Abstract: Based on pseudo-entropy of weights,a new method is proposed to integrate pruning algorithm with sensitivity weights.The pruning algorithm introduces the pseudo-entropy of weights as a penalty term into the normal objective function,and the distribution of weights is automatically constrained by a multilayer feed-forward neural network during the training process. The weight sensitivity is served as the simplification criteria of pruning to avoid the pruning randomicity caused by only using the weights.Meanwhile,for the problems of heavy computation burden and low efficiency of pruning algorithm in optimizing the multi-input and multi-output networks,a fast constructive algorithm is put forward,which is based on the Cascade-Correlation(CC) algorithm and constructs the new neural network from a proper network structure.The simulation results show that this fast constructive algorithm is a better choice in terms of convergence rate,computational efficiency and even generalization performance.