Rebalancing Multi-Label Class-Incremental Learning
Authors: Kaile Du, Yifan Zhou, Fan Lyu, Yuyang Li, Junzhou Xie, Yixi Shen, Fuyuan Hu, Guangcan Liu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our comprehensive experiments on the PASCAL VOC and MS-COCO datasets demonstrate that this rebalancing strategy significantly improves performance, achieving new state-of-the-art results even with a vanilla CNN backbone. Extensive experiments conducted across multiple MLCIL scenarios using the PASCAL VOC and MS-COCO demonstrate that Reb LL achieves new SOTA results. The results for PASCAL VOC in {B4-C2 and B5-C3} are presented in Table 1. The results for MS-COCO in {B20-C4 and B0-C5} are presented in Table 2. In Table 4, we use Fine-Tuning (Lbce) and KD (Lbce + Lkd) as baselines to validate the effectiveness of AKD in scenario {B4-C2} of PASCAL VOC. As shown in Table 5, we conduct ablation of OR based on AKD (Lcls + Lakd) in scenario {B4-C2} of PASCAL VOC. |
| Researcher Affiliation | Academia | 1School of Automation, Southeast University, China 2Key Laboratory of Measurement and Control of CSE, Ministry of Education, China 3NLPR, MAIS, CASIA, China 4School of Electronic and Information Engineering, Suzhou University of Science and Technology, China 5Suzhou Key Laboratory of Intelligent Low Carbon Technology Application, China 6Jiangsu Industrial Intelligent Low Carbon Technology Engineering Center, China |
| Pseudocode | No | The paper describes its methodology using textual explanations and mathematical formulations, but it does not include any explicitly labeled pseudocode blocks or algorithms. |
| Open Source Code | Yes | Code https://github.com/Kaile-Du/Reb_LL |
| Open Datasets | Yes | We adopt the experimental framework from (Dong et al. 2023; Du et al. 2024b), utilizing the MS-COCO 2014 (Lin et al. 2014) and PASCAL VOC 2007 (Everingham et al. 2010) datasets to validate the efficacy of our method. |
| Dataset Splits | Yes | For the VOC 2007 dataset, we use two challenging scenarios {B4-C2 and B5-C3}. Similarly, for the MS-COCO dataset, we assess Reb LL with two challenging scenarios {B20-C4 and B0-C5}. The CIL process adheres to the lexicographical order of class names, as described in (Dong et al. 2023; Du et al. 2024b). |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It mentions using 'TRes Net M' and 'vanilla CNN' as backbones but not the underlying hardware used for training or inference. |
| Software Dependencies | No | The paper mentions optimizing the network using the Adam optimizer (Kingma and Ba 2015) and using TRes Net M (Ridnik et al. 2021b) as a feature extractor, but it does not specify version numbers for any key software components like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | The training is conducted with a batch size of 64 for 20 epochs on both the PASCAL VOC and MS-COCO datasets. We optimize the network using the Adam optimizer (Kingma and Ba 2015) with a weight decay of 1e-4. We apply a consistent learning rate of 4e-5 across all tasks for the PASCAL VOC dataset. For the MS-COCO dataset, we use an initial learning rate of 3e-5 for the base task and 5e-5 for subsequent tasks. We employ the same data augmentation methods as detailed in (Du et al. 2024b; Dong et al. 2023). Figure 5 presents an analysis of hyperparameters α and β in {B4-C2} of PASCAL VOC. The balancing parameters λakd and λer are set to 0.15 and 0.30, respectively. |