Research map / Vision-language and multimodal / CLIP and zero-shot
MED-LCDS: Multi-Expert-Domain CLIP Classification via Logit Calibration
Zheng Zeng, Deepak Sridhar, Nuno Vasconcelos
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 domain supervision and logit scaling to MoLE | MoLE · Cross-domain, base split | mean accuracy (%): 71.80 -> 85.51 | Improved |
| add domain supervision to softmax MoE gating (ablation) | MoLE · Cross-domain, all split | mean accuracy (%): 65.85 -> 71.61 | Improved |
| add domain-wise logit scaling on top of domain supervision (ablation) | MED-DS · Cross-domain, all split | mean accuracy (%): 71.61 -> 71.80 | Improved |
| add domain-wise logit scaling on top of domain supervision (ablation) | MED-DS · Cross-domain, base split | mean accuracy (%): 85.43 -> 85.51 | No clear effect |
| use softmax instead of sigmoid gating (ablation) | MED-DSLC · Cross-domain, all split | mean accuracy (%): 69.97 -> 71.80 | Improved |
| add logit scaling under extreme target-domain imbalance (k=2) (ablation) | MED-DS · target-domain imbalance sweep | target-domain accuracy: +2.2 points | Improved |
| add logit scaling under target-domain imbalance (k=4) (ablation) | MED-DS · target-domain imbalance sweep | target-domain accuracy: +1.3 points | Improved |
| one-shot training of gating and temperatures only (ablation) | MED-DSLC · Cross-domain, all split | mean accuracy (%): 71.80 -> 71.71 | No clear effect |
| use SigLIP backbone instead of CLIP (ablation) | MED-DSLC · Cross-domain, all split | mean accuracy (%): 71.80 -> 62.27 | Hurt |
| increase number of classes per domain from k=5 to 40 (First-k) (ablation) | MED-DSLC · All split First-k | mean accuracy (%): 91.37 -> 79.39 | Hurt |
| add MED-DSLC over MoLE with more classes (First-k, k=40) (ablation) | MoLE · All split First-k | mean accuracy (%): 66.99 -> 79.39 | Improved |
| use Single-LoRA with 10x parameters instead of expert LoRAs | Single-LoRA · Cross-domain, base split | mean accuracy (%): 66.22 -> 42.35 | Hurt |
Closest papers by meaning
- CLIPCEIL: Domain Generalization through CLIP via Channel rEfinement and Image-text aLignment (NeurIPS 2024 · similarity 0.87)
- Leveraging Vision-Language Models for Improving Domain Generalization in Image Classification (CVPR 2024 · similarity 0.87)
- Interpreting and Analysing CLIP's Zero-Shot Image Classification via Mutual Knowledge (NeurIPS 2024 · similarity 0.86)
- Bridge the Modality and Capability Gaps in Vision-Language Model Selection (NeurIPS 2024 · similarity 0.86)
- UMFC: Unsupervised Multi-Domain Feature Calibration for Vision-Language Models (NeurIPS 2024 · similarity 0.85)
- Unknown Prompt the only Lacuna: Unveiling CLIP's Potential for Open Domain Generalization (CVPR 2024 · similarity 0.85)
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.
