Continual Learning (CL), also known as Lifelong Learning, refers to the process in which a model learns a sequence of tasks T_1, T_2, ..., T_n without losing knowledge about previous tasks through learning new tasks.
Formally, the problem of "Catastrophic Forgetting" is addressed, where optimization for new tasks deteriorates performance on previous tasks.
Approaches to Continual Learning can be classified into: (1) Regularization-based methods that introduce a penalty for changes to important parameters (e.g., Elastic Weight Consolidation by Kirkpatrick et al., 2017), (2) Replay methods that use previous data or generated pseudo-data to preserve memory, and (3) Architecture methods that use separate subnetworks or modular design.
The central challenge is to solve the stability-plasticity dilemma, i.e., the balance between preserving old knowledge and integrating new knowledge.
Continual Learning is essential for adaptive systems operating in dynamic environments and is actively researched in areas such as robotics, personalized assistance systems, and autonomous vehicles.
Definition:
“Continual Learning is a learning paradigm in which a model sequentially learns tasks while preserving knowledge about previous tasks to avoid Catastrophic Forgetting.”
Source:
Kirkpatrick, J., Pascanu, R., Rabinowitz, N., et al. (2017). Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13), 3521-3526.