Research map / Label-efficient and robust learning / Test-time adaptation
DETA: Denoised Task Adaptation for Few-Shot Learning
Ji Zhang, Lianli Gao, Xu Luo, Hengtao Shen, Jingkuan Song
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add DETA image-denoising to adapter/finetuning task adaptation | TSA on URL (RN-18) · Meta-Dataset image-denoising (avg of 10) | accuracy (%): 75.5 -> 77.3 | Improved |
| add DETA to F-NCC on Swin-T | F-NCC on Swin-T · Meta-Dataset image-denoising (avg of 10) | accuracy (%): 81.3 -> 85.4 | Improved |
| add DETA label-denoising under 30% label noise | F-NCC on Swin-T · Meta-Dataset label noise 30% (avg of 10) | accuracy (%): 71.6 -> 77.3 | Improved |
| add DETA on top of state-of-the-art TSA (standard protocol, single-domain) | TSA on URL (RN-18) · Meta-Dataset standard protocol, single-domain (avg of 10) | accuracy (%): 70.7 -> 72.8 | Improved |
| add DETA to eTT on Aircraft | eTT on DINO (ViT-S) · Aircraft (Meta-Dataset) | accuracy (%): 68.1 -> 67.7 | Hurt |
| use cropped regions as plain augmentation without DETA components (ablation) | MoCo F-NCC · Meta-Dataset image-denoising (avg of 10) | accuracy (%): 73.0 -> 73.8 | No clear effect |
| add out-of-class relevance aggregation to CoRA (ablation) | DETA · Meta-Dataset image-denoising (avg of 10) | accuracy (%): 74.1 -> 74.4 | Improved |
| add global dispersion loss to CoRA + local compactness loss (ablation) | DETA · Meta-Dataset label noise 30% (avg of 10) | accuracy (%): 65.4 -> 67.1 | Improved |
| add local compactness loss to CoRA + global dispersion loss (ablation) | DETA · Meta-Dataset image-denoising (avg of 10) | accuracy (%): 75.2 -> 75.8 | Improved |
| add momentum accumulator for image weights at inference (ablation) | DETA · Meta-Dataset label noise 30% (avg of 10) | accuracy (%): 67.1 -> 68.0 | Improved |
| increase ratio of label-noisy support samples from 10% to 70% | TSA on URL (RN-18) · Meta-Dataset label noise (avg of 10) | accuracy (%): 72.8 -> 38.3 | Hurt |
Closest papers by meaning
- DiffusionMTL: Learning Multi-Task Denoising Diffusion Model from Partially Annotated Data (CVPR 2024 · similarity 0.88)
- Denoising Task Routing for Diffusion Models (ICLR 2024 · similarity 0.88)
- Task-Adaptive Prompted Transformer for Cross-Domain Few-Shot Learning (AAAI 2024 · similarity 0.87)
- Noisy Test-Time Adaptation in Vision-Language Models (ICLR 2025 · similarity 0.87)
- Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.86)
- The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models (NeurIPS 2025 · similarity 0.86)
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.
