Continual Learning (CL) is a learning paradigm in machine learning for multiple tasks. Although deep neural network models perform very well in many tasks, they often suffer from severe interference and forgetting of previously learned knowledge when updating parameters for new tasks. Continual learning is the study of how to solve these interference and forgetting issues by simulating the learning process of the brain. The characteristic of continual learning lies in efficiently transforming and utilizing previously learned knowledge to complete the learning of new tasks, significantly reducing the problems caused by forgetting, and is of great significance for intelligent computing systems to adaptively adapt to environmental changes. This forum will invite four experts in relevant fields to delve into continual learning from the advances in incremental learning, knowledge-guided continual learning, continuous image recognition under micromemory, and continual learning for real-world visual perception, hoping to provide more inspiration and promote the development of the field.
持续学习是多个任务的机器学习的一种学习范式。尽管深度神经网络模型在很多任务中都表现得非常好，但当模型在新的任务中进行参数更新时，往往会经受非常严重的干扰以及遗忘之前学习过的知识。而持续学习就是研究怎样通过模拟大脑学习的过程解决这种干扰和遗忘问题。持续学习的特点在于高效地转化和利用已经学过的知识来完成新任务的学习，并且能够极大程度地降低遗忘带来的问题，对智能计算系统自适应地适应环境改变具有重要的意义。本论坛将邀请四位相关领域的专家从增量学习的优势、知识引导的持续学习、为存储器下的连续图像识别、持续学习下的现实世界视觉感知等方面深入探讨持续学习，期望提供更多的启迪和推动领域的发展。