A Regularizer Approach for RBF Networks Under the Concurrent Weight Failure Situation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Existing fault-tolerant algorithms primarily address single weight failure scenarios.
• Real-world trained networks often encounter multiple types of weight failures simultaneously.

Purpose of the Study:

• To investigate the performance degradation of radial basis function (RBF) networks under open weight faults and multiplicative weight noise.
• To develop novel fault-tolerant training algorithms for RBF networks capable of handling multiple fault types.
• To provide a method for estimating the test set error of these fault-tolerant networks.

Main Methods:

• Analysis of performance degradation caused by open weight faults and multiplicative weight noise in RBF networks.
• Definition of a new objective function for training fault-tolerant RBF networks.
• Development of batch and online learning algorithms based on the defined objective function.
• Investigation of the convergence conditions for the online learning algorithm.
• Formulation of a method to estimate the test set error for networks trained with the proposed approach.

Main Results:

• Quantification of performance impact from specific fault types (open weight, multiplicative noise).
• Successful development of two distinct learning algorithms (batch and online) for fault-tolerant RBF network training.
• Establishment of convergence criteria for the online training algorithm.
• Introduction of a practical formula for estimating test set error in faulty RBF networks.

Conclusions:

• The developed algorithms enhance the robustness of RBF networks against multiple weight failure types.
• The error estimation formula aids in optimizing network parameters, such as RBF width, for improved generalization.
• This work contributes to building more reliable neural network systems in the presence of complex fault conditions.