On the importance of severely testing deep learning models of cognition