Generic Back-Propagation in Arbitrary FeedForward Neural Networks
In this paper, we describe a general mathematical model for feedforward neural networks. The final form of the network is a vectorial function f of two variables, x (the input of the network) and w (the weight vector). We show that the differential of f can be computed with an extended back-propagation algorithm or with a direct method. By evaluating the time needed to compute the differential with the help of both methods, we show how to chose the best one. We introduce also input sharing and output function which allow us to implement efficiently a multilayer perceptron with our model.
KeywordsNeural Network Weight Vector Input Space Feedforward Neural Network Weight Space
Unable to display preview. Download preview PDF.
- Léon Bottou and Patrick Gallinari. A Framework for the Cooperation of Learning Algorithms. In R.P. Lippmann, J.E. Moody, and D.S. Touretzky, editors, Neural Information Processing Systems, volume 3, pages 781–788. Morgan Kauffman, 1991.Google Scholar
- Cédric Gégout, Bernard Girau, and Fabrice Rossi. Les réseaux d’opérateurs. Technical report, Thomson- CSF/SDC, Juillet 1993.Google Scholar
- Cédric Gégout, Bernard Girau, and Fabrice Rossi. NSK, an Object-Oriented Simulator Kernel for Arbitrary Feedforward Neural Networks. In Int. Conf on Tools with Artificial Intelligence, pages 93–104, New Orleans (Louisiana), November 1994. IEEE.Google Scholar