Machine Learning for Intelligent Identification of Information


B.H. Juang

Georgia Institute of Technology


Monday, June 16

4:00-5:00 pm

Room 1005 EECS



Abstract

Modern informatics has started to address the issue of intelligence, which in the lineage of information theory (as opposed to artificial intelligence) can be traced back to Harry Nyquist. In Shannon's information theory, nevertheless, information is a measure arisen from uncertainty, or probability, and is mostly independent of the notion of significance. To be able to handle intelligence in informatics, one must consider, along with statistics and probability, the component of significance for various tasks such as information classification and identification. In this talk, we discuss how information of varying significance can be classified in a statistical approach, consistent with Bayes' teaching. We embed the notion of significance in the cost of classification error; i.e., misclassifying information of higher signifycance would incur a heavier cost. We propose a particular methodology, called performance-based learning, which allows active learning from labeled information patterns to minimize the cost of error. We elaborate formalism for the proposed machine-learning scheme in contrast to some earlier heuristics involved in cost-sensitive learning.


Biography

Biing Hwang (Fred) Juang joined Georgia Institute of Technology in 2002 after two decades of career with Bell Laboratories. Prof. Juang has published extensively, including the book "Fundamentals of Speech Recognition", co-authored with L.R. Rabiner. He has served as Editor-in-Chief for the IEEE Transactions on Speech and Audio Processing, and a number of positions in the IEEE Signal Processing Society. He is a Fellow of the IEEE, a Fellow of Bell Laboratories, a member of the US National Academy of Engineering, an Eminent Scholar in the Eminent Scholar Academy of Georgia, and an Academician of the Academia Sinica (Taiwan).