Research map / Vision-language and multimodal / Cross-modal learning
Missing Modality Prediction for Unpaired Multimodal Learning via Joint Embedding of Unimodal Models
Taesup Kim, Donggeun Kim
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 |
|---|---|---|---|
| predict missing-modality embedding with read-only prompts instead of CLS token (ablation) | unimodal-encoder framework · MM-IMDb (train 100% image/30% text) | F1-Macro: 44.91 -> 48.99 | Improved |
| use prompts for feature prediction (cosine similarity of image feature prediction) (ablation) | unimodal-encoder framework · MM-IMDb | cosine similarity of image feature prediction: 0.4 -> 0.71 | Improved |
| add variance and covariance terms (VICReg) to invariance loss (ablation) | feature predictor · MM-IMDb (complete training, text-missing) | F1-Macro: 25.72 -> 40.88 | Improved |
| use stop-gradient in invariance loss (ablation) | feature predictor · MM-IMDb (complete training, text-missing) | F1-Macro: 25.72 -> 22.72 | Hurt |
| add read-only attention masking to prompts (ablation) | prompt-based predictor · MM-IMDb (complete training, text-missing) | F1-Macro: 34.72 -> 40.88 | Improved |
| increase prompt length (peak at 6) | read-only prompts · MM-IMDb (complete training, text-missing) | F1-Macro | Mixed |
| use BitFit instead of LN tuning, prefix tuning or adapters (ablation) | unimodal encoders · MM-IMDb (complete training, text-missing) | F1-Macro: 39.07 (LN) / 37.75 (prefix) -> 40.88 | Improved |
| raise VICReg coefficients from 5/5/1 to 50/50/1 (ablation) | feature prediction loss · MM-IMDb (complete training) | F1-Macro: 36.11 -> 40.88 | Improved |
| use larger unimodal encoders | framework · complete training | robustness to missing modalities | Improved |
| ours vs unimodal baseline when train/test missing pattern is aligned | framework · Food-101 (train 30% image/100% text, test 30%/100%) | accuracy: 87.44 -> 87.11 | No clear effect |
Closest papers by meaning
- Do More Modalities Always Help? A Geometric Perspective on Missing-Modality Robustness (NeurIPS 2026 · similarity 0.91)
- Mitigating Pervasive Modality Absence Through Multimodal Generalization and Refinement (AAAI 2025 · similarity 0.91)
- Deep Correlated Prompting for Visual Recognition with Missing Modalities (NeurIPS 2024 · similarity 0.90)
- Retrieval-Augmented Dynamic Prompt Tuning for Incomplete Multimodal Learning (AAAI 2025 · similarity 0.89)
- Beyond Modality Fusion: Deep Ensembles for Multimodal Classification (NeurIPS 2026 · similarity 0.89)
- MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal Data (NeurIPS 2026 · 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.
