The conventional linear back-propagation algorithm is replaced by a non-linear version, which avoids the necessity for calculating the derivative of the activation function. This may be exploited in hardware realizations of neural processors. In this paper we derive the non-linear back-propagation algorithms in the framework of recurrent back-propagation and present some numerical simulations of feed-forward networks on the NetTalk problem. A discussion of implementation in analog VLSI electronics concludes the paper.
|Journal||I E E E Transactions on Neural Networks|
|Publication status||Published - 1997|
Hertz, J., Krogh, A. S., Lautrup, B., & Lehmann, T. (1997). Non-Linear Back-propagation: Doing Back-Propagation withoutDerivatives of the Activation Function. I E E E Transactions on Neural Networks, 8(6), 1321-1327.