Homotopy continuation methods for neural networks

The application of the homotopy continuation method for finding the global minimum during the training phase of a multilayer neural network is presented. A brief description of the theory of the homotopy continuation methods is given. The backward error propagation algorithm used for training neural networks is summarized. The reformulation of the error minimization problem in the learning algorithm in a framework suitable for the continuation method and the procedure for training neural networks using the homotopy continuation method are described. Results of comparing the performance of the proposed method with traditional training methods are given