Increment Vector Transformation for Class Incremental Learning

Zihuan Qiu, Fanman Meng, Yi Xu, Linfeng Xu, Qingbo Wu, Hongliang Li

Keywords: Class incremental learning

Abstract:

Class Incremental Learning (CIL) presents a major challenge due to the phenomenon of catastrophic forgetting.
Recent studies on Linear Mode Connectivity (LMC) reveal that Naive-SGD oracle, trained with all historical data, connects to previous task minima through low-loss linear paths---a property generally absent in current CIL methods.
In this paper, we explore whether LMC holds for the CIL oracle. Our empirical results confirm the presence of LMC in the CIL oracle, showing that models can retain performance on earlier tasks by following the discovered low-loss linear paths. Motivated by this finding, we propose Increment Vector Transformation (IVT), which leverages the diagonal of the Fisher Information Matrix to approximate Hessian-based transformation, uncovering low-loss linear paths for incremental updates.
Our method is orthogonal to existing CIL approaches, serving as a plug-in with minor extra computational costs.
Extensive experiments on CIFAR-100, ImageNet-Subset, and ImageNet-Full demonstrate significant performance improvements when integrating IVT with representative CIL methods.

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: 11206