Nonlinear stretching for enhanced learning of the backpropagation algorithm in target recognition

In spite of its relatively slow learning speed, backpropagation (BP) is one of the most popular neural network training algorithms. Here, a method based on nonlinear stretching is presented that modifies the activation function in a BP algorithm to speed up the convergence in training. A target recognition system that incorporates the approach and moment invariants is formulated and tested. The test results indicate that the speed of convergence can be effectively increased through appropriate selection of a stretch factor. © 1997 John Wiley & Sons, Inc. Microwave Opt Technol Lett 16: 334–337, 1997.