Research map / Continual learning / Domain and prompt-based
DualCP: Rehearsal-Free Domain-Incremental Learning via Dual-Level Concept Prototype
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 DualCP, a rehearsal-free domain-incremental learning method that builds coarse- and fine-grained concept prototypes (ETF-based, from CLIP text features) and aligns image features to them via a coarse-to-fine calibrator trained with a dual dot-regression loss.
Problem
Rehearsal-free domain-incremental learning must balance learning new domains against forgetting old ones; existing methods treat this as a zero-sum trade-off and features from different domains crowd the same space.
Approach
Treat same-class images across domains as one concept, build simplex-ETF concept prototypes from class-name text features, and hierarchically group classes by semantic similarity so coarse prototypes separate groups and fine prototypes separate classes within groups; a C2F calibrator aligns image features to these prototypes with a DDR loss.
Key results
- On DomainNet, DualCP reaches 60.13% A_T and -1.96 F_T, beating the best rehearsal-free C-Prompt (58.68%) with ViT-B/16.
- On CDDB Hard, DualCP reports 82.16% A_T versus 74.51% for S-iP among rehearsal-free methods; the text says 82.46%, inconsistent with the table.
- With a CLIP image encoder and CLIP text guidance, DualCP reaches 72.46% on DomainNet and 92.34% on CDDB, and 90.59% on CORe50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DualCP instead of C-Prompt | RFDIL on DomainNet · DomainNet | average accuracy: 58.68 -> 60.13 | Helped |
| Use DualCP instead of S-iP | RFDIL on CDDB HARD · CDDB HARD | average accuracy: 74.51 -> 82.16 | Helped |
| Use DualCP instead of CODA-P | RFDIL on CORe50 · CORe50 | average accuracy: 85.68 -> 88.10 | Helped |
| Replace single-level prototypes with dual-level prototypes (ablation) | DualCP framework · DomainNet | average accuracy: 56.22 -> 60.13 | Helped |
| Replace single-level prototypes with dual-level prototypes (ablation) | DualCP framework · CDDB HARD | average accuracy: 82.16 -> 82.16 | No clear effect |
| Replace single-level prototypes with dual-level prototypes (ablation) | DualCP framework · CORe50 | average accuracy: 86.27 -> 88.10 | Helped |
| Set DDR loss α to 0.5 rather than 0.1, 0.25, 0.75, or 0.9 (ablation) | DualCP on DomainNet · DomainNet | average accuracy: 58.92, 59.75, 60.13, 59.32, 58.06 | Helped |
| Use CLIP-Text rather than ViT-B/16-BD guidance with ViT-B/16 image features (ablation) | DualCP · DomainNet | average accuracy: 60.13 -> 62.73 | Helped |
| Use CLIP-Text rather than ViT-B/16-BD guidance with ViT-B/16 image features (ablation) | DualCP · CDDB HARD | average accuracy: 82.16 -> 83.05 | Helped |
| Use CLIP-Text rather than ViT-B/16-BD guidance with ViT-B/16 image features (ablation) | DualCP · CORe50 | average accuracy: 88.10 -> 88.55 | Helped |
| Replace ViT-B/16 image features with CLIP-Image features while retaining ViT-B/16-BD guidance (ablation) | DualCP · DomainNet | average accuracy: 60.13 -> 69.31 | Helped |
| Use CLIP-Image features and CLIP-Text guidance (ablation) | DualCP · DomainNet | average accuracy: 60.13 -> 72.46 | Helped |
Surprising findings
- Table 2 reports DualCP with ViT-B/16 at 82.16% on CDDB, which is the same number as the VanillaCP entry, and the text claims 82.46%, so reported numbers are internally inconsistent.
- Several reported DualCP gains are obtained with a CLIP backbone, and S-Prompts with CLIP (67.78% on DomainNet) is far above the ViT-only DualCP result, so much of the gain tracks backbone choice.
Limitations
- DualCP cannot be applied to CDDB's two-class setting, so the dual-level design is unused there.
- Grouping depends on a fixed similarity threshold p (0.85) and on CLIP text features, with gains tied to the backbone choice.
- Theoretical result (Theorem 1) only compares angles between prototypes, not end-to-end generalization.
Experimental setup
Datasets: DomainNet CDDB CORe50
Benchmarks: DomainNet CDDB (Hard track) CORe50
Metrics: average accuracy A_T forgetting degree F_T
Baselines: DyTox DARE EWC LwF SimCLR BYOL Barlow Twins SupCon L2P DualPrompt S-iP CODA-P
Builds on: neural collapse simplex ETF CLIP S-Prompts
Closest papers by meaning
- Advancing Cross-domain Discriminability in Continual Learning of Vision-Language Models (NeurIPS 2024 · similarity 0.86)
- Dual Consolidation for Pre-Trained Model-Based Domain-Incremental Learning (CVPR 2025 · similarity 0.85)
- Dual Prototype Evolving for Test-Time Generalization of Vision-Language Models (NeurIPS 2024 · similarity 0.84)
- Gradual Divergence for Seamless Adaptation: A Novel Domain Incremental Learning Method (ICML 2024 · similarity 0.84)
- Addressing Imbalanced Domain-Incremental Learning through Dual-Balance Collaborative Experts (ICML 2025 · similarity 0.83)
- D3T: Distinctive Dual-Domain Teacher Zigzagging Across RGB-Thermal Gap for Domain-Adaptive Object Detection (CVPR 2024 · similarity 0.83)
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.