Research map / Label-efficient and robust learning / Recognition and long-tail
Easy2Hard: From Partially to Fully Unmatched Modalities as Negative Samples in Contrastive Learning
Zhicheng Yang, Yichen Liu, Chang Ge, Xiaopeng Jiang
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 |
|---|---|---|---|
| use Easy2Hard partial/full negative split with curriculum instead of Symile uniform negatives | Symile (TC/MIP) · EH-MIMIC | Acc@1: 0.477 -> 0.503 | Improved |
| use Easy2Hard instead of Symile | Symile (TC/MIP) · Channel | Acc@1: 0.567 -> 0.573 | No clear effect |
| use Easy2Hard instead of Symile | Symile (TC/MIP) · Symile-MIMIC | Acc@1: 0.434 -> 0.462 | Improved |
| use Easy2Hard instead of Symile | Symile (TC/MIP) · MM-IMDb | Acc@1: 0.404 -> 0.421 | Improved |
| use Easy2Hard instead of Symile with 5 modalities | Symile (TC/MIP) · HoloAssist (5-modal) | Acc@1: 0.833 -> 0.882 | Improved |
| use Easy2Hard instead of pivot-based ImageBind-like alignment | ImageBind-like baseline · HoloAssist (5-modal) | Acc@1: 0.439 -> 0.882 | Improved |
| use sigmoid curriculum instead of linear schedule (ablation) | Easy2Hard · Channel | Acc@1: 0.494 -> 0.573 | Improved |
| use sigmoid curriculum instead of linear schedule (ablation) | Easy2Hard · EH-MIMIC | Acc@1: 0.441 -> 0.503 | Improved |
| increase sigmoid slope k from 0.5 to 0.7 (tm=15) (ablation) | Easy2Hard · EH-MIMIC | Acc@1: 0.488 -> 0.446 | Hurt |
| vary midpoint tm from 11 to 19 (k=0.3) (ablation) | Easy2Hard · Channel | Acc@1: 0.562 -> 0.562 (peak 0.576 at tm=15) | No clear effect |
Closest papers by meaning
- Connect, Collapse, Corrupt: Learning Cross-Modal Tasks with Uni-Modal Data (ICLR 2024 · similarity 0.88)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.88)
- On the Value of Cross-Modal Misalignment in Multimodal Representation Learning (NeurIPS 2025 · similarity 0.88)
- On the Similarities of Embeddings in Contrastive Learning (ICML 2025 · similarity 0.88)
- Contrasting with Symile: Simple Model-Agnostic Representation Learning for Unlimited Modalities (NeurIPS 2024 · similarity 0.87)
- A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics (NeurIPS 2025 · 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.
