On the Learnability of Physical Concepts: Can a Neural Network Understand What's Real?