Addressing Imbalanced Domain-Incremental Learning through Dual-Balance Collaborative Experts

Authors: Lan Li, Da-Wei Zhou, Han-Jia Ye, De-Chuan Zhan

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results on four benchmark datasets demonstrate DCE s state-of-the-art performance.
Researcher Affiliation Academia 1School of Artificial Intelligence, Nanjing University, China 2National Key Laboratory for Novel Software Technology, Nanjing University, China. Correspondence to: Da-Wei Zhou <EMAIL>.
Pseudocode Yes Algorithm 1 Incremental Training of DCE
Open Source Code Yes Code is released at https://github.com/Lain810/DCE.
Open Datasets Yes Office Home(Venkateswara et al., 2017) (4 domains), Domain Net(Peng et al., 2019) (6 domains), CORe50(Lomonaco & Maltoni, 2017) (11 domains), and CDDB-Hard(Li et al., 2023) (5 domains).
Dataset Splits Yes we randomly select 10 samples from each class to construct the testing set and adopt (20, 60) and (20, 100) training samples as the demarcation for many-shot, medium-shot, and few-shot classes on Office-Home and Domain Net datasets respectively. For CORe50, we follow the setting in Cui et al. (2019), creating different imbalanced tasks based on varying imbalance ratios ρ = Nmax/Nmin, where Nmax represents the number of samples in the majority class and Nmin represents the number of samples in the minority class. Among the 8 training tasks, 4 are assigned an imbalance ratio of 100, and the other 4 are assigned a ratio of 50. For the CDDB-Hard dataset, we construct class-imbalanced data by specifying the number of positive and negative samples as follows: gaugan: 3000 negatives and 300 positives; biggan: 240 negatives and 1200 positives; wild: 310 negatives and 3115 positives; whichfaceisreal: 600 negatives and 120 positives; san: 130 negatives and 130 positives
Hardware Specification Yes All experiments are conducted on an RTX 3090 GPU
Software Dependencies No While Py Torch (Paszke et al., 2019) and Pilot (Sun et al., 2025) are mentioned, specific version numbers for these software packages are not provided.
Experiment Setup Yes We optimize DCE using SGD optimizer with a batch size of 128 for 20 or 30 epochs. The learning rate is set to 0.001. In the first training stage, we train the model for 20 epochs on the Office-Home, CORe50, and CDDB-Hard datasets, and for 30 epochs on Domain Net. In the second stage, the model is trained for 10 epochs on all datasets. During VPT training, we set the number of prompts to 10. The batch size is fixed at 128, and we use a cosine learning rate decay schedule with an initial learning rate of 0.01.