Research map / Label-efficient and robust learning / Domain adaptation and generalization
Concept-Based 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 CUDA, a concept-based unsupervised domain adaptation framework for concept bottleneck models that uses adversarial alignment of concept embeddings with a relaxed discriminator loss threshold, with a generalization bound, and reports gains over CBM/CEM/PCBM and UDA baselines on Waterbirds, digits, and SkinCON.
Problem
Concept Bottleneck Models assume training and test distributions match, so their concept and label accuracy degrade under domain shift; naive combinations with domain adaptation perform poorly because uniform alignment over-constrains concept distributions.
Approach
Adversarially align concept embeddings across domains, but cap the discriminator loss at a threshold tau (relaxed alignment) so that minor domain-specific concept differences are tolerated; the framework is backed by a target-error bound that links concept prediction error to the source error with ground-truth concepts.
Key results
- CUDA reaches 81.15% average class accuracy on the Waterbirds datasets versus 67.53% for CBM and 66.34% for CEM.
- On digit adaptation CUDA averages 91.25% class accuracy versus 64.03% for CEM and 58.82% for CBM.
- On SkinCON skin-tone shifts CUDA averages 78.65% class accuracy, above the best DA baseline GH++ at 76.59%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CUDA instead of the strongest listed baseline | Waterbirds domain adaptation · Waterbirds datasets | average class accuracy: 68.33 -> 81.15 | Helped |
| Use CUDA instead of the strongest listed baseline | Digit domain adaptation · Digit datasets | average class accuracy: 77.29 -> 91.25 | Helped |
| Use CUDA instead of the strongest listed baseline | SkinCON domain adaptation · SkinCON datasets | average class accuracy: 77.11 -> 78.65 | Helped |
| Add relaxed alignment by capping the discriminator loss at a threshold, instead of enforcing uniform alignment (ablation) | CUDA · Domain adaptation with concept shifts | classification accuracy | Helped |
| Intervene on predicted concepts using corrected ground-truth concepts | CUDA · Waterbirds datasets | classification accuracy | Helped |
Surprising findings
- Zero-shot and CONDA baselines collapse to near-chance accuracy on several Waterbirds shifts, e.g. CONDA 0.79% and 0.43% on Waterbirds-200 and Waterbirds-CUB.
- Naive combination of CBMs with DA is reported to perform poorly, motivating the relaxed alignment design.
Limitations
- Theory is restricted to binary classification; extension to multi-class is deferred to future work.
- Lipschitz continuity of hypotheses is assumed and its validity is discussed only in the appendix.
- Relaxation threshold tau must be chosen, and the paper gives no principled way to set it beyond the theoretical range.
Experimental setup
Datasets: Waterbirds CUB MNIST MNIST-M SVHN USPS SkinCON Fitzpatrick 17k
Benchmarks: Waterbirds-2 Waterbirds-200 Waterbirds-CUB MNIST to MNIST-M SVHN to MNIST MNIST to USPS SkinCON skin tone shift
Metrics: concept accuracy concept F1 class accuracy
Baselines: CBM CEM PCBM CONDA DANN MCD SRDC UTEP GH++
Builds on: Concept Bottleneck Models Concept Embedding Models DANN Post-hoc Concept Bottleneck Models CONDA
Closest papers by meaning
- Hybrid Concept Bottleneck Models (CVPR 2025 · similarity 0.89)
- Interpretable Generative Models through Post-hoc Concept Bottlenecks (CVPR 2025 · similarity 0.88)
- CONDA: Adaptive Concept Bottleneck for Foundation Models Under Distribution Shifts (ICLR 2025 · similarity 0.88)
- Sample-efficient Learning of Concepts with Theoretical Guarantees: from Data to Concepts without Interventions (NeurIPS 2025 · similarity 0.88)
- Concepts Worth Having: Refining VLM-Guided Concept Bottleneck Models with Minimal Annotations (NeurIPS 2026 · similarity 0.87)
- Memory-Assisted Sub-Prototype Mining for Universal Domain Adaptation (ICLR 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.