Natural gradient learning algorithms for RBF networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Radial Basis Function (RBF) networks are widely used for function approximation and classification.
• Traditional RBF network learning faces challenges like slow convergence and plateaus.

Purpose of the Study:

• To enhance the learning efficiency and stability of RBF networks.
• To address the limitations of conventional gradient descent methods in RBF network training.

Main Methods:

• Introduced natural gradient learning to RBF networks, assuming Gaussian probability density functions for input and activation.
• Developed an adaptive method to efficiently compute the Fisher Information Matrix and its inverse for large-scale RBF networks.
• Formulated an explicit adaptive natural gradient learning algorithm.

Main Results:

• The proposed adaptive natural gradient method effectively avoids learning plateaus.
• Simulations demonstrate superior performance of the adaptive natural gradient method compared to conventional gradient descent for nonlinear function approximation.

Conclusions:

• The adaptive natural gradient learning algorithm offers an effective solution for improving RBF network training.
• This method enhances learning speed and stability, making RBF networks more robust for complex approximation tasks.