AI Paper Atlas

Research map / Vision-language and multimodal / CLIP and zero-shot

UP-DP: Unsupervised Prompt Learning for Data Pre-Selection with Vision-Language Models

Xin Li, Sima Behpour, Thang Doan, Wenbin He, Liang Gou, Liu Ren

NeurIPS 2023 · Vision-language and multimodal › CLIP and zero-shot · approach: Prompt learning for CLIP · 2 citations (OpenAlex)

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 changedWhereMeasuredEffect
use UP-DP cluster-head selection instead of USL-M k-means selection on same learned-prompt featuresUP-DP · 7-dataset average, CLIP ViTG14accuracy: 0.300 -> 0.371Improved
use multimodal features with learned prompt (USL-M) instead of image features (USL-I)USL · 7-dataset average, CLIP ViTG14accuracy: 0.290 -> 0.300Improved
shorten learnable context from 16 to 4 tokens (ablation)UP-DP · 7-dataset average, CLIP ViTG14accuracy: 0.322 -> 0.354Improved
shorten learnable context from 16 to 4 tokens (ablation)UP-DP · 7-dataset average, CLIP RN50accuracy: 0.210 -> 0.240Improved
apply prompt learned on Flowers102 to all datasets instead of image-only featuresBLIP-2 KNN features · 7-dataset average KNNaccuracy: 0.760 -> 0.864Improved
apply prompt learned on EuroSAT instead of image-only featuresBLIP-2 KNN features · OxfordPets KNNaccuracy: 0.681 -> 0.501Hurt
use empty prompt instead of image-only featuresBLIP-2 KNN features · 7-dataset average KNNaccuracy: 0.760 -> 0.606Hurt
use initial (untrained) prompt instead of image-only featuresBLIP-2 KNN features · 7-dataset average KNNaccuracy: 0.760 -> 0.592Hurt
use cluster head g_C features instead of backbone f features (ablation)UP-DP · 7-dataset average, CLIP ViTG14accuracy: 0.354 -> 0.371Improved
use instance head g_I features instead of backbone f features (ablation)UP-DP · 7-dataset average, CLIP ViTG14accuracy: 0.354 -> 0.357No clear effect
increase annotation budget by 50 percent (300 labels)UP-DP vs Random · 7-dataset average, CLIP ViTG14accuracy: 0.400 -> 0.423Mixed
UP-DP versus random selection on Caltech101UP-DP · Caltech101, CLIP ViTG14accuracy: 0.334 -> 0.339No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy