Research map / Label-efficient and robust learning / Domain adaptation and generalization
A2XP: Towards Private Domain Generalization
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 A2XP, a privacy-preserving domain generalization method that trains per-source-domain input prompts (experts) on a frozen CLIP ViT and mixes them per image with attention-based embedders.
Problem
Domain generalization to unseen target domains typically requires modifying the architecture or parameters of the objective network, which raises privacy concerns; the paper seeks a method that keeps the network frozen and private.
Approach
Adapt visual prompts per source domain (expert adaptation), then combine them per input image via cross-attention between image embeddings and expert embeddings, with L2-normalized experts and tanh-scaled weights, without changing the network's weights.
Key results
- A2XP reaches 95.07% average target accuracy on PACS, a 4.74% gain over the best DART-augmented baseline.
- On VLCS, A2XP reaches 83.15% average accuracy, a 4.99% gain over the best compared method.
- Expert normalization contributes about 39.09% average accuracy gain in the leave-one-domain-out ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use A2XP instead of the strongest reported baseline | CLIP ViT domain generalization · PACS target domain (avg.) | accuracy: 90.33 -> 95.07 | Helped |
| Use A2XP instead of the strongest reported baseline | CLIP ViT domain generalization · VLCS target domain (avg.) | accuracy: 79.24 -> 83.15 | Helped |
| Add meta prompt initialization instead of zero initialization (ablation) | A2XP expert adaptation and generalization · Office-Home attention-based generalization (avg.) | accuracy: 65.85 -> 77.06 | Helped |
| Add meta prompt initialization instead of uniform initialization (ablation) | A2XP expert adaptation and generalization · Office-Home attention-based generalization (avg.) | accuracy: 66.20 -> 77.06 | Helped |
| Add meta prompt initialization instead of normal initialization (ablation) | A2XP expert adaptation and generalization · Office-Home attention-based generalization (avg.) | accuracy: 66.29 -> 77.06 | Helped |
| Add expert normalization (ablation) | A2XP · PACS leave-one-domain-out | accuracy: 49.35 -> 88.01 | Helped |
| Add softmax scaling of attention weights (ablation) | A2XP · PACS leave-one-domain-out | accuracy: 49.35 -> 46.96 | Hurt |
| Add tanh scaling of attention weights (ablation) | A2XP · PACS leave-one-domain-out | accuracy: 49.35 -> 57.55 | Helped |
| Add expert normalization and tanh scaling, without softmax (ablation) | A2XP · PACS leave-one-domain-out | accuracy: 49.35 -> 95.07 | Helped |
| Add softmax and tanh scaling without expert normalization (ablation) | A2XP · PACS leave-one-domain-out | accuracy: 49.35 -> 88.19 | Helped |
| Add A2XP to linear probing (ablation) | CLIP ViT · PACS target domains (avg.) | accuracy: 87.72 -> 95.07 | Helped |
| Add A2XP to full tuning (ablation) | CLIP ViT · PACS target domains (avg.) | accuracy: 38.04 -> 32.84 | Hurt |
Surprising findings
- Full fine-tuning combined with A2XP reduces accuracy (32.84% vs 38.04% for FT alone on the ablation), while A2XP with linear probing improves it substantially.
- Softmax normalization of attention weights decreased average accuracy by 4.35% compared with tanh.
Limitations
- Requires well-trained experts; some datasets are hard to adapt with input prompts.
- Full tuning combined with A2XP is unstable and counterproductive.
- Evaluation limited to PACS, VLCS and Office-Home image benchmarks with CLIP ViT backbone.
Experimental setup
Datasets: PACS VLCS Office-Home
Benchmarks: PACS VLCS Office-Home
Metrics: accuracy
Baselines: ERM SAM SagNet DANN MIRO DART
Builds on: Visual Prompting (Bahng et al.) DAM-VP DART CLIP
Closest papers by meaning
- Domain-Controlled Prompt Learning (AAAI 2024 · similarity 0.85)
- Domain Generalization in CLIP via Learning with Diverse Text Prompts (CVPR 2025 · similarity 0.85)
- PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization (CVPR 2024 · similarity 0.85)
- Rethinking the Evaluation Protocol of Domain Generalization (CVPR 2024 · similarity 0.85)
- Enhancing Domain Adaptation through Prompt Gradient Alignment (NeurIPS 2024 · similarity 0.85)
- CLIPCEIL: Domain Generalization through CLIP via Channel rEfinement and Image-text aLignment (NeurIPS 2024 · similarity 0.84)
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.