A Theory of Human-Like Few-Shot Learning