多任务局部进化学习——消减负迁移与参数冗余OA

Local evolutionary learning(LEL)is proposed to mitigate parameter negative transfer caused by low inter-task correlation in multi-task learning(MTL).Drawing on bionic evolutionary learning,LEL constructs local evolutionary units that incorporate inter-task correlation information to perform localized evolution and parameter updates during model training,thereby avoiding ineffective parameter interference among tasks.Upon introduction of a new task,LEL dynami-cally activates existing units according to their correlation with the new task,and applies adaptive evolutionary strategies to optimize the learning process of each subtask.Compared with existing MTL models,LEL preserves parameter sharing among highly correlated tasks while effectively reducing negative transfer between low-correlation tasks.Experimental results on multiple real-world datasets demonstrate that LEL enhances collaborative learning ability,improves generaliza-tion performance,and reduces parameter redundancy.