Modified kernel-based nonlinear feature extraction.
- Junshui
- Simon J.
- James P.
- Stanley
Feature Extraction (FE) techniques are widely used in many applications to pre-process data in order to reduce the complexity of subsequent processes. A group of Kernel-based nonlinear FE ( H E ) algorithms has attracted much attention due to their high performance. However, a serious limitation that is inherent in these algorithms -- the maximal number of features extracted by them is limited by the number of classes involved -- dramatically degrades their flexibility. Here we propose a modified version of those KFE algorithms (MKFE), This algorithm is developed from a special form of scatter-matrix, whose rank is not determined by the number of classes involved, and thus breaks the inherent limitation in those KFE algorithms. Experimental results suggest that MKFE algorithm is .especially useful when the training set is small.
- Research Organization:
- Los Alamos National Laboratory (LANL), Los Alamos, NM (United States)
- Sponsoring Organization:
- USDOE
- OSTI ID:
- 975954
- Report Number(s):
- LA-UR-02-0292; LA-UR-02-292; TRN: US201018%%1039
- Resource Relation:
- Conference: Submitted to: International Conference on Machine Learning and Applications 2002
- Country of Publication:
- United States
- Language:
- English
Similar Records
Scalable and efficient learning from crowds with Gaussian processes
Optimizing Kernel Machines Using Deep Learning