Research map / Label-efficient and robust learning / Domain adaptation and generalization
Domain Adaptation with Adversarial Training on Penultimate Activations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply adversarial training to penultimate activations rather than input images | UDA target adaptation · Office-Home | accuracy (Avg.): 66.1 -> 73.8 | Helped |
| Apply adversarial training to penultimate activations rather than input images | UDA target adaptation · DomainNet | accuracy (Avg.): 74.76 -> 83.08 | Helped |
| Add activation normalization to adversarial training (ablation) | APA · Office-Home | accuracy (Avg.): 73.8 -> 75.0 | Helped |
| Add activation normalization to adversarial training (ablation) | APA · DomainNet | per-class average accuracy (Avg.): 83.1 -> 83.9 | Helped |
| Add FixMatch loss (ablation) | APAⁿ · Office-Home | accuracy (Avg.): 75.0 -> 75.7 | Helped |
| Add perturbation projection to constrain perturbed activations to the unit sphere (ablation) | APAⁿ · DomainNet | accuracy (Avg.): 83.0 -> 83.9 | Helped |
| Freeze the classifier during adaptation (ablation) | APAᵘ · DomainNet | accuracy (Avg.): 83.08 -> 83.32 | No clear effect |
| Use top-k pseudo-label terms to construct the perturbation instead of the full objective (ablation) | APAᵘ · DomainNet | accuracy: 83.08 -> 83.44 | No clear effect |
| Use top-k pseudo-label terms to construct the normalized perturbation instead of the full objective (ablation) | APAⁿ · DomainNet | accuracy: 83.88 -> 83.54 | No clear effect |
| Apply adversarial training to normalized rather than un-normalized penultimate activations (ablation) | APA · Office-Home RS-UT | per-class average accuracy (Avg.): 67.2 vs 66.9 | Mixed |
| Apply adversarial training to normalized rather than un-normalized penultimate activations (ablation) | APA · VisDA | accuracy (Avg.): 85.2 -> 84.1 | Hurt |
| Apply adversarial training to normalized rather than un-normalized penultimate activations (ablation) | APA · DomainNet | per-class average accuracy (Avg.): 83.1 -> 83.9 | Helped |
Closest papers by meaning
- f-Domain Adversarial Learning: Theory and Algorithms (ICML 2021 · similarity 0.91)
- Adversarially Adaptive Normalization for Single Domain Generalization (CVPR 2021 · similarity 0.90)
- Adversarial Discriminative Domain Adaptation (CVPR 2017 · similarity 0.90)
- Can We Evaluate Domain Adaptation Models Without Target-Domain Labels? (ICLR 2024 · similarity 0.89)
- CASUAL: Conditional Support Alignment for Domain Adaptation with Label Shift (AAAI 2025 · similarity 0.88)
- Toward Improving Robustness and Accuracy in Unsupervised Domain Adaptation (AAAI 2025 · similarity 0.88)
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.