Radial basis function neural networks for nonlinear Fisher discrimination and Neyman-Pearson classification

Area of Science:

• Machine Learning
• Artificial Intelligence
• Pattern Recognition

Background:

• Radial Basis Function (RBF) neural networks (NNs) are widely used for classification tasks.
• Optimizing RBF NN parameters for specific classification goals remains a challenge.
• Existing methods often focus on overall classification accuracy, potentially neglecting performance on critical data subsets.

Purpose of the Study:

• To propose a novel technique for designing RBF NNs.
• To enhance classification performance for specific data classes.
• To approximate Neyman-Pearson classification and enable controlled performance.

Main Methods:

• Utilizing class membership information of training samples to create new cluster classes.
• Adjusting RBF parameters based on these new cluster classes.
• Analyzing the hidden-to-output layer for Fisher discrimination analysis and the full system for nonlinear Fisher analysis.

Main Results:

• The proposed method allows for emphasis on classification performance for certain class data over overall classification.
• The RBF hidden to output layer performs Fisher discrimination analysis when output neuron levels are properly chosen.
• The full RBF system performs a nonlinear Fisher analysis.
• Effectiveness confirmed on an agricultural product inspection problem and synthetic data.

Conclusions:

• The novel RBF NN design technique effectively improves targeted classification performance.
• The method offers control over classification performance and approximates Neyman-Pearson classification.
• The underlying mechanisms involve Fisher discrimination analysis, providing theoretical grounding.