Research map / Label-efficient and robust learning / Recognition and long-tail
Multi-Label Cluster Discrimination for Visual Representation Learning
Xiang An, Kaicheng Yang, Xiangzi Dai, Ziyong Feng, Jiankang Deng
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 |
|---|---|---|---|
| replace instance discrimination (OpenCLIP) with multi-label cluster discrimination | ViT-L/14 LAION-400M · ImageNet finetune | accuracy: 86.2 -> 87.1 | Improved |
| assign top-8 positive centers per image instead of a single label (ablation) | MLCD ViT-B/32 · ImageNet linear probe | accuracy: 71.4 -> 75.2 | Improved |
| increase positive centers from 8 to 32 (ablation) | MLCD ViT-B/32 · ImageNet linear probe | accuracy: 75.2 -> 68.7 | Hurt |
| use adaptive similarity-threshold labels instead of fixed top-l (ablation) | MLCD ViT-B/32 · ImageNet linear probe | accuracy: 75.2 -> 73.3 | Hurt |
| add separate positive and negative loss terms (MLCD vs MLC) (ablation) | MLC loss · ImageNet linear probe | accuracy: 76.9 -> 78.1 | Improved |
| sample 10% of negative class centers instead of all (ablation) | MLCD ViT-B/32 · ImageNet linear probe | accuracy: 63.2 -> 75.2 | Improved |
| sample only 1% of negative centers (ablation) | MLCD ViT-B/32 · ImageNet linear probe | accuracy: 75.2 -> 73.4 | Hurt |
| increase number of classes from 1M to 5M (ablation) | MLCD ViT-B/32 · ImageNet linear probe | accuracy: 75.2 -> 74.7 | Hurt |
| increase number of classes from 100K to 1M (ablation) | MLCD ViT-B/32 · ImageNet linear probe | accuracy: 66.9 -> 75.2 | Improved |
| replace single-label UNICOM with MLCD on LAION-400M | ViT-B/32 · 26-dataset linear probe average | average accuracy: 79.2 -> 81.4 | Improved |
| replace CLIP vision tower with MLCD in LLaVA-style VLM | Qwen2-7B VLM · TextVQA | accuracy: 48.98 -> 42.52 | Hurt |
| replace CLIP vision tower with MLCD in LLaVA-style VLM | Qwen2-72B VLM · MathVista mini | accuracy: 38.3 -> 46.5 | Improved |
Closest papers by meaning
- Region-based Cluster Discrimination for Visual Representation Learning (ICCV 2025 · similarity 0.88)
- Multi-Label Supervised Contrastive Learning (AAAI 2024 · similarity 0.87)
- DICNet: Deep Instance-Level Contrastive Network for Double Incomplete Multi-View Multi-Label Classification (AAAI 2023 · similarity 0.87)
- Towards Effective Visual Representations for Partial-Label Learning (CVPR 2023 · similarity 0.87)
- Correlative and Discriminative Label Grouping for Multi-Label Visual Prompt Tuning (CVPR 2025 · similarity 0.87)
- Image Clustering via the Principle of Rate Reduction in the Age of Pretrained Models (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.
