Research map / Continual learning / Domain and prompt-based
Addressing Imbalanced Domain-Incremental Learning through Dual-Balance Collaborative Experts
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Dual-Balance Collaborative Experts (DCE), a frequency-aware multi-expert framework with a dynamic expert selector trained on Gaussian-sampled pseudo-features, for imbalanced domain-incremental learning with pre-trained ViTs.
Problem
Domain-incremental learning under class imbalance, where intra-domain class imbalance causes underfitting of few-shot classes and cross-domain class distribution shifts cause forgetting of many-shot classes while limiting transfer to few-shot classes.
Approach
Three frequency-aware experts trained with CE, balanced softmax, and an inverse-distribution loss, plus a dynamic MLP expert selector trained on balanced Gaussian samples from stored per-class feature statistics of a frozen pre-trained ViT, enabling soft cross-domain expert fusion without replaying data.
Key results
- On DomainNet (ViT-B/16 IN1K, five task orders), DCE reaches 64.3 average and 63.5 last accuracy versus 59.0 / 57.9 for the runner-up S-iPrompt.
- DCE achieves 79.4 few-shot accuracy on Office-Home versus 73.7 for the best baselines (SimpleCIL, L2P), with 84.6 average accuracy.
- On CORe50 DCE reaches 80.1 average and 84.8 last accuracy, and on CDDB-Hard 74.6 average and 71.8 last accuracy, outperforming all compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DCE instead of the strongest listed baseline | Office-Home · Office-Home | average accuracy: 83.4 vs 84.6 | Helped |
| Use DCE instead of the strongest listed baseline | Office-Home · Office-Home | last accuracy: 83.5 vs 84.4 | Helped |
| Use DCE instead of the strongest listed baseline | DomainNet · DomainNet | average accuracy: 59.0 vs 64.3 | Helped |
| Use DCE instead of the strongest listed baseline | DomainNet · DomainNet | last accuracy: 57.9 vs 63.5 | Helped |
| Use DCE instead of the strongest listed baseline | CORe50 · CORe50 | average accuracy: 76.7 vs 80.1 | Helped |
| Use DCE instead of the strongest listed baseline | CORe50 · CORe50 | last accuracy: 81.7 vs 84.8 | Helped |
| Use DCE instead of the strongest listed baseline | CDDB-Hard · CDDB-Hard | average accuracy: 67.9 vs 74.6 | Helped |
| Use DCE instead of the strongest listed baseline | CDDB-Hard · CDDB-Hard | last accuracy: 66.6 vs 71.8 | Helped |
| Add a balanced-softmax expert to a single cross-entropy expert (ablation) | Frequency-aware expert design on DomainNet · DomainNet | overall accuracy | Helped |
| Add an inverse-distribution-loss expert to the two-expert setup (ablation) | Frequency-aware expert design on DomainNet · DomainNet | overall accuracy | Helped |
| Add a fourth expert with the additional frequency-weighted loss (ablation) | Three-expert DCE on DomainNet · DomainNet | overall accuracy | Helped |
| Use cross-entropy loss rather than inverse-distribution loss for the single expert (ablation) | Single-expert configuration on DomainNet · DomainNet | many-shot accuracy | Helped |
Surprising findings
- Naively re-implementing baselines with balanced softmax degrades some of them, per the authors.
- DUCT, a model-merging DIL method, performs poorly in this imbalanced setting (18.0 last accuracy on DomainNet).
Limitations
- Stored per-class statistics and per-task experts grow memory with the number of tasks (selector size scales with 3b).
- Gains from more than three experts are marginal, so the three-expert design is a fixed choice rather than adaptive.
- Covariance estimation for classes with few samples is unreliable, requiring OAS shrinkage and domain-level averaging approximations.
Experimental setup
Datasets: Office-Home DomainNet CORe50 CDDB-Hard
Benchmarks: Office-Home DomainNet CORe50 CDDB-Hard
Metrics: average accuracy last accuracy many-shot accuracy medium-shot accuracy few-shot accuracy class performance drift
Baselines: iCaRL MEMO SimpleCIL RanPAC L2P DualPrompt CODA-Prompt S-iPrompt ROW DUCT
Builds on: L2P S-iPrompt balanced softmax Visual Prompt Tuning Oracle Approximating Shrinkage
Closest papers by meaning
- Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · similarity 0.89)
- Advancing Cross-domain Discriminability in Continual Learning of Vision-Language Models (NeurIPS 2024 · similarity 0.86)
- Defying Imbalanced Forgetting in Class Incremental Learning (AAAI 2024 · similarity 0.86)
- Dual Consolidation for Pre-Trained Model-Based Domain-Incremental Learning (CVPR 2025 · similarity 0.85)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.85)
- Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.