A case for the self-adaptation of activation functions in FFANNs

Abstract The universal approximation results for sigmoidal feedforward artificial neural networks do not recommend a preferred activation function. We investigated the null hypothesis that there is no preferred activation function from a class of activation functions. Our result indicates that there may exist a preferred activation that depends on the task to be solved and the specific data set used for training. The results allow us to conjecture that training algorithms that adapt the activation function may lead to faster training than those that do not.