Causal Learning and Explanation of Deep Neural Networks via Autoencoded Activations