Continuous learning / life long learning

         Continuous learning , Generalized knowledge Ability is an important feature of strong artificial intelligence . The important research goal of artificial intelligence is to build Continuous learning Capable agent , The agent can build a complex understanding of the world based on its own experience , Thus using more complex knowledge Autonomous and progressive Training and improvement .

         The learning process of deep learning model is usually a fixed data set in local or isolated tasks , This makes it difficult for the model to produce more complex and autonomous intelligent behavior . Continuous learning （Continual Learning）, Also known as lifelong learning （Life-long learning） It is the research direction to solve such problems , Its goal is to extend the adaptability of the model , So that the model can learn the knowledge of different tasks at different times , At the same time, the characteristics of previous tasks will not be forgotten .

         There are two goals for continuous learning on deep Neural Networks ： One is to deal with the natural existence of neural network due to its own design Catastrophic forgetting (McCloskey and Cohen, 1989), The second is to make the training model more general , Even if the model has Plasticity （ The ability to learn new knowledge ） And stability （ The ability to remember old knowledge ）.

         The difficulty of a deep learning system that can continue to learn is to design a system that is sensitive to new input but will not be disrupted . The current mainstream continuous learning methods for neural network models can be divided into the following categories ：

• Regularization： Add restrictions when updating network parameters , So that the network does not affect the previous knowledge when learning new tasks . In this kind of method , The most typical algorithm is EWC.
• EWC https://github.com/GMvandeVen/continual-learning
• Ensembling： When the model learns new tasks , Add a new model （ It can be a display or an implicit method ）࿰