A learning algorithm for multi-layer perceptron networks with nondifferentiable nonlinearities
E.R. Buhrke, J.L. LoCicero · 2003
A learning algorithm is proposed for neural networks with hard limiting nonlinearities. The algorithm is gradient-based, where the gradient is related to the average network response rather than to its instantaneous value. This gradient is well defined and computable. The algorithm was demonstrated on a vowel discrimination problem, where good results were achieved.>