Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Enhancing CLIP with CLIP: Exploring Pseudolabeling for Limited-Label Prompt Tuning
Cristina Menghini, Andrew Delworth, Stephen Bach
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 |
|---|---|---|---|
| GRIP textual prompts vs CLIP zero-shot (SSL, avg over datasets) | CLIP ViT-B/32 · Flowers102/RESICS45/FGVCAircraft/MNIST/EuroSAT/DTD | accuracy: 20.7 points | Improved |
| GRIP textual prompts vs CoOp | CLIP ViT-B/32 · SSL and TRZSL | accuracy: 9.6 points in SSL, and 26.6 in TRZSL | Improved |
| GRIP visual prompts vs VPT | CLIP ViT-B/32 · SSL and TRZSL | accuracy: 12.9 points in SSL, and 20.8 in TRZSL | Improved |
| GRIP multimodal prompts vs UPT | CLIP ViT-B/32 · SSL and TRZSL | accuracy: 4.7 in semi-supervised learning, and 19.5 in transductive zero-shot | Improved |
| use iterative IFPL instead of static FPL (ablation) | FPL · Flowers102/RESICS45/DTD | accuracy: 5.6 points in SSL, 1.7 in UL, and 5.6 in TRZSL | Improved |
| use GRIP (growing pseudolabels) instead of FPL (ablation) | FPL · Flowers102/RESICS45/DTD | accuracy: 7.8 points in SSL, 3.1 in UL, and 9.7 in TRZSL | Improved |
| use GRIP instead of FPL on MNIST/EuroSAT/FGVCAircraft (ablation) | FPL · MNIST/EuroSAT/FGVCAircraft | accuracy: 16.7 points | Improved |
| use IFPL instead of FPL on MNIST/EuroSAT/FGVCAircraft (ablation) | FPL · MNIST/EuroSAT/FGVCAircraft | accuracy: 4.4 points | Hurt |
| GRIP visual prompts vs CLIP zero-shot on FGVCAircraft UL | CLIP ViT-B/32 · FGVCAircraft UL | accuracy: 17.58 -> 15.22 | Hurt |
| GRIP vs CoOp on EuroSAT SSL | CoOp · EuroSAT SSL | accuracy: 59.51 -> 58.66 | Hurt |
| use larger image encoder ViT-L/14 instead of ViT-B/32 (textual, SSL) | GRIP · Flowers102/RESICS45/DTD | avg improvement over CLIP: 17.46 -> 15.85 | Mixed |
| use prompt tuning (GRIP) instead of linear probing on rich classes (ablation) | CLIP · UL, six datasets | rich-class accuracy change vs CLIP: 9.39 vs 0.33 average drop | Improved |
Closest papers by meaning
- Enhancing Zero-Shot Vision Models by Label-Free Prompt Distribution Learning and Bias Correcting (NeurIPS 2024 · similarity 0.90)
- ViLT-CLIP: Video and Language Tuning CLIP with Multimodal Prompt Learning and Scenario-Guided Optimization (AAAI 2024 · similarity 0.90)
- Candidate Pseudolabel Learning: Enhancing Vision-Language Models by Prompt Tuning with Unlabeled Data (ICML 2024 · similarity 0.89)
- Learning to Prompt with Text Only Supervision for Vision-Language Models (AAAI 2025 · similarity 0.89)
- PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 2024 · similarity 0.88)
- Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification (CVPR 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.
