"Bayes Optimal Classification and Feature Extraction"
Dr. Onur Hamsici
Qualcomm Research, San Diego
Friday, September 19, 2014 · 11:00AM · HEC 101
Recent advances in technology allow most measurement devices to capture abundant digital data by densely sampling the analog world, e.g. digital cameras usually
capture high-resolution images. Similarly, scientific experiments usually generate vast amount of data, e.g. in medicine patients are subjected over an extensive
number of tests or are investigated for a large number of gene expression levels to find the causes of a sickness. Manual analysis of these data sets is impossible,
while an automatic analysis of the raw data is computational and very sensitive to variations in the imaging or measurement conditions. My research focuses on
algorithms that can extract robust features from various types of data, learn from supervised or unsupervised data labels and make accurate estimations on the
future observations. For instance, in classification of spherical data, literature has been avoiding the use of spherical distributions due to required approximations
and computational difficulties. We have addressed this problem with the theory of spherical-homoscedasticity that defines the conditions to achieve the performance of
the Bayes optimal classification accuracy for spherical distributions with a simple linear classifier. We have generalized this theory to nonlinear problems with the
definition of Rotation Invariant Kernels (RIK). RIK allows one to obtain effective nonlinear models for the shape variations in Active Appearance Models and Non-Rigid
Structure from Motion problems. A key component in these and other vision algorithms is to extract the most significant features. The complexity of this problem has
left the literature to suboptimal algorithms. We have derived the Bayes optimal feature extraction algorithm, which can extract the smallest number of features (1 or
2 most important features) that has the minimum classification error. We have shown that this algorithm outperforms the literature in various high-dimensional inference
problems such as face and object recognition.
Onur C. Hamsici received the BS degree in Electrical and Electronics Engineering and a minor degree in Mechatronics in Mechanical Engineering both from Middle East
Technical University, Ankara, Turkey in 2003. He received the MS and PhD degrees in Electrical and Computer Engineering from The Ohio State University (OSU), in 2005
and 2008, respectively. He was a Postdoctoral Researcher at OSU and Visiting Faculty at Qualcomm in 2009. He is currently a researcher at the Office of the Chief
Scientist at Qualcomm Research, San Diego. His research interests are statistical pattern recognition, machine learning, and vision.