Advances and Trends of Continual Learning

With the development and successful application of deep learning, continual learning has attracted increasing attention and has been a hot topic in the field of machine learning, especially in the resource-limited and data-security scenarios with the increasing requirements of quickly learning sequential tasks and data. Different from humans who enjoy the ability of continually learning and transferring knowledge, the existing deep learning models are prone to easily suffering from a catastrophic forgetting problem in a sequential learning process. Therefore, how to continually learn new knowledge and retain old knowledge at the same time on dynamic and non-stationary sequential task and streaming data, is the core of continual learning. Firstly, through the investigation and summary of the related work of continual learning at home and abroad in recent years, continual learning methods can be roughly divided into three categories: replay-based, constraint-based, and architecture-based. We will further subdivide these three types of methods. Specifically, the replay-based methods are subdivided into three categories: sample replay, generation replay, and pseudo-sample replay, according to the sample’s sources used; the constraint-based methods are subdivided into parameter constraints, gradient constraints, and data constraints, according to the constraint’s sources; the architecture-based methods are subdivided into two categories: parameter isolation and model expansion, according to how the model structure is used. By comparing the innovation points of related works, the advantages and disadvantages of various methods are summarized. Secondly, the research progress at home and abroad is analyzed. Finally, the future development direction of continual learning combined with other fields is simply prospected.