Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Fair Training with Zero Inputs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes zero uniformity training (ZUT), which feeds a category-agnostic all-zero image in parallel during training and adds a zero uniformity loss that pushes its classifier features toward uniformity, to improve classification fairness across image classification, person re-ID and semantic segmentation.
Problem
Classifiers show unfairness both from long-tailed training distributions (bias toward head classes) and from the clever Hans effect (reliance on non-discriminative features).
Approach
Category-agnostic zero-valued inputs reveal both long-tail and clever Hans biases in classifier outputs; a loss that unifies the classifier features for these zero inputs (std, mean, or entropy based) suppresses those biases during training, with no inference-time change.
Key results
- ZUT-ResNet improves VOC2007 mAP from 90.29 to 92.06 (+1.77%) and reaches 92.0 mAP versus 86.0 for Riformer.
- ZUT-VAN-b0 raises ADE20K mAcc from 49.41 to 52.56 and mIoU from 37.46 to 37.87 over VAN-b0.
- On PRCC, ZUT-ResNet reaches 62.2% R1, improving over CVSL (57.5%) by 4.7%; ZUT-Riformer reaches 48.9% accuracy on CIFAR100-LT versus CSA 46.6%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply ZUT to ResNet-50 (ablation) | ResNet-50 · VOC2007 | mAP: 90.29±0.15 -> 92.06±0.07 | Helped |
| Apply ZUT to Riformer | Riformer · CIFAR100-LT | accuracy: 47.4 -> 48.9 | Helped |
| Apply ZUT to ResNet-50 (ablation) | ResNet-50 · PRCC | R1: 59.93±0.81 -> 62.17±0.15 | Helped |
| Apply ZUT to VAN-b0 | VAN-b0 · ADE20K | mAcc: 49.41 -> 52.56 | Helped |
| Replace zero-valued input with one-valued input (ablation) | ZUT-ResNet-50 · PRCC | R1: 62.17±0.15 -> 60.77±0.51 | Hurt |
| Replace zero-valued input with normalization mean (ablation) | ZUT-ResNet-50 · PRCC | R1: 62.17±0.15 -> 59.00±1.08 | Hurt |
| Replace zero-valued input with random values (ablation) | ZUT-ResNet-50 · PRCC | R1: 62.17±0.15 -> 60.53±0.99 | Hurt |
| Replace standard-deviation ZUL with mean-based ZUL (ablation) | ZUT-ResNet-50 · PRCC | R1: 62.17±0.15 -> 61.73±1.15 | Hurt |
| Replace standard-deviation ZUL with entropy-based ZUL (ablation) | ZUT-ResNet-50 · PRCC | R1: 62.17±0.15 -> 61.30±1.73 | Hurt |
| Apply ZUT across backbones (ablation) | Classification models · VOC2007 and CIFAR100-LT | mAP and accuracy: 1.77%; 1.34% | Helped |
| Apply ZUT to ResNet-50 (ablation) | ResNet-50 · ADE20K | mAcc: 1.10% | Helped |
| Apply ZUT to VAN-b0 (ablation) | VAN-b0 · ADE20K | mAcc: 3.15% | Helped |
Surprising findings
- Using an all-zero input reveals both long-tail and clever Hans bias in the classifier outputs of models trained on VOC2007.
- Using the mean-value-of-normalization input decreased CC-ReID performance, attributed to interaction with random erasing augmentation.
Limitations
- Loss weight alpha is task-dependent and must be tuned per task.
- Improvements over some SOTA methods are only comparable rather than clearly superior (e.g., CC-ReID and ADE20K results).
Experimental setup
Datasets: VOC2007 CIFAR100-LT PRCC Celeb-ReID VCClothes CCVID ADE20K ImageNet
Benchmarks: VOC2007 CIFAR100-LT PRCC Celeb-ReID VCClothes CCVID ADE20K
Metrics: mAP accuracy rank-1 mAcc mIoU
Baselines: RNN-frequency Eva02 Riformer CR-CE CSA CVSL CLIP3DReID GI-ReID 3DSL AFDNet ACID 3DInvarReID
Builds on: CAL random erasing MMPretrain
Closest papers by meaning
- Classes Are Not Equal: An Empirical Study on Image Recognition Fairness (CVPR 2024 · similarity 0.86)
- Leveling Down in Computer Vision: Pareto Inefficiencies in Fair Deep Classifiers (CVPR 2022 · similarity 0.84)
- Navigating Towards Fairness with Data Selection (AAAI 2025 · similarity 0.84)
- Class Distribution Shifts in Zero-Shot Learning: Learning Robust Representations (NeurIPS 2024 · similarity 0.83)
- FairDD: Fair Dataset Distillation (NeurIPS 2025 · similarity 0.83)
- Efficient Fairness-Performance Pareto Front Computation (NeurIPS 2025 · similarity 0.83)
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.