Research map / Label-efficient and robust learning / Recognition and long-tail
Unbiased Supervised Contrastive Learning
Carlo Alberto Barbano, Benoit Dufumier, Enzo Tartaglione, Marco Grangetto, Pietro Gori
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 SupCon with ε-SupInfoNCE | ResNet-50 · CIFAR-10 | top-1 accuracy: 95.64\std{0.02} -> 96.14\std{0.01} | Improved |
| replace SupCon with ε-SupInfoNCE | ResNet-50 · CIFAR-100 | top-1 accuracy: 75.41\std{0.19} -> 76.04\std{0.01} | Improved |
| replace ε-SupCon with ε-SupInfoNCE | ResNet-50 · ImageNet-100 | top-1 accuracy: 82.83\std{0.11} -> 83.3\std{0.06} | Improved |
| add FairKL to ε-SupInfoNCE | ε-SupInfoNCE · Biased-MNIST | top-1 accuracy: 33.16\std{3.57} -> 90.51\std{1.55} at ρ=0.999; 73.86\std{0.81} -> 96.19\std{0.23} at ρ=0.997; 83.65\std{0.36} -> 97.00\s | Improved |
| add FairKL to ε-SupInfoNCE | ε-SupInfoNCE · Corrupted CIFAR-10 and bFFHQ | top-1 accuracy: 33.33\std{0.38}, 36.53\std{0.38}, 41.45\std{0.42}, 50.73\std{0.90}; bFFHQ 64.8\std{0.43} | Mixed |
| add FairKL to ε-SupInfoNCE | ε-SupInfoNCE · 9-Class ImageNet UNB and ImageNet-A | top-1 accuracy: UNB 94.1\std{0.3} -> 94.8\std{0.3}; ImageNet-A 34.1\std{0.6} -> 35.7\std{0.5} | Improved |
| add a projection head (ablation) | ε-SupInfoNCE + FairKL · Corrupted CIFAR-10 | top-1 accuracy: 33.33\std{0.38} -> 30.85\std{0.19}; 36.53\std{0.38} -> 32.75\std{0.57}; 41.45\std{0.42} -> 37.95\std{0.14}; 50.73\std{0. | Hurt |
| use full FairKL, including distance-distribution variance constraints, rather than mean-only FairKL (ablation) | ε-SupInfoNCE + FairKL · Corrupted CIFAR-10 and bFFHQ | top-1 accuracy: 32.37\std{1.72} -> 33.33\std{0.38}; 35.65\std{0.75} -> 36.53\std{0.38}; 39.94\std{0.50} -> 41.45\std{0.42}; 50.25\std{0. | Improved |
| use ε-SupInfoNCE instead of ε-SupCon | supervised contrastive learning · Biased-MNIST ρ=0.999 | top-1 accuracy: 24.36\std{3.23} -> 33.16\std{3.57} | Improved |
| use ε-SupInfoNCE instead of ε-SupCon | supervised contrastive learning · Biased-MNIST ρ≤0.997 | top-1 accuracy: ρ=0.997: 74.35\std{0.09} vs 73.86\std{0.81}; ρ=0.995: 84.13\std{1.31} vs 83.65\std{0.36}; ρ=0.99: 91.12\std{0.35} vs 91. | Mixed |
| add a small cross-entropy contribution during end-to-end training (ablation) | FairKL debiasing training · biased datasets | accuracy | No clear effect |
| use a larger batch size and weight decay in the bFFHQ setup (ablation) | ε-SupInfoNCE debiasing · bFFHQ | accuracy: 54.8\% without FairKL; 56.87\% for the vanilla model | No clear effect |
Closest papers by meaning
- Unbiased Classification through Bias-Contrastive and Bias-Balanced Learning (NeurIPS 2021 · similarity 0.88)
- Fighting Fire with Fire: Contrastive Debiasing without Bias-free Data via Generative Bias-transformation (ICML 2023 · similarity 0.88)
- Fair Contrastive Learning for Facial Attribute Classification (CVPR 2022 · similarity 0.88)
- A Conservative Approach for Unbiased Learning on Unknown Biases (CVPR 2022 · similarity 0.87)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.85)
- A Tale of Two Classes: Adapting Supervised Contrastive Learning to Binary Imbalanced Datasets (CVPR 2025 · similarity 0.85)
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.
