Research map / Label-efficient and robust learning / Domain adaptation and generalization
Multi-Granularity Class Prototype Topology Distillation for Class-Incremental Source-Free Unsupervised Domain Adaptation
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 GROTO, a source-free method for class-incremental unsupervised domain adaptation that mines positive classes, uses multi-granularity prototypes for pseudo-labels, and distills prototype topology from the source model to mitigate forgetting.
Problem
Class-incremental source-free unsupervised domain adaptation (CI-SFUDA), where unlabeled target classes arrive in sessions as a subset of the source label space without source data, causing interference from similar source classes and forgetting of old target classes.
Approach
Combines hybrid knowledge-driven positive-class mining (source-similarity and target-probability accumulation distributions) with coarse- and fine-grained class prototypes for pseudo-labeling, and a prototype topology distillation loss that aligns source and target prototype structures to reduce forgetting across sessions.
Key results
- GROTO reaches 88.8% average final accuracy on ImageNet-Caltech-CI and 82.5% on Office-Home-CI, the best among compared methods.
- On Office-31-CI, GROTO averages 93.0% final accuracy versus 91.6% for ProCA-B (ViT-B backbone, with exemplar replay).
- Ablation on Office-31-CI shows the source-only model at 79.8% rising to 93.0% with the full set of losses and modules.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add exemplar replay loss Lrep (ablation) | GROTO · Office-31-CI | accuracy: 79.8 -> 88.4 | Helped |
| Add positive-class target feature self-organization loss Lptfs (ablation) | GROTO with Lrep · Office-31-CI | accuracy: 88.4 -> 92.0 | Helped |
| Add prototype topology distillation loss Lptd (ablation) | GROTO with Lrep and Lptfs · Office-31-CI | accuracy: 92.0 -> 93.0 | Helped |
| Use coarse- and fine-grained prototypes together instead of either prototype type alone (ablation) | GROTO · Office-31-CI | accuracy: 90.7; 91.3 -> 93.0 | Helped |
| Add contrastive loss Lcon (ablation) | GROTO with Lrep, HKPCM, and PTFS · Office-31-CI | accuracy: 91.5 -> 92.0 | Helped |
| Add compactness loss Lcom (ablation) | GROTO · Office-31-CI | accuracy: 92.3 -> 93.0 | Helped |
| Add separability loss Lsep (ablation) | GROTO · Office-31-CI | accuracy: 92.4 -> 93.0 | Helped |
| Add prototype topology distillation (PTD) (ablation) | GROTO · Office-31-CI | accuracy: 92.0 -> 93.0 | Helped |
| Add positive-class target feature self-organization (PTFS) (ablation) | GROTO · Office-31-CI | accuracy: 88.0 -> 93.0 | Helped |
| Add hybrid knowledge-driven positive class mining (HKPCM) (ablation) | GROTO · Office-31-CI | accuracy: 91.0 -> 93.0 | Helped |
| Set incremental classes γ to 13 rather than 10 (ablation) | GROTO · ImageNet-Caltech-CI | accuracy: 88.8 -> 89.1 | Helped |
| Vary the initial contrastive-loss coefficient µ0c (ablation) | GROTO · ImageNet-Caltech-CI | accuracy: 88.5; 88.8; 87.5; 85.4; 88.0 | No clear effect |
Surprising findings
- Several source-free universal DA baselines (e.g., LEAD-B, DIFO) perform worse than the source-only ViT-B model on some sub-tasks, suggesting domain alignment without forgetting prevention disrupts old knowledge.
Limitations
- Relies on exemplar memory replay, which stores a small number of exemplars from target data.
- Hyperparameter sensitivity: number of incremental classes per session and prototype thresholds were chosen by heuristics.
- Evaluated only on image classification benchmarks with ViT-B backbone.
Experimental setup
Datasets: Office-31 Office-Home ImageNet-Caltech
Benchmarks: Office-31-CI Office-Home-CI ImageNet-Caltech-CI
Metrics: final accuracy session accuracy
Baselines: ViT-B source-only CIDA ProCA-B PLUE TPDS LCFD DIFO LEAD-B
Builds on: ProCA LEAD DIFO SHOT-style pseudo-labeling unsupervised contrastive learning
Closest papers by meaning
- Topology-aware Knowledge Preservation for Class-Incremental Learning (AAAI 2026 · similarity 0.87)
- Divide and Contrast: Source-free Domain Adaptation via Adaptive Contrastive Learning (NeurIPS 2022 · similarity 0.87)
- A Prototype-Oriented Framework for Unsupervised Domain Adaptation (NeurIPS 2021 · similarity 0.87)
- Unveiling the Unknown: Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation (CVPR 2024 · similarity 0.86)
- M2SD:Multiple Mixing Self-Distillation for Few-Shot Class-Incremental Learning (AAAI 2024 · similarity 0.85)
- Dynamic Target Distribution Estimation for Source-Free Open-Set Domain Adaptation (AAAI 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.