One-Pixel Shortcut: on the Learning Preference of Deep Neural Networks