A new learning method using local and global information for neural networks

A new learning method is proposed, which can be free from local minima of error function by using prior information. Because prior information can describe some features of teach function, neural networks also must have the features after learning. For this, learning using the prior information must attain two targets: learning of the features of teach function and a good approximation accuracy. The proposed method is very promising for solving the generalization ability problem of neural networks and avoiding the convergence to local minima. A bound on learning rate is also given for stability of the proposed method. The simulation results indicate usefulness of the proposed method.