Adaptive Learning for Robust Radial Basis Function Networks.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Signal Processing

Background:

• Least-squares estimation is common for Radial Basis Function Network (RBFN) output layer parameter estimation.
• This method is equivalent to maximum-likelihood estimation under Gaussian noise assumptions.
• Outliers in data can significantly degrade the performance of standard estimation techniques.

Purpose of the Study:

• To develop a robust estimation method for RBFN output layer parameters.
• To improve network performance in the presence of non-Gaussian noise and outliers.
• To enhance the adaptability of RBFNs to diverse noise distributions.

Main Methods:

• Proposed a novel approach using a generalized Kullback-Leibler (KL) divergence.
• Developed a surrogate-likelihood function robust to various noise distributions.
• Implemented an adaptive learning algorithm for RBFN training.

Main Results:

• The proposed method demonstrated superior performance compared to standard approaches in signal processing experiments.
• Consistent improvements in mean-squared error (MSE) were observed across artificially generated and real-world datasets.
• The algorithm showed enhanced robustness against different noise conditions and outliers.

Conclusions:

• The generalized KL divergence-based method offers a more robust alternative for RBFN parameter estimation.
• The adaptive learning algorithm effectively handles broader classes of noise, outperforming existing robust training methods.
• This work contributes to more reliable machine learning models in real-world applications with imperfect data.