Invited Speaker---Prof. Sung-Ho Kim


Department of Mathematical Sciences, KAIST, South Korea

Biography: Sung-Ho Kim is a Professor of Statistics in the Department of Mathematical Sciences at Korea Advanced Institute of Science and Technology (KAIST). He received his BA degree from Seoul National University and MA and PhD degrees in Statistics from Carnegie Mellon University. He held a position as a research scientist at Educational Testing Service (ETS), Princeton, from 1989 to 1993, where he worked on statistical modeling for problem solving. Dr. Kim has taught Statistics and Mathematics at KAIST since he joined the university in 1993 and has published over 50 research articles. The journals where his articles have appeared include Journal of the American Statistical Association, Computational Statistics and Data Analysis, Biometrika, Decision Support Systems, J. of Neuroscience Methods, and Signal Processing. He has worked as a consultant for government and private institutes on statistical technologies for educational testing, biology, and mechanic engineering in South Korea. He is an associate editor of Intelligent Data Analysis (IDA) since 2003 and did the same job for Journal of the Korean Statistical Society (JKSS) and is on the reviewer panel of the American Mathematical Society. His present research interests include large-scale modeling for graphical models, structure learning, information optimization, and statistical analysis of neuro-activity data. His other activities include innovative education for pre-college Mathematics.

Speech Title: A note on effect of dimension reduction on ANN
Abstract: We will present a method of dimension reduction for the input variables to an ANN by applying a graphical model approach. In the graphical model approach, we find a Markov blanket of variables for the output variable of the ANN. The Markov blanket variables are then used as the input variables to the ANN. This reduces the dimension of the input vector with the performance well comparable with the traditional ANN model. The performance comparison was made, in terms of prediction accuracy, in favor of the proposed method with 488 stocks in S\&P500. Discussions will be made concerning further possible improvements on the proposed method.