Research map / Vision-language and multimodal / CLIP and zero-shot
Generate, Transduct, Adapt: Iterative Transduction with VLMs
Oindrila Saha, Logan Lawrence, Grant Van Horn, Subhransu Maji
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 Adapt (fine-tune CLIP) after dynamic-attribute transduction (ablation) | GTA-CLIP (ViT-B/16) · 5-dataset average | top-1 accuracy: 65.56 -> 67.52 | Improved |
| add dynamic (confusion-driven) attributes instead of static ones, with transduct and adapt (ablation) | GTA-CLIP (ViT-B/16) · 5-dataset average | top-1 accuracy: 66.76 -> 67.52 | Improved |
| use dynamic instead of static attributes without Adapt (ablation) | GTA-CLIP (ViT-B/16) · 5-dataset average | top-1 accuracy: 65.60 -> 65.56 | No clear effect |
| add Transduct to inductive CLIP (ablation) | GTA-CLIP (ViT-B/16) · 5-dataset average | top-1 accuracy: 60.56 -> 64.26 | Improved |
| add Adapt to static-attribute transduction (ablation) | GTA-CLIP (ViT-B/16) · 5-dataset average | top-1 accuracy: 65.60 -> 66.76 | Improved |
| add static attributes to transduction (ablation) | GTA-CLIP (ViT-B/16) · 5-dataset average | top-1 accuracy: 64.26 -> 65.60 | Improved |
| add static attributes to inductive CLIP only (ablation) | CLIP (ViT-B/16) · 5-dataset average | top-1 accuracy: 60.56 -> 61.59 | Improved |
| replace TransCLIP-ZS with GTA-CLIP | GTA-CLIP · 12-dataset average, zero-shot (ViT-B/16) | top-1 accuracy: 69.80 -> 73.81 | Improved |
| replace TransCLIP-ZS with GTA-CLIP on UCF101 with ViT-L/14 | GTA-CLIP · UCF101 (ViT-L/14) | top-1 accuracy: 81.73 -> 81.28 | Hurt |
| replace TransCLIP-FS with GTA-CLIP at 4 shots | GTA-CLIP · 12-dataset average, 4-shot (ViT-B/16) | top-1 accuracy: 75.22 -> 79.08 | Improved |
| replace TransCLIP-FS with GTA-CLIP at 1 shot | GTA-CLIP · 12-dataset average, 1-shot (ViT-B/16) | top-1 accuracy: 72.08 -> 75.49 | Improved |
| replace TransCLIP-ZS with GTA-CLIP using seen-class pretrained CLIP | GTA-CLIP · Cars unseen classes | top-1 accuracy: 78.04 -> 82.17 | Improved |
Closest papers by meaning
- Boosting Vision-Language Models with Transduction (NeurIPS 2024 · similarity 0.93)
- Transductive Zero-Shot and Few-Shot CLIP (CVPR 2024 · similarity 0.89)
- A Closer Look at the Few-Shot Adaptation of Large Vision-Language Models (CVPR 2024 · similarity 0.89)
- Building Vision-Language Models on Solid Foundations with Masked Distillation (CVPR 2024 · similarity 0.89)
- Conformal Prediction for Zero-Shot Models (CVPR 2025 · similarity 0.89)
- Test-Time Adaptation with CLIP Reward for Zero-Shot Generalization in Vision-Language Models (ICLR 2024 · 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.
