Research map / Label-efficient and robust learning / Domain adaptation and generalization
SVasP: Self-Versatility Adversarial Style Perturbation for Cross-Domain Few-Shot Learning
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 SVasP, a single-source cross-domain few-shot learning method that ensembles crop-level style gradients into global style gradients to form stabilized adversarial style perturbations, plus a discrepancy-and-consistency loss (DCO), reaching new SOTA on BSCD-FSL and mini-CUB benchmarks.
Problem
Single-source cross-domain few-shot learning suffers from gradient instability and convergence to poor sharp minima when adversarial style perturbations are used without access to target domains.
Approach
Aggregates localized random-crop style gradients into the global style gradient (self-versatility) to stabilize adversarial style perturbation, with random per-dimension magnitudes kappa1/kappa2, and a DCO objective that pushes adversarial features away from the source domain while keeping global-crop-adversarial predictions consistent.
Key results
- ResNet-10 1-shot average accuracy 49.26% (no finetune) vs StyleAdv 48.06%; 5-shot 65.09% vs 63.77%.
- ViT-small 5-shot finetuned average 74.42% vs StyleAdv 74.06%.
- Ablation: full SVasP reaches 65.09% vs baseline 62.07% average (5-shot, ResNet-10).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SVasP instead of StyleAdv | ResNet-10 · BSCD-FSL + mini-CUB, 5-way 5-shot, no fine-tuning | average accuracy: 63.77 -> 65.09 | Helped |
| use SVasP instead of StyleAdv | ViT-small · BSCD-FSL + mini-CUB, 5-way 5-shot, no fine-tuning | average accuracy: 72.44 -> 72.84 | Helped |
| add the SV Gradient Ensemble module to the baseline (ablation) | SVasP · BSCD-FSL + mini-CUB, 5-way 5-shot | average accuracy: 62.07 -> 62.61 | Helped |
| add domain discrepancy loss to SV (ablation) | SVasP · BSCD-FSL + mini-CUB, 5-way 5-shot | average accuracy: 62.61 -> 63.69 | Helped |
| add crop consistency loss to SV (ablation) | SVasP · BSCD-FSL + mini-CUB, 5-way 5-shot | average accuracy: 62.61 -> 64.05 | Helped |
| add both domain discrepancy and crop consistency losses (ablation) | SVasP · BSCD-FSL + mini-CUB, 5-way 5-shot | average accuracy: 62.07 -> 65.09 | Helped |
| set crop count k to 2 rather than 0 or 1 (ablation) | SVasP · BSCD-FSL + mini-CUB, 5-way 5-shot | accuracy | Helped |
| increase decay factor xi from 0 to 0.1 (ablation) | SVasP · BSCD-FSL + mini-CUB, 5-way 5-shot | average accuracy: 62.77 -> 65.09 | Helped |
| set decay factor xi to 1 instead of 0.1 (ablation) | SVasP · BSCD-FSL + mini-CUB, 5-way 5-shot | average accuracy: 65.09 -> 63.42 | Hurt |
| set consistency weight lambda to 0.2 instead of 0 (ablation) | SVasP · BSCD-FSL + mini-CUB, 5-way 5-shot | average accuracy: 63.79 -> 65.09 | Helped |
| set consistency weight lambda to 1 instead of 0.2 (ablation) | SVasP · BSCD-FSL + mini-CUB, 5-way 5-shot | average accuracy: 65.09 -> 63.67 | Hurt |
| allow kappa1 and kappa2 to differ instead of forcing the same value (ablation) | SVasP · BSCD-FSL + mini-CUB, 5-way 5-shot | average accuracy: 63.87 -> 65.09 | Helped |
Surprising findings
- Crop number k=2 is optimal; more crops (3-5) degrade accuracy, attributed to overfitting to source-domain style.
- Gains over StyleAdv on ViT-small are small (0.84% 1-shot, 0.40% 5-shot average).
Limitations
- Gains over StyleAdv on ViT-small are marginal.
- Evaluated only on single source miniImageNet with a fixed set of benchmark datasets.
- Many hyperparameters (xi, lambda, crop scale, kappa set) tuned on the benchmark results shown.
Experimental setup
Datasets: miniImageNet
Benchmarks: BSCD-FSL (ChestX, ISIC, EuroSAT, CropDisease) mini-CUB (CUB, Cars, Places, Plantae)
Metrics: accuracy average accuracy
Baselines: GNN FWT ATA SET-RCL StyleAdv Fine-tune BSR NSAE RDC PMF
Builds on: StyleAdv AdaIN GNN ProtoNet FWT
Closest papers by meaning
- Generate Universal Adversarial Perturbations for Few-Shot Learning (NeurIPS 2024 · similarity 0.86)
- Style Adaptation and Uncertainty Estimation for Multi-Source Blended-Target Domain Adaptation (NeurIPS 2024 · similarity 0.85)
- Flatten Long-Range Loss Landscapes for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.84)
- Mixture of Adversarial LoRAs: Boosting Robust Generalization in Meta-Tuning (NeurIPS 2024 · similarity 0.84)
- Few-Shot Adversarial Prompt Learning on Vision-Language Models (NeurIPS 2024 · similarity 0.84)
- StyDeSty: Min-Max Stylization and Destylization for Single Domain Generalization (ICML 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.