Research map / Label-efficient and robust learning / Recognition and long-tail
A Decade's Battle on Dataset Bias: Are We There Yet?
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.
Revisits the "Name That Dataset" dataset classification task with modern neural networks on large web datasets, finding that networks classify source dataset with high accuracy (84.7% on YCD 3-way), and that such features partly transfer.
Problem
Whether modern neural networks trained on large, diverse, presumably less biased internet image datasets can still identify which dataset an image comes from, revealing dataset bias, and whether the learned features generalize.
Approach
Revisits the dataset classification ("Name That Dataset") experiment at web scale with modern architectures, showing high held-out accuracy, robustness to corruptions, dependence on data quantity and augmentation in a generalization-like way (contrasted with a pseudo-dataset memorization control), and transfer of learned features to ImageNet linear probing.
Key results
- ConvNeXt-T reaches 84.7% accuracy on 3-way YFCC/CC/DataComp classification (chance 33.3%), and 69.2% on 6-way.
- Human participants (n=20) average 45.4% on the YCD task, far below the network.
- MAE linear probing on YCD reaches 78.4% when pre-trained on YCD itself and 76.2% when pre-trained on ImageNet-1K.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add random cropping to no augmentation (ablation) | ConvNeXt-T dataset classifier · YCD, 10K training images per dataset | accuracy: 43.2 -> 66.1 | Helped |
| Add RandAug to random cropping (ablation) | ConvNeXt-T dataset classifier · YCD, 10K training images per dataset | accuracy: 66.1 -> 70.2 | Helped |
| Add MixUp and CutMix to random cropping and RandAug (ablation) | ConvNeXt-T dataset classifier · YCD, 10K training images per dataset | accuracy: 70.2 -> 72.4 | Helped |
| Use the full augmentation pipeline rather than no augmentation (ablation) | ConvNeXt-T dataset classifier · YCD, 100K training images per dataset | accuracy: 71.9 -> 80.1 | Helped |
| Add MixUp and CutMix to random cropping and RandAug (ablation) | ConvNeXt-T dataset classifier · YCD, 1M training images per dataset | accuracy: 85.0 -> 84.7 | Hurt |
| Increase the number of training images (ablation) | Dataset classifier · YCD | validation accuracy | Helped |
| Increase the number of datasets from three to six (ablation) | ConvNeXt-T dataset classifier · Dataset classification | accuracy: 84.7 -> 69.2 | Hurt |
| Use MAE pre-trained on YCD instead of ImageNet (ablation) | MAE linear probe · YCD dataset classification | accuracy: 76.2 -> 78.4 | Helped |
| Add WIT to Y+C+D pre-training data (ablation) | Dataset-classification pre-training and ImageNet-1K linear probing · ImageNet-1K | accuracy: 27.7 -> 34.2 | Helped |
| Add LAION to Y+C+D+W pre-training data (ablation) | Dataset-classification pre-training and ImageNet-1K linear probing · ImageNet-1K | accuracy: 34.2 -> 34.2 | No clear effect |
| Apply stronger image corruption to the training and validation data (ablation) | ConvNeXt-T dataset classifier · YCD dataset classification | accuracy: 84.7 -> 75.1 | Hurt |
| Train on the combined datasets rather than a single dataset (ablation) | MoCo v3 · Cross-dataset generalization, average validation loss | validation loss: 2.251 -> 1.871 | Helped |
Surprising findings
- Even a ConvNeXt with about 7K parameters reaches 72.4% accuracy on YCD classification.
- Self-supervised features with a linear probe on frozen features still classify datasets at 76-78% accuracy.
Limitations
- Concrete forms of the captured bias remain unclear.
- Only six image datasets are studied; video and language domains are not covered.
Experimental setup
Datasets: YFCC100M CC12M DataComp-1B WIT LAION ImageNet-1K ImageNetV2
Benchmarks: YCD dataset classification ImageNet-1K linear probing
Metrics: accuracy validation loss
Baselines: AlexNet VGG-16 ResNet-50 ViT-S MAE MoCo v3
Builds on: Name That Dataset (Torralba & Efros 2011) ConvNeXt MAE MoCo v3
Closest papers by meaning
- Understanding Bias in Large-Scale Visual Datasets (NeurIPS 2024 · similarity 0.88)
- Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers (CVPR 2025 · similarity 0.86)
- SELECT: A Large-Scale Benchmark of Data Curation Strategies for Image Classification (NeurIPS 2024 · similarity 0.86)
- Common Sense Bias Modeling for Classification Tasks (AAAI 2025 · similarity 0.85)
- Classes Are Not Equal: An Empirical Study on Image Recognition Fairness (CVPR 2024 · similarity 0.85)
- Low-Biased General Annotated Dataset Generation (CVPR 2025 · similarity 0.84)
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.