Comparative analysis of activation functions in neural networks

Although the impact of activations on the accuracy of neural networks has been covered in the literature, there is little discussion about the relationship between the activations and the geometry of neural network model. In this paper, we examine the effects of various activation functions on the geometry of the model within the feature space. In particular, we investigate the relationship between the activations in the hidden and output layers, the geometry of the trained neural network model, and the model performance. We present visualizations of the trained neural network models to help researchers better understand and intuit the effects of activation functions on the models.