Research map / Vision-language and multimodal / Cross-modal learning
Multimodality Helps Unimodality: Cross-Modal Few-Shot Learning With Multimodal Models
Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, Deva Ramanan
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 class-name text samples to image few-shot training | CLIP linear probing · 11-dataset average, 1-shot | top-1 accuracy: 36.7 -> 64.1 | Improved |
| Add class-name text samples and partially finetune the image encoder | CLIP few-shot adaptation · 11-dataset average, 1-shot | top-1 accuracy: 59.6 -> 64.7 | Improved |
| Add audio training samples to image classification | ImageNet-ESC-19 image classifier · ImageNet-ESC-19, 1-shot | accuracy: 68.0 -> 69.3 | Improved |
| Add audio training samples to image classification | ImageNet-ESC-27 image classifier · ImageNet-ESC-27, 4-shot | accuracy: 79.0 -> 78.9 | Hurt |
| Add image training samples to audio classification | ImageNet-ESC-19 audio classifier · ImageNet-ESC-19, 1-shot | accuracy: 31.2 -> 35.7 | Improved |
| Add image training samples to audio classification | ImageNet-ESC-27 audio classifier · ImageNet-ESC-27, 1-shot | accuracy: 28.2 -> 35.0 | Improved |
| Use cross-modal training rather than image-only training | WiSE-FT · 11-dataset average, 1-shot | top-1 accuracy: 59.1 -> 63.8 | Improved |
| Use cross-modal training rather than image-only training | CoOp prompting · 11-dataset average, 1-shot | top-1 accuracy: 59.6 -> 62.0 | Improved |
| Use cross-modal training rather than image-only training | Tip-Adapter · 11-dataset average, 1-shot | top-1 accuracy: 63.3 -> 64.4 | Improved |
| Use text-feature initialization instead of random initialization (ablation) | Cross-Modal Linear Probing · 11-dataset average, 1-shot | top-1 accuracy: 48.37_{(1.58)} -> 63.66_{(1.25)} | Improved |
| Use text-feature initialization instead of random initialization (ablation) | Cross-Modal Partial Finetuning · 11-dataset average, 1-shot | top-1 accuracy: 42.03_{(1.91)} -> 64.27_{(0.96)} | Improved |
| Finetune image attention pooling while freezing text encoder (ablation) | Cross-modal partial finetuning · 11-dataset average, 16-shot | top-1 accuracy: 75.54_{(0.53)} -> 76.65_{(0.44)} | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Tip-Adapter — Tip-Adapter: Training-Free Adaption of CLIP for Few-Shot Classification (ECCV 2022) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Adaptive Cross-Modal Few-shot Learning (NeurIPS 2019 · similarity 0.89)
- Everything is a Video: Unifying Modalities through Next-Frame Prediction (ICCV 2025 · similarity 0.89)
- Context-Aware Multimodal Pretraining (CVPR 2025 · similarity 0.89)
- MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal Data (NeurIPS 2026 · similarity 0.88)
- Learning Unseen Modality Interaction (NeurIPS 2023 · similarity 0.88)
- Connect, Collapse, Corrupt: Learning Cross-Modal Tasks with Uni-Modal Data (ICLR 2024 · similarity 0.87)
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.
