Ensemble of convolutional neural networks trained with different activation functions

Abstract Activation functions play a vital role in the training of Convolutional Neural Networks. For this reason, developing efficient and well-performing is crucial problem deep learning community. The idea these approaches to allow reliable parameter learning, avoiding vanishing gradient problems. goal work propose an ensemble Networks trained using several different activation functions. Moreover, novel function here proposed for first time. Our aim improve performance small/medium sized biomedical datasets. results clearly show that outperforms with standard ReLU as function. p-value 0.01 each tested stand-alone function; comparison we our approach on more than 10 datasets, two well-known Networks: Vgg16 ResNet50.