Rethinking the Influence of Distribution Adjustment in Incremental Segmentation
Keywords: Incremental segmentation, knowledge distribution
Abstract: In an ever-changing world, incremental segmentation learning faces challenges due to the need for pixel-level accuracy and the practical application of gradually obtained samples.
While most existing methods excel in stability by freezing model parameters or employing
other regularization techniques to preserve the distribution of old knowledge, these approaches often fall short of achieving satisfactory plasticity.
This phenomenon arises from the limited allocation of parameters for learning new knowledge.
Meanwhile, in such a learning manner, the distribution of old knowledge cannot be optimized as new knowledge accumulates.
As a result, the feature distribution of newly learned knowledge overlaps with old knowledge, leading to inaccurate segmentation performance on new classes and insufficient plasticity.
This issue prompts us to explore how both old and new knowledge representations can be dynamically and simultaneously adjusted in the feature space during incremental learning.
To address this, we conduct a mathematical structural analysis, which indicates that compressing the feature subspace and promoting sparse distribution is beneficial in allocating more space for new knowledge in incremental segmentation learning.
Following compression principles, high-dimensional knowledge is projected into a lower-dimensional space in a contracted and dimensionally reduced manner. Regarding sparsity, the exclusivity of multiple peaks in Gaussian mixture distributions across different classes is preserved.
Through effective knowledge transfer, both up-to-date and long-standing knowledge can dynamically adapt within a unified space, facilitating efficient adaptation to continuously incoming and evolving data.
Extensive experiments across various incremental settings consistently demonstrate the significant improvements provided by our proposed method. In particular, regarding the plasticity of in the incremental stage, our approach outperforms the state-of-the-art method by 11.7% in MIoU scores for the challenging 10-1 setting. Source code is available in the supplementary materials.
Supplementary Material: zip
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: 4345
Loading