Neural Networks Are Decision Trees (opens in new tab)

The author shows that any neural network having piece-wise linear activation functions can be represented as a decision tree

That's interesting. Is it still that the Rectified Linear Unit (ReLU) is the prevailing activation function in deep neural networks, because of the the vanishing gradients with activation functions like tanh? If so the conclusions from the paper would apply to a very wide range of deep neural networks.