An Optimal Brain Surgeon Procedure Algorithm for Improving Neural Network Generalization Performance

The optimal brain surgeon(OBS)procedures have higher pruning-weight accuracy and node-compression,but potential post-training operation impact its application.Based on the thought of trust region and the method of regularization,the optimal brain surgeon optimized model embeded constraint was designed by introducing constraint term into training target function.Its convergence was proved theoretically.The model was implemented by the means of Levenberg-Marquardt methods,and the simulating experiments realized the parallel of training process and OBS pruning,and supported the consistency of trust region method and Levenberg-Marquardt method theoretically.