Research map / Label-efficient and robust learning / Recognition and long-tail
LPT: Long-tailed Prompt Tuning for Image Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 group-specific prompts and phase 2 to the shared-prompt phase (ablation) | LPT · Places-LT | overall accuracy: 49.41% -> 50.07% | Helped |
| Train phases 1 and 2 separately instead of jointly (ablation) | LPT · Places-LT | overall accuracy: 47.48% -> 50.07% | Helped |
| Use the phase-1 class token rather than the pretrained ViT-B token as query (ablation) | LPT phase 2 · Places-LT | overall accuracy: 49.81% -> 50.07% | Helped |
| Replace GCL with asymmetric GCL loss (ablation) | LPT phase 2 · Places-LT | overall accuracy: 49.58% -> 50.07% | Helped |
| Use dual sampling with γ=0.5 rather than balanced sampling alone (ablation) | LPT phase 2 · Places-LT | overall accuracy: 49.90% -> 50.07% | Helped |
| Set dual-sampling γ to 1.0 rather than 0.5 (ablation) | LPT phase 2 · Places-LT | overall accuracy: 50.07% -> 49.62% | Hurt |
| Use 20 group-specific prompts rather than 5, 10, or 40 (ablation) | LPT phase 2 · Places-LT | overall accuracy: 49.81% (5) -> 50.07% (20); 49.87% (40) | Mixed |
| Use K=6 shared-only blocks rather than K=8 or K=4 (ablation) | LPT phase 2 · Places-LT | overall accuracy: 49.77% (K=8), 50.07% (K=6), 49.92% (K=4) | Mixed |
| Ensemble two matched group prompts rather than one (ablation) | LPT phase 2 · Places-LT | few-shot accuracy: 46.32% -> 46.88% | Helped |
| Increase ensemble size beyond two matched prompts (ablation) | LPT phase 2 · Places-LT | overall accuracy: 50.07% (k=2), 50.00% (k=3), 49.93% (k=4) | No clear effect |
| Replace shared-prompt-only phase 1 with prompt tuning over a fixed feature extractor (ablation) | Places-LT · overall accuracy | overall accuracy: 41.33% -> 49.10% | Helped |
| Use asymmetric GCL instead of cross-entropy in the prompt-tuned phase-1 variant (ablation) | Places-LT type (b) · Places-LT | few-shot accuracy: 43.25% -> 47.32% | Helped |
Closest papers by meaning
- Improving Visual Prompt Tuning by Gaussian Neighborhood Minimization for Long-Tailed Visual Recognition (NeurIPS 2024 · similarity 0.90)
- Decoupled Optimisation for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.87)
- Facing the Elephant in the Room: Visual Prompt Tuning or Full finetuning? (ICLR 2024 · similarity 0.86)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.86)
- Quick-Tune: Quickly Learning Which Pretrained Model to Finetune and How (ICLR 2024 · similarity 0.86)
- AutoVP: An Automated Visual Prompting Framework and Benchmark (ICLR 2024 · 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.