Generalized M-sparse algorithms for constructing fault tolerant RBF networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Radial Basis Function (RBF) networks face challenges in RBF center selection and overfitting.
• Ensuring fault tolerance in trained RBF networks against noise or faults is critical for reliable performance.
• Existing algorithms often fail to address RBF node selection, overfitting, and fault tolerance simultaneously.

Purpose of the Study:

• To propose novel fault-tolerant training algorithms for RBF networks.
• To simultaneously address RBF node selection, output weight training, and fault tolerance.
• To provide explicit control over the number of RBF nodes without complex parameter tuning.

Main Methods:

• Definition of a fault-tolerant objective function incorporating a term to mitigate weight faults and noise.
• Formulation of the training process as a generalized M-sparse problem with an ℓ0-norm constraint for explicit node control.
• Introduction of the noise-resistant iterative hard thresholding (NR-IHT) algorithm and its momentum-enhanced variant (NR-IHT-Mom).

Main Results:

• The proposed algorithms effectively select RBF nodes and train output weights.
• Algorithms demonstrate improved test set performance by utilizing more RBF nodes without overfitting.
• The NR-IHT and NR-IHT-Mom algorithms show superior performance compared to state-of-the-art methods in simulations.

Conclusions:

• The developed fault-tolerant training algorithms offer a comprehensive solution to key RBF network construction issues.
• Explicit control over RBF node count and enhanced fault tolerance lead to robust network performance.
• The NR-IHT and NR-IHT-Mom algorithms represent significant advancements in RBF network training.