Make Domain Shift a Catastrophic Forgetting Alleviator in Class-Incremental Learning
Authors: Wei Chen, Yi Zhou
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our comprehensive studies demonstrate that incorporating domain shift leads to a clearer separation in the feature distribution across tasks and helps reduce parameter interference during the learning process. ... Experimental results show that incorporating our method into various CIL methods achieves substantial performance improvements... These findings illustrate that Dis Co can serve as a robust fashion for future research in class-incremental learning. ... 5 Experiments |
| Researcher Affiliation | Academia | Wei Chen1,2, Yi Zhou1,2 * 1School of Computer Science and Engineering, Southeast University, China 2Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications, Ministry of Education, China EMAIL, EMAIL |
| Pseudocode | No | The paper describes the method using equations and text descriptions in Section 4 'Method', but does not contain a dedicated pseudocode or algorithm block. |
| Open Source Code | Yes | Code https://github.com/Pixel Chen24/Dis Co |
| Open Datasets | Yes | We perform experiments on CIFAR100, Fashion-MNIST, and Tiny-Image Net. Details of these datasets and continual task split are in Appendix C.1. ... We start with an empirical study where we simulate the combination of CIL and domain shift by splitting Domain Net (Peng et al. 2019) dataset, or manually add domain shift using a style transfer GAN to the original CIFAR-100 dataset (Krizhevsky and Hinton 2009) to construct Domain CIFAR-100. |
| Dataset Splits | Yes | Details of these datasets and continual task split are in Appendix C.1. ... Fig. 5: Ablation study on incremental task length. B{X}{Y } means there are X classes in task 0 and the rest are evenly distributed in Y tasks. |
| Hardware Specification | No | No specific hardware details (like GPU or CPU models, or memory amounts) are provided for the experimental setup in the main text. |
| Software Dependencies | No | The paper mentions 'Our code is implemented in Py Torch and based on LAMDA-PILOT(Sun et al. 2023)', but does not specify version numbers for these software dependencies. |
| Experiment Setup | Yes | For both Res Net (He et al. 2016) and Vi T (Dosovitskiy et al. 2020) -based models, we train all tasks for 100 epochs, 60th and 80-th epochs being milestones. For models built on Res Net, we set weight decay w = 5e 4 and learning rate lr = 0.1 with 0.1 at milestones. For Vi T-based models, w = 2e 4, lr = 1e 3 with 0.1 at milestones. |