Continual Learning has many names. You may know it under the names of Incremental Learning, Lifelong Learning, or Online Learning. Those names may defines the same concept, or a slightly different concept depending of the speaker.
To put it simple: it’s the act of learning an ever-growing amount of knowledge, while trying to forget as little as possible. Look this article for more details.
One implementation difficulty in Continual Learning is how to create the stream of data. Continuum aims to solve this problem. Do not waste time anymore to reproduce the data setting, and starts directly to work on the model.
In this article, we aim to catalog some of the main settings of Continual Learning. We detail the different settings from research papers, and show how to easily code them with our library Continuum.
Three broads settings
The field of Continual Learning can be crudely split in three settings (Lomonaco and Maltoni, 2017):
- Learning new classes (NC)
- Learning new instances (NI)
- Learning new instances and new classes (NIC)
Learning New Classes
In the NC setting,