Efficient Representation of the Activation Space in Deep Neural Networks