Towards Robust Continual Learning: A Multi-Head Approach with Online Prototype Equilibrium and Adaptive Prototypical Feedback

Continual learning is an approach in machine learning that aims to learn different tasks sequentially, still performing well on all of them. This manner is akin to human learning. However, continual learning faces a significant challenge known as catastrophic forgetting. This refers to as a decrease in the model’s performance on previously learned tasks when learning a new task. Memory-based replay method is one approach that has been proven effective in addressing this issue. After completing each task, the model stores a small amount of data to combine with new data for training an encountering task. This approach is associated with the size of the memory buffer and data security concerns. In this paper, instead of storing actual data points, we only store prototypes representing classes in learned tasks. To this end, we also employ two techniques, i.e. Online Prototype Equilibrium (OPE) and Adaptive Prototypical Feedback (APF) to enhance the quality of prototypes’ hidden representation. Furthermore, to enhance accurate classification, we do not use a single shared head for all classes. Instead, for each task, we add a within-task prediction head and a task-ID prediction head. Experimental results on benchmark datasets demonstrate that our method outperforms several state-of-the-art methods in terms of well-studied average accuracy.