增量学习研究综述

Literature Review of Incremental Learning

增量学习是一种通过从新数据中学习来增加现有知识的范式,通常用于向现有模型添加新类或学习新域。与一次性获得所有训练数据的批量学习相比,增量学习与人类学习的过程更为相似,因此具有更高的现实意义。增量学习面临的最大问题是“灾难性遗忘”,即学习新任务忘记旧任务的现象。近年来,人们已提出大量的方法来缓解该问题,主要包括三类技术,分别是:基于回放、基于正则化及基于参数隔离。现通过对计算机视觉中增量学习的研究进展做出阐述,讨论其特点及子领域,针对当前三类主流技术,介绍其特点及代表性方法,最后总结增量学习目前存在的问题,并对研究方向做出展望。

Incremental learning is a paradigm for expanding the existing knowledge by learning from new data,which is often used to add new classes to existing models or learn new domains.Incremental learning is more similar to human learning than batch learning where obtains all training data at once,so it has more realistic significance.The biggest dilemma faced by incremental learning is"catastrophic forgetting",the phenomenon of learning new tasks and forgetting old tasks.Recently,a large number of methods have been proposed to alleviate this problem,including replay based,regularization based and parameter isolation based methods.By presenting a comprehensive review of the research progress of incremental learning in computer vision,discussing its characteristics and sub fields.And then introduces its characteristics and representative methods for the current three types of mainstream technologies.Finally,it summarizes the existing problems of incremental learning,and looks forward to the research direction.