Search

Feature selection and dimensionality reduction is important for high dimensional signal processing and pattern recognition problems. Feature selection can be achieved by filter approach, in which certain criteria must be optimized. By using mutual information (MI) between feature vectors and class labels as the criterion, we proposed an ICA-MI framework for feature selection. In this paper, we will compare the linear ICA and local linear ICA for the accuracy of MI estimation, and study the bias-variance trade-off on feature projections and ranking. [ABSTRACT FROM AUTHOR]

Independent component analysis is an important statistical tool in machine learning, pattern recognition, and signal processing. Most of these applications require on-line learning algorithms. Current on-line ICA algorithms use the stochastic gradient concept, drawbacks of which include difficulties in selecting the step size and generating suboptimal estimates. In this paper a recursive generalized eigendecomposition algorithm is proposed that tracks the optimal solution that one would obtain using all the data observed. [ABSTRACT FROM AUTHOR]