ACCEPTED TUTORIALS

Generative and Discriminative Feature Extraction

Motivation and Abstract: Although the classification performance is heavily depending upon features, the unsupervised feature extraction algorithms such as PCA, ICA, and NMF have not been optimized for the classification task. On the other hand the supervised feature extraction algorithms such as LDA have a tendency of overfitting to the training data and do not generalize well to test data. Therefore, it is important to learn good feature extraction algorithm with both discriminative and generative performance.

In this tutorial we will first review existing unsupervised feature extraction algorithms such as PCA, ICA, and NMF for the generative performance, and the supervised algorithm such as LDA to maximize discriminative performance. Then, we will introduce new approaches to combine both the generative and discriminative performance such as hybrid feature extraction, feature selection, and feature adaptation algorithms. Several experimental results will also be shown for the extracted features and classification performance.
- Unsupervised feature extraction algorithms for generative performance (PCA, ICA, and NMF)
- Supervised feature extraction algorithms for discriminative performance (LDA and CSP)
- Hybrid generative-discriminative feature extraction algorithms (discriminative ICA and discriminative NMF)
- Feature selection algorithms (Fisher discriminant score, Mutual Information, game theory, etc.)
- Feature adaptation algorithm to combine feature extractor and classifier (NMF-SLP, etc.)

PRESENTER:

Soo-Young Lee

Professor at Department of Electrical Engineering and Director at Brain Science and Technology Applications
Korea Advanced Institute of Science and Technology (KAIST)
Republic of Korea
Tel: +82-42-350-3431, Fax: +82-42-350-8490
E-mail: sylee at kaist.ac.kr