Research map / Label-efficient and robust learning / Noisy labels
Weakly Supervised Image Classification Through Noise Regularization
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train clean and residual nets jointly so the residual net regularizes the clean classifier (ablation) | Proposed method · OpenImage | mAP: 67.34 -> 69.02 | Helped |
| Train clean and residual nets jointly so the residual net regularizes the clean classifier (ablation) | Proposed method · Clothing1M | top-1 accuracy: 75.13 -> 79.93 | Helped |
| Train only the noisy-label classifier h, without clean-label training of g (ablation) | Proposed method · OpenImage | mAP: 61.82 -> 64.06 | Hurt |
| Train only the noisy-label classifier h, without clean-label training of g (ablation) | Proposed method · Clothing1M | top-1 accuracy: 68.94 -> 67.92 | Hurt |
| Use tanh rather than sigmoid activation (ablation) | Proposed method · MS COCO2014, 5%, 10%, and 20% clean labels | mAP: 58.28 -> 57.95; 59.51 -> 59.19; 60.55 -> 60.11 | No clear effect |
| Use ReLU rather than sigmoid activation (ablation) | Proposed method · MS COCO2014, 5%, 10%, and 20% clean labels | mAP: 58.28 -> 58.16; 59.51 -> 59.29; 60.55 -> 60.19 | No clear effect |
| Use the residual-net method instead of the strongest listed fine-tuning baseline | Proposed method · MS COCO2014, 5% clean labels | mAP: 55.10 -> 58.28 | Helped |
| Use the residual-net method instead of the strongest listed fine-tuning baseline | Proposed method · MS COCO2014, 10% clean labels | mAP: 56.77 -> 59.51 | Helped |
| Use the residual-net method instead of the strongest listed fine-tuning baseline | Proposed method · MS COCO2014, 20% clean labels | mAP: 59.03 -> 60.55 | Helped |
| Use the residual-net method rather than the strongest listed baseline | Proposed method · OpenImage | mAP: 65.66 -> 69.02 | Helped |
| Use the proposed method rather than Forward fine-tuning | Proposed method · Clothing1M | top-1 accuracy: 80.38 -> 79.93 | No clear effect |
Closest papers by meaning
- Large Loss Matters in Weakly Supervised Multi-Label Classification (CVPR 2022 · similarity 0.89)
- CleanNet: Transfer Learning for Scalable Image Classifier Training With Label Noise (CVPR 2018 · similarity 0.89)
- Losses over Labels: Weakly Supervised Learning via Direct Loss Construction (AAAI 2023 · similarity 0.89)
- Multi-Evidence Filtering and Fusion for Multi-Label Classification, Object Detection and Semantic Segmentation Based on Weakly Supervised Learning (CVPR 2018 · similarity 0.88)
- Toward Robustness against Label Noise in Training Deep Discriminative Neural Networks (NeurIPS 2017 · similarity 0.87)
- Learning From Semi-Supervised Weak-Label Data (AAAI 2018 · 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.