Learning Dynamics of Attention: Human Prior for Interpretable Machine Reasoning