On the Diversity and Stability of Internal Representations in Deep Neural Networks