报告题目：PRS and DRS for Pattern Recognition

报告时间： 2020年11月12日，星期四，14:00-16:00；2020年11月19日，星期四，19:00-21:00

报告人：Sang-Woon Kim教授，韩国明知大学

腾讯会议： 333 318 499，密码：1112；806 817 687，密码：1112

报告摘要：

In this talk, two pattern recognition techniques, PRS (prototype reduction schemes) and DRS (dimension reduction schemes), are briefly introduced. Currently, machine learning, including deep learning, has become the main tool used in most academic/industry fields, including information technology. Sufficient amounts of data should be provided for successful learning in machine learning. On the other hand, efficient calculations require only a small amount of core data that contains only the necessary information for learning. In general, training data are prepared in a (n x d)-dimensional matrix of D. Here, the row, n, corresponds to the number of samples (or objects) and the column, d, to the number of dimensions (or variables). Both the techniques, PRS and DRS, are algorithms with which the dimension and the cardinality of the matrix D can be reduced to those of other (n' x d')-dimensional matrix of D' (where, n'<n and d'<d) while keeping the most of the essential information of D required for classification (regression, prediction, unmixing, simplification, outlier detection, variable selection, etc.) to the maximum. The goal of PRS is to obtain n'<<n, and the goal of DRS is to find d'<<d. This talk reviews the progress of various PRS and DRS algorithms, including neural network implementation, and introduces what is currently being used. In addition, after summarizing convolutional neural network (CNN) and its application, it will be addressed that CNN-based feature extractor can be applied to domains where both the dimension d and the cardinality n are very large. The details of the talk are as follows: Introduction to PRS and DRS; Various PRSs (condensed nearest neighbor, selective nearest neighbor, learning vector quantification, etc.); Various DRSs (PCA, locality preserving projection, neighborhood preserving embedding, etc.); Case studies (CNN-based extractor and its use in seismic patch classification).

报告人简介：

Sang-Woon Kim received the BE degree from Hankook Aviation University, Gyeonggi, Korea in 1978, and the ME and the PhD degrees from Yonsei University, Seoul, Korea in 1980 and 1988, respectively, both in Electronic Engineering. In 1989, he joined the Department of Computer Engineering at Myongji University and conducted research and lectures on topics related to Pattern Recognition/Visual Media Computing/Computer Organization and Architecture. Since 2019, he has continued his research as an Emeritus Professor there. His research interests include Statistical Pattern Recognition, Machine Learning, and Avatar Communications in Virtual Worlds. He is the author or coauthor of 49 regular papers and 13 books. He is a Senior Member of the IEEE and a life Member of the IEEK.