Abstract: The structure of neutron-rich nuclei in the neutron drip line region is one of the frontiers of the Radioactive Ion Beam physics. By directly detecting the neutrons emitted during their decay, the multi-neutron correlations of the nucleus can be extracted, which also provides critical information for the study of the properties of neutron-rich nuclear matter. In order to meet the requirements of conducting multi-neutron detection experiments, we developed a machine-learning-based multi-neutron recognition algorithm. We constructed a deep neural network to determine the number of incident neutrons event by event, and to further select the real neutron signals. The results of this work indicate that the detection efficiency of the machine learning algorithm is ~15%, whereas that of the traditional algorithm is ~1%. The machine learning algorithm can significantly improve four-neutron detection efficiency by more than 10 times, and is expected to be applied to multi-neutron detection experiments.