Research map / Vision-language and multimodal / CLIP and zero-shot
Bayesian Prompt Learning for Image-Language Model Generalization
Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor G. Turrisi da Costa, Cees G.M. Snoek, Georgios Tzimiropoulos, Brais Martinez
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 |
|---|---|---|---|
| model prompts variationally (conditional Bayesian prompt learning) instead of CoCoOp | CLIP ViT-B/16 prompt learning · Task I: unseen classes (11 datasets, avg) | accuracy (%): 71.69 -> 74.94 | Improved |
| use prompt-space distribution instead of ProDA | CLIP ViT-B/16 prompt learning · Task I: unseen classes (11 datasets, avg) | accuracy (%): 72.30 -> 74.94 | Improved |
| use Bayesian prompt learning instead of CoCoOp for cross-dataset transfer | CLIP ViT-B/16 prompt learning · Task II: cross-dataset (ImageNet to 10 datasets, avg) | accuracy (%): 65.59 -> 65.95 | Improved |
| use Bayesian prompt learning instead of CoCoOp for cross-domain robustness | CLIP ViT-B/16 prompt learning · Task III: cross-domain (ImageNet variants, avg) | accuracy (%): 59.88 -> 60.44 | Improved |
| use Bayesian prompt learning instead of CoOp for in-domain accuracy | CLIP ViT-B/16 prompt learning · In-domain performance, Task I average | accuracy (%): 82.66 -> 80.10 | Hurt |
| add variational formulation (unconditional) over CoOp baseline (ablation) | CoOp · Task I: DTD unseen | accuracy (%): 41.18 -> 58.70 | Improved |
| add image conditioning to the latent distribution (ablation) | Bayesian prompt learning · Task I: unseen classes (5 datasets) | accuracy (%): Flowers102 75.90 -> 70.40; DTD 58.70 -> 60.80 | Mixed |
| use learned posterior N(mu(x), Sigma(x)) instead of N(0,I) sample (ablation) | Bayesian prompt learning · Task I: DTD unseen (one seed) | accuracy (%): 26.60 -> 56.40 | Improved |
| increase Monte Carlo samples from 1 to 20 at inference (ablation) | Bayesian prompt learning · Task I: EuroSAT unseen | accuracy (%): 64.50 -> 74.80 | Improved |
| increase Monte Carlo samples beyond 20 (ablation) | Bayesian prompt learning · Task I: unseen classes (5 datasets) | accuracy (%): EuroSAT 74.80 -> 76.20 at 80 samples; FGVC 35.00 -> 34.70 | No clear effect |
| initialize context tokens with 'An image of a {class}' instead of random (ablation) | Bayesian prompt learning · Task I: EuroSAT unseen (one seed) | accuracy (%): 75.10 -> 77.90 | Improved |
| use variational sampling instead of naive prompt ensemble on CoCoOp (ablation) | CoCoOp · Task I: unseen classes (5 datasets) | accuracy (%): EuroSAT 66.10 -> 75.30; DTD 52.17 -> 60.80 | Improved |
Closest papers by meaning
- Make Prompts Adaptable: Bayesian Modeling for Vision-Language Prompt Learning with Data-Dependent Prior (AAAI 2024 · similarity 0.93)
- Understanding prompt engineering may not require rethinking generalization (ICLR 2024 · similarity 0.91)
- Any-Shift Prompting for Generalization over Distributions (CVPR 2024 · similarity 0.90)
- CoPL: Contextual Prompt Learning for Vision-Language Understanding (AAAI 2024 · similarity 0.90)
- Learning to Prompt with Text Only Supervision for Vision-Language Models (AAAI 2025 · similarity 0.90)
- IPO: Interpretable Prompt Optimization for Vision-Language Models (NeurIPS 2024 · similarity 0.89)
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.
