Disentangling Abstraction from Statistical Pattern Matching in Human and Machine Learning