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