Compared with the method of automatic control theory, the self-organized control for collective motion is more robust and flexible. The strong self-organizing collective motion of biological species is related with individual hierarchical interactions, which is characterized by the asymmetrical influence of the pairwise interaction. Due to the complexity of interaction information, the construction of analytical models of hierarchical interactions is still full of challenges. Based on the deep learning technology, the experimental data of the collective motion of Hemigrammus rhodostomus fish is analysed to construct the individual interaction model with multi-parameter inputs. A deep network structure for pair interaction is designed, and the interaction model is obtained by means of proper training. Based on visual pressure, the individual identifies the key neighbour, which is used for hierarchical interaction built by deep neural networks. Compared with other neighbour selection methods, the macro characteristics are more consistent between the proposed intelligent control method and real fish collective mention experiment. Simulation shows that the proposed method can be extended to larger-scale groups for aggregation control with collective motion, so that the individual can take advantage of local information to achieve large-scale collective motion. The proposed control method is simple to use, flexible for different scale, and fast for calculation. Thus, it has broad application prospects in the fields of multi-robot control, intelligent transportation systems, saturated cluster attacks, and multi-agent logistics.