On the Symmetries of Deep Learning Models and their Internal Representations