This paper aims to provide a review of the progress in continual learning research and its applications on unmanned platforms in recent years,with a prospectus for future research.Firstly,the concept of catastrophic forgetting and the basic setup of continual learning are introduced.Subsequently,the relevance and differences between continual learning and other similar fields are analyzed,and typical evaluation metrics for continual learning are presented.Then,a review of classic continual learning methods is provided,covering methods based on regularization,rehearsal with data integration,generative replay,and dynamic architecture approaches.Finally,the latest research achievements of continual learning methods on unmanned platforms are summarized.The review indicates that the theories and methods of continual learning require further research,with special attention needed for continual learning under conditions of sample imbalance,continual learning with small samples,semi-supervised continual learning,and online continual learning;as well as how to combine multiple continual learning methods to simulate the learning and memory mechanisms of humans.
维普
万方数据