Performance of Neural Networks Having Non-Monotonous Activation Functions

Abstract

The performance of neural networks having a non-monotonous activation function is described. The possibility to improve two difficulties, convergence to local minima and slow learning speed, is examined. Simulations are performed for the exclusive-or and the binary addition problems. The network is also applied to the acoustic diagnosis for a compressor as a practical pattern recognition task. The obtained results show that the three-layered neural networks having the non-monotonous activation function are effective for the above problems and have the same generalization performance as the three-layered network having the sigmoidal activation function.