidth) in Santa Clara CA during 1999-2001. In 2002, he joined Electrical-Electronics Engineering Department of Koc University in Istanbul, Turkey where he is currently an associate professor. Dr. Erdogan served as an associate editor for IEEE Transactions on Signal Processing and as a member of IEEE Signal Processing Theory and Methods Technical Committee. He is a recipient of TUBITAK Encouragement Award, Werner Von Siemens Award and Turkish Academy of Sciences Outstanding Young Scholar Award. His research interest is on Adaptive Signal Processing, Machine Learning, Communication Systems, Computational Neuroscience and Convex Optimization Applications.