Continual Learning for Multiple Task Complexities

Continual learning aims to acquire knowledge from sequentially incoming tasks while mitigating the forgetting of knowledge from previous tasks. Existing continual learning methods have primarily focused on tasks with a consistent objective, such as a sequential image classification task, which limits their effectiveness when extended to tasks with different complexities in terms of objectives (e.g., modalities). This limitation makes the model prone to forgetting knowledge from relatively simple tasks when learning more complex tasks. To address this issue, we propose a continual learning framework for tasks of different complexities. We allocate a separate set of learnable parameters based on task complexity, thereby mitigating forgetting caused by overwritten knowledge from other tasks. Experimental results demonstrate that the proposed method outperforms other continual learning methods for tasks with varying complexities.