Research map / Vision-language and multimodal / CLIP and zero-shot
Explaining CLIP Zero-shot Predictions Through Concepts
Onat Ozdemir, Anders Christensen, Stephan Alaniz, Zeynep Akata, Emre Akbas
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 |
|---|---|---|---|
| project CLIP into a learned concept space (EZPC) instead of plain CLIP | CLIP RN50 · ImageNet-100 GZS | harmonic mean H: 0.693 -> 0.682 | No clear effect |
| project CLIP into a learned concept space instead of plain CLIP | CLIP RN50 · ImageNet-1k GZS | harmonic mean H: 0.530 -> 0.481 | Hurt |
| use EZPC instead of Z-CBM and SpLiCE | concept-based zero-shot CLIP · CUB GZS | harmonic mean H: 0.189 -> 0.465 vs Z-CBM | Improved |
| use EZPC instead of Z-CBM | concept-based zero-shot CLIP · Places365 GZS | harmonic mean H: 0.357 -> 0.352 | No clear effect |
| increase reconstruction weight lambda from 0.01 to 10 (ablation) | EZPC · ImageNet-100 GZS | harmonic mean H: 0.358 -> 0.695 | Improved |
| increase lambda beyond 10 (to 100 or 1000) (ablation) | EZPC · ImageNet-100 GZS | harmonic mean H: 0.695 -> 0.694 | No clear effect |
| increase lambda to 100 for explanations (ablation) | EZPC · qualitative concept relevance | explanation quality | Hurt |
| use a larger backbone (ViT-L/14, SigLIP) for EZPC (ablation) | EZPC · ImageNet-100 GZS | harmonic mean H: 0.682 (RN50) -> 0.878 (SigLIP) | Improved |
| train EZPC on ImageNet-100 and test on CIFAR-100 | EZPC vs CLIP · CIFAR-100 cross-dataset GZS | harmonic mean H: 0.380 -> 0.409 | Improved |
| train EZPC on ImageNet-100 and test on CUB | EZPC vs CLIP · CUB cross-dataset GZS | harmonic mean H: 0.526 -> 0.515 | No clear effect |
| replace per-image optimization (SpLiCE, Z-CBM) with single linear projection | EZPC · ImageNet-100 inference | latency (ms/img): 338.51 / 542.34 -> 5.90 | Improved |
Closest papers by meaning
- Interpreting and Analysing CLIP's Zero-Shot Image Classification via Mutual Knowledge (NeurIPS 2024 · similarity 0.94)
- PerceptionCLIP: Visual Classification by Inferring and Conditioning on Contexts (ICLR 2024 · similarity 0.91)
- Transductive Zero-Shot and Few-Shot CLIP (CVPR 2024 · similarity 0.88)
- Cascade-CLIP: Cascaded Vision-Language Embeddings Alignment for Zero-Shot Semantic Segmentation (ICML 2024 · similarity 0.88)
- TPR: Topology-Preserving Reservoirs for Generalized Zero-Shot Learning (NeurIPS 2024 · similarity 0.88)
- Exploring Regional Clues in CLIP for Zero-Shot Semantic Segmentation (CVPR 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.
