Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Rethinking Few Shot CLIP Benchmarks: A Critical Analysis in the Inductive Setting
Alexey Kravets, Da Chen, Vinay P. Namboodiri
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 |
|---|---|---|---|
| unlearn benchmark classes from CLIP (inductive evaluation) | CoOp · 7 datasets avg, inductive (unlearned CLIP) | accuracy: 71.4 -> 15.3 | Hurt |
| unlearn benchmark classes from CLIP (inductive evaluation) | PromptSRC · 7 datasets avg, inductive (unlearned CLIP) | accuracy: 74.7 -> 20.1 | Hurt |
| unlearn benchmark classes from CLIP (inductive evaluation) | CLIPLora · 7 datasets avg, inductive (unlearned CLIP) | accuracy: 74.9 -> 19.2 | Hurt |
| unlearn benchmark classes from CLIP (inductive evaluation) | SEPRES · 7 datasets avg, inductive (unlearned CLIP) | accuracy: 75.4 -> 56 | Hurt |
| add residual textual features to SEP (RES) (ablation) | SEP · 7 datasets avg, inductive (unlearned CLIP) | accuracy: 11.511 -> 56.428 | Improved |
| add SEP to residual features (ablation) | only RES · 7 datasets avg, inductive (unlearned CLIP) | accuracy: 21.492 -> 56.428 | Improved |
| add visual residual features (vRES) (ablation) | SEPRES · 7 datasets avg, inductive (unlearned CLIP) | accuracy: 56.428 -> 56.17 | No clear effect |
| set residual weight alpha to 0 (no residual features) (ablation) | SEPRES · inductive and partially transductive | accuracy | Hurt |
| increase residual weight alpha from 0.1 towards 1 (ablation) | SEPRES · partially transductive | accuracy | Hurt |
| use SSD unlearning instead of SalUn (ablation) | CLIP unlearning · Unlearning Dogs, Cars, Flowers | avg knowledge lost: 0.3617 -> 0.0283 | Improved |
| use SSD unlearning instead of TextProj (ablation) | CLIP unlearning · Unlearning Dogs, Cars, Flowers | avg knowledge lost: 0.0967 -> 0.0283 | Improved |
| unlearn subset vs train CLIP without the subset (ablation) | CLIP oracle · Birds subset accuracy | accuracy on subset: 0.043 vs 0.082 | No clear effect |
Closest papers by meaning
- Transductive Zero-Shot and Few-Shot CLIP (CVPR 2024 · similarity 0.90)
- AMU-Tuning: Effective Logit Bias for CLIP-based Few-shot Learning (CVPR 2024 · similarity 0.89)
- Demystifying CLIP Data (ICLR 2024 · similarity 0.89)
- Interpreting and Analysing CLIP's Zero-Shot Image Classification via Mutual Knowledge (NeurIPS 2024 · similarity 0.88)
- Does CLIP’s generalization performance mainly stem from high train-test similarity? (ICLR 2024 · similarity 0.88)
- Few-Shot Recognition via Stage-Wise Retrieval-Augmented Finetuning (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.
