Research map / Detection and segmentation / CNN features
Reinforce Data, Multiply Impact: Improved Model Accuracy and Robustness with Dataset Reinforcement
Fartash Faghri, Hadi Pouransari, Sachin Mehta, Mehrdad Farajtabar, Ali Farhadi, Mohammad Rastegari, Oncel Tuzel
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 |
|---|---|---|---|
| reinforce ImageNet with ensemble-teacher outputs and augmentations (ImageNet+) instead of plain ImageNet | ResNet-50 · ImageNet-1k val (1000 epochs) | top-1 accuracy (%): 79.6 -> 81.7 | Improved |
| use ImageNet+ instead of ImageNet for pretraining | SwinTransformer-Tiny · ImageNet-R | top-1 accuracy (%): 37.7 -> 58.0 | Improved |
| use ImageNet+ pretraining for detection and segmentation | ResNet-50 · MS-COCO / ADE20K | mAP / mIoU: 42.2 -> 42.5 mAP; 42.8 -> 44.2 mIoU | Improved |
| reinforce both pretraining and fine-tuning dataset | MobileNetV3-Large on Food-101 · Food-101 | top-1 accuracy (%): 86.1 -> 89.5 | Improved |
| use harder reinforcement (RRC+M*+R*) for light-weight CNN (ablation) | MobileNetv2-0.25 · ImageNet val (150 epochs) | top-1 accuracy (%): 54.3 -> 52.5 | Hurt |
| use harder reinforcement (RRC+M*+R*) for transformer (ablation) | ViT-Small · ImageNet val (150 epochs) | top-1 accuracy (%): RRC 77.0 -> 80.8 | Improved |
| use ensemble of IG-ResNext teachers instead of single most accurate model (ablation) | dataset reinforcement teacher · KD students on ImageNet | student accuracy | Improved |
| increase teacher accuracy within an architecture family (ablation) | KD teacher choice · KD students on ImageNet | student accuracy | Mixed |
| swap IG-ResNext ensemble teacher for CLIP-pretrained ensemble (ablation) | ImageNet+ · ImageNet val (1000 epochs), ResNet-50 | top-1 accuracy (%): 81.7 -> 81.1 | Hurt |
| reduce stored samples to 100 per image with top-5 sparsity (ImageNet+-Small) (ablation) | ImageNet+ · ImageNet val, ResNet-50 (300 ep) | top-1 accuracy (%): 80.6 -> 80.6 | No clear effect |
| use ImageNet+ rather than FKD for dataset reinforcement | ResNet-50 · ImageNet val (300 epochs) | top-1 accuracy (%): 80.1 -> 80.6 | Improved |
Closest papers by meaning
- Does progress on ImageNet transfer to real-world datasets? (NeurIPS 2023 · similarity 0.87)
- When Dynamic Data Selection Meets Data Augmentation: Achieving Enhanced Training Acceleration (ICML 2025 · similarity 0.86)
- SELECT: A Large-Scale Benchmark of Data Curation Strategies for Image Classification (NeurIPS 2024 · similarity 0.85)
- Data Filtering Networks (ICLR 2024 · similarity 0.85)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.85)
- Unlocking Dataset Distillation with Diffusion Models (NeurIPS 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.
