Deep Neural Networks are characterized by the weight, bias, and activation function. The activation functions decide whether a neuron should be activated or not by computing weighted sums and biases. In this book, I represent an experimental review on the eight different activation functions for the Convolutional layers in Neural Networks.
For my experiment, I selected eight activation functions for three different datasets. The activation functions are – Sigmoid, Softmax, tanh, Softplus, Softsign, ReLU, ELU, and SELU. I also experimented on the networks where I did not use any activation function for the convolutional layers. After analyzing the results I see that the models with three different activation functions achieved higher performance for the three datasets. The interesting thing is, the best average performance for the three datasets is achieved by using the Softmax activation function for the Convolutional layer. At present time ReLu and ELU are the most used activation functions, but I see that tanh, Softplus, and Softsign achieved better average performance for the three datasets.
In this book, I just focus only on the activation function for the Convolutional layer to test the performance of the Convolutional Neural Networks.