Finding Invariants in Deep Neural Networks