The study of biological collaborative intelligence can be used to inspire the regulation of artificial complex systems. However, most existing automated modelling approaches do not match the characteristics of biological collective information processing. The information mechanisms for modelling individual interaction are still full of challenges. Without loss of generality, a deep network model conforming to the biological hard attention mechanism was designed based on the collective motion data of hemigrammus rhodostomus. The designed model structure forced the individual to consider less than two neighbors information. In spite of such little information, the model could reveal the hidden place, where the high-impact neighbors located frequently. Thus, the hard attention model coincided with the information processing mechanism of biological collective interaction. The experimental results showed that the proposed hard attention model had better sparse information decoupling capability, more robust collective motion metrics and better collective scale generalization performance. The proposed model provides a powerful tool to support multi-levels behavior analysis of complex systems. Moreover, this method had strong inspirational significance for distributed controller design of swarm robotics.