Relaxed conditions for convergence analysis of online back-propagation algorithm with L2 regularizer for Sigma-Pi-Sigma neural network

摘要

AbstractThe properties of a boundedness estimations are investigated during the training of online back-propagation method withL2regularizer for Sigma-Pi-Sigma neural network. This brief presents a unified convergence analysis, exploiting theorems of White for the method of stochastic approximation. We apply the method of regularizer to derive estimation bounds for Sigma-Pi-Sigma network, and also give conditions for determinating convergence ensuring that the back-propagation estimator converges almost surely to a parameter value which locally minimizes the expected squared error loss. Besides, some weight boundedness estimations are derived through the squared regularizer, after that the boundedness is exploited to prove the convergence of the algorithm. A simulation is also given to verify the theoretical findings.