New method for comparing neural networks exposes how artificial intelligence works: Adversarial training makes it harder to fool the networks