Research map / Vision-language and multimodal / CLIP and zero-shot
Learning Robust Visual-Semantic Embeddings
Yao-Hung Hubert Tsai, Liang-Kang Huang, Ruslan Salakhutdinov
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 unsupervised reconstruction and MMD objectives on labeled training data (ablation) | ReViSE (supervised only) · Inductive zero-shot recognition average | top-1 accuracy (%): 3.4% recognition improvement | Improved |
| add unsupervised reconstruction and MMD objectives on labeled training data (ablation) | ReViSE (supervised only) · Inductive zero-shot retrieval average | mAP (%): 3.9% retrieval improvement | Improved |
| add unlabeled training data to unsupervised loss (ablation) | ReViSE^b · AwA+CUB inductive zero-shot | top-1 accuracy (%): 2.6% recognition improvement and 0.4% retrieval improvement | Improved |
| add unsupervised-data adaptation inference (ablation) | ReViSE^c · Inductive zero-shot recognition average | top-1 accuracy (%): 5.0% average recognition improvement | Improved |
| add unsupervised-data adaptation inference (ablation) | ReViSE^c · Inductive zero-shot retrieval average | mAP (%): 5.8% average retrieval improvement | Improved |
| add unsupervised-data adaptation inference (ablation) | ReViSE^c · Transductive zero-shot recognition average | top-1 accuracy (%): 53.6 -> 68.1 | Improved |
| add MMD loss between visual and textual codes (ablation) | ReViSE^dagger · Transductive zero-shot recognition average | top-1 accuracy (%): 66.8 -> 68.1 | Improved |
| add contractive loss (ablation) | ReViSE^dagger-dagger · Transductive zero-shot recognition average | top-1 accuracy (%): 65.8 -> 66.8 | Improved |
| apply unsupervised-data adaptation inference to DeViSE | DeViSE · Transductive zero-shot recognition average | top-1 accuracy (%): 49.3 -> 60.6 | Improved |
| apply unsupervised-data adaptation inference to CMT | CMT · Transductive zero-shot recognition average | top-1 accuracy (%): 50.5 -> 60.5 | Improved |
| apply unlabeled-data adaptation to CMT in inductive setting on AwA test images | CMT · AwA test images inductive zero-shot recognition | top-1 accuracy (%): 3.1% recognition deterioration | Hurt |
| apply unlabeled-data adaptation to DeViSE in inductive setting on AwA test images | DeViSE · AwA test images inductive zero-shot retrieval | mAP (%): 2.7% retrieval deterioration | Hurt |
Closest papers by meaning
- Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation (CVPR 2025 · similarity 0.89)
- Towards Universal Image Embeddings: A Large-Scale Dataset and Challenge for Generic Image Representations (ICCV 2023 · similarity 0.88)
- Image-Text Embedding Learning via Visual and Textual Semantic Reasoning (TPAMI 2022 · similarity 0.88)
- Beyond Supervised vs. Unsupervised: Representative Benchmarking and Analysis of Image Representation Learning (CVPR 2022 · similarity 0.88)
- VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks (ICLR 2025 · similarity 0.88)
- Learning Shared Representations from Unpaired Data (NeurIPS 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.
