Continual Learning After Model Deployment

Derda Kaymak, Gyuhak Kim, Tomoya Kaichi, Tatsuya Konishi, Bing Liu

Keywords: Open-World, Continual Learning

Abstract: This paper studies continual learning after model deployment. A real-world application environment is often an open world filled with novel or out-of-distribution (OOD) objects that have not been seen before. We can call continual learning in such an environment *open-world continual learning* (OWCL). OWCL incrementally performs two main tasks: (1) detecting OOD objects, and (2) continually learning the OOD or new objects on the fly. Although OOD detection and continual learning have been extensively studied separately, their combination for OWCL has barely been attempted. This is perhaps because in addition to the existing challenges of OOD detection and continual learning such as *catastrophic forgetting* (CF), OWCL also faces the challenge of data scarcity. As novel objects appear sporadically, when an object from a new/novel class is detected, it is difficult to learn it from one or a few samples to give good accuracy. This paper proposes a novel method called OpenLD to deal with these problems based on *linear discriminant analysis* (LDA) and a pre-trained model. This method enables OOD detection and incremental learning of the detected samples on the fly with no CF. Experimental evaluation demonstrates the effectiveness of OpenLD.

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 7678