RBF neural network center selection based on Fisher ratio class separability measure

Area of Science:

• Machine Learning
• Artificial Intelligence
• Pattern Recognition

Background:

• Radial basis function (RBF) neural networks map nonlinear data to a linear space via hidden layer neurons.
• The discriminative power of RBF networks is determined by the selection of RBF centers.
• Existing methods may not optimally select centers for maximum class separability.

Purpose of the Study:

• To propose a novel method for selecting RBF centers based on the Fisher ratio class separability measure.
• To achieve maximum discriminative power in classification tasks.
• To develop a parsimonious RBF network architecture with enhanced class separation.

Main Methods:

• Utilizing the Fisher ratio as a measure of class separability for RBF center selection.
• Employing an orthogonal transform to decouple correlations among hidden layer neuron responses.
• Implementing a multistep procedure combining Fisher ratio, orthogonal transform, and forward selection search.

Main Results:

• The proposed method effectively selects RBF centers that maximize class separability.
• The orthogonal transform allows for independent evaluation of class separability for individual RBF neurons.
• The method results in a parsimonious network architecture.

Conclusions:

• The Fisher ratio-based RBF center selection method enhances classification performance.
• This approach leads to improved feature vectors and greater class separation.
• The developed technique offers a robust strategy for designing effective RBF neural networks.