Research map / Label-efficient and robust learning / Domain adaptation and generalization
ADU: Adaptive Detection of Unknown Categories in Black-Box 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 ADU, a black-box domain adaptation framework for open-partial and open-set settings that uses selective amplification knowledge distillation and entropy-driven label differentiation with an average-entropy threshold to detect unknown target classes.
Problem
Black-box domain adaptation assumes identical label sets across domains, which fails when the target domain contains unknown classes; existing methods suffer from noisy pseudo-labels in knowledge distillation and fixed thresholds for detecting unknown categories.
Approach
Scales the knowledge distillation loss by a confidence-dependent modulating term (1+p_s)^θ to emphasize high-confidence source-API pseudo-labels, refines pseudo-labels by EMA, and separates high- and low-quality target predictions by entropy with an adaptive mean-entropy threshold, using negative learning on high-quality labels to widen the gap for unknown classes.
Key results
- In OPBDA on VisDA, ADU reaches 51.0% average H-score on the S→R-style task set versus 49.3% for UB2DA, a 3.1% gain stated in the abstract and text (text reports 3.5% for VisDA in one place).
- On OfficeHome OPBDA, ADU averages 72.2% H-score versus 70.4% for UB2DA and 64.5% for No Adapt.
- On Office31 OPBDA, ADU averages 88.1% H-score versus 86.9% for UB2DA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace UB2DA with ADU | Black-box domain adaptation · OfficeHome OPBDA (Avg.) | H-score (%): 70.4 -> 72.2 | Helped |
| replace UB2DA with ADU | Black-box domain adaptation · Office31 OPBDA (Avg.) | H-score (%): 86.9 -> 88.1 | Helped |
| replace UB2DA with ADU | Black-box domain adaptation · VisDA OPBDA | H-score (%): 45.2 -> 48.7 | Helped |
| replace UB2DA with ADU | Black-box domain adaptation · DomainNet OPBDA (Avg.) | H-score (%): 49.3 -> 51.0 | Helped |
| replace UB2DA with ADU | Black-box domain adaptation · OfficeHome OSBDA (Avg.) | H-score (%): 69.1 -> 71.1 | Helped |
| replace UB2DA with ADU | Black-box domain adaptation · Office31 OSBDA (Avg.) | H-score (%): 87.1 -> 88.1 | Helped |
| replace UB2DA with ADU | Black-box domain adaptation · VisDA OSBDA | H-score (%): 48.1 -> 50.8 | Helped |
| add negative loss to LHQ (ablation) | SAKD-only ADU · Office31 OPBDA ablation (Avg. of listed tasks) | H-score (%): 73.7 -> 74.6 | Helped |
| add entropy loss to LHQ (ablation) | SAKD-only ADU · Office31 OPBDA ablation (Avg. of listed tasks) | H-score (%): 73.7 -> 74.5 | Helped |
| add low-quality-label loss (ablation) | SAKD-only ADU · Office31 OPBDA ablation (Avg. of listed tasks) | H-score (%): 73.7 -> 74.8 | Helped |
| include all EDLD loss components (ablation) | SAKD-only ADU · Office31 OPBDA ablation (Avg. of listed tasks) | H-score (%): 73.7 -> 76.3 | Helped |
| add negative loss to LHQ (ablation) | ADU without negative loss · Office31 OPBDA ablation (Avg. of listed tasks) | H-score (%): 75.2 -> 76.3 | Mixed |
Surprising findings
- Existing BDA methods (DINE, BETA, SEAL) fall below the No Adapt. baseline on Office31 and OfficeHome in the OPBDA setting.
Limitations
- Relies on a preset-free but still heuristic mean-entropy threshold that may not fit all category-shift regimes
- Source model is a self-fine-tuned ResNet50 rather than a realistic commercial API, so evaluation is simulated
Experimental setup
Datasets: Office31 OfficeHome VisDA DomainNet
Benchmarks: Office31 OfficeHome VisDA DomainNet
Metrics: H-score
Baselines: No Adapt. DINE BETA SEAL UB2DA
Builds on: DINE UB2DA Negative Learning knowledge distillation H-score evaluation
Closest papers by meaning
- Category-Aware Active Domain Adaptation (ICML 2024 · similarity 0.91)
- Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 2025 · similarity 0.90)
- A Separation and Alignment Framework for Black-Box Domain Adaptation (AAAI 2024 · similarity 0.89)
- Reviewing the Forgotten Classes for Domain Adaptation of Black-Box Predictors (AAAI 2024 · similarity 0.89)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.87)
- Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias (CVPR 2024 · similarity 0.87)
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.