In the last decade, continual learning research has evolved with various methods such as parameter regularization and rehearsal strategies that help mitigate forgetting, but challenges remain when applying these solutions to real-world lifelong learning tasks.
Our work challenges the robustness of state-of-the-art continual learners, proposing a novel benchmark, the idSprites dataset, which enables the generation of thousands of tasks, reflecting true human-like learning capabilities.
Collection
[
|
...
]