Towards Interpretable Continual Learning Through Controlling Concepts
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Continual Learning, Interpretability, Catastrophic Forgetting
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Continual learning is a challenging task in machine learning as models can learn new tasks easily but suffer from catastrophic forgetting of previous tasks. In this work, we propose a novel framework called "Concept Controller" that addresses the issue of catastrophic forgetting by systematically controlling interpretable concepts in deep neural networks. Our method has several advantages: (1) High Performance: empirical results show that our method outperforms exemplar-free methods and is comparable with exemplar-based methods in the standard metrics such as average accuracy and average forgetting. Moreover, combining our method with exemplar-based methods can further improve the performance of exemplar-based methods. (2) Light: our method does not need extra memory space to store previous tasks' samples unlike the exemplar-based methods. (3) Interpretable: the procedure of controlling concept units is transparent.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4845
Loading