Seminar Announcement
These events are organized by various sub-sets of the IEEE Toronto Section.
The contact person listed below is the volunteer who has arranged this event.
Please use the e-mail link provided if you have any questions, suggestions,
or concerns.
| Title
|
Minimum Noiseless Description Length (MNDL)
|
| Speaker
|
Soosan Beheshti, Ph.D.
Assistant Professor
ELCE Department
Ryerson University
|
| Day and Time
|
Wednesday, February 28, 2007, 6:00 p.m. – 7:30 p.m.
6:00 Arrival & Networking
6:30 Talk
7:30 Conclusion
NOTE: the WIE Affinity Group has scheduled a social networking event to follow this seminar.
|
| Location
|
Room BA 1240
Bahen Centre
for Information Technology
University of Toronto - St. George Campus
40 St. George Street map - code BA |
| Organizer
|
Signals & Computational Intelligence Joint Chapter |
| Contact
|
Bruno Di Stefano, E-mail:
|
| Abstract
|
The purpose of statistical modeling is to discover regularities and structure in observed data. As any finite length data set can be represented by a string of symbols from a finite alphabet, any regularity in a given data set can be used to compress the data. In this process, Occam's razor is interpreted as counseling the use of simpler models rather than complex ones and fewer symbols rather than more symbols. A well known approach to this modeling problem is Minimum Description Length (MDL). I have recently developed a new approach to the statistical modeling of noisy data denoted by Minimum Noiseless Description Length (MNDL). The main difference between these two approaches is that the conventional MDL compares the description length of the ``noisy" data, while the MNDL compares the description length of the desired "noiseless" data.
In this presentation, we review the basics of MDL approach and present the fundamentals of MNDL statistical modeling. The application of MNDL in best basis selection and compression will be presented. We will compare MNDL thresholding with existing thresholding methods with an example in wavelet image denoising and demonstrate its effective performance for frequency resolution improvement in nonparametric power spectral density (PSD) estimation. We present the advantages and drawbacks of MNDL and discuss its potential for applications in various areas. |
| Biography
|
Soosan Beheshti received the B.S. degree from Isfahan Institute of Technology, and the M.S. and Ph.D. degrees from Massachusetts Institute of Technology (MIT) in 1996 and 2002, respectively, all in electrical engineering. During her graduate studies, she was member of Digital Signal Processing Group and Laboratory for Information and Decision Systems and received the MIT EECS Carlton E. Tucker Award for Teaching Excellence. From 2002 to 2005, she was postdoctoral associate and lecturer at MIT. She has been with the ELCE Department of Ryerson University as an Assistant Professor since July 2005. Her research interests include information processing and statistical learning theory. |
|
