Research map / Image generation / Autoregressive and latent generation
Active Image Synthesis for Efficient Labeling
Jialei Chen, Yujia Xie, Kan Wang, Chuck Zhang, Mani A. Vannan, Ben Wang, Zhen Yu Qian
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 uniform random feature sampling with AISEL sampling (400 virtual images) (ablation) | GIN-based augmentation · Fashion (400 training images) | accuracy: 76.4% -> 81.9% | Improved |
| replace uncertainty-only top-k active learning selection with AISEL explore plus exploit sampling (ablation) | GIN-based augmentation · Fashion (400 training images) | accuracy: 78.2% -> 81.9% | Improved |
| add oracle-labeled AISEL virtual data to the native model | CNN trained on 400 images · Fashion (400 training images) | accuracy: 72.8% -> 81.9% | Improved |
| add oracle-labeled AISEL virtual data to the native model | CNN trained on 400 images · MNIST (400 training images) | accuracy: 88.2% -> 91.2% | Improved |
| use ACGAN augmentation labeled by the small training set | CNN trained on 400 images · Fashion (400 training images) | accuracy: 72.8% -> 72.3% | No clear effect |
| use ImageNet pretrained ResNet18 with last layer fine-tuned | native model · Fashion (400 training images) | accuracy: 72.8% -> 74.3% | No clear effect |
| increase random virtual images from 400 to 5000 (ablation) | GIN-based random sampling · Fashion (400 training images) | accuracy: 76.4% -> 80.7% | Improved |
| use AISEL (+400) instead of random sampling with 5000 virtual images (ablation) | GIN-based augmentation · Fashion (400 training images) | accuracy: 80.7% -> 81.9% | Improved |
| use AISEL virtual data in place of 400 extra actual labeled images | native model · Fashion (400 training images) | accuracy: 81.3% -> 81.9% | No clear effect |
| use AISEL (+1134) instead of random generation of the same size (ablation) | CNN on aortic stenosis CT · Aortic stenosis CT, 4-fold CV average | accuracy (%): 67.27 -> 75.60 | Improved |
| add AISEL (+1134) virtual data to the native model | CNN on aortic stenosis CT · Aortic stenosis CT, 4-fold CV average | accuracy (%): 59.48 -> 75.60 | Improved |
| increase random virtual data from 1134 to 10000 (ablation) | random generation · Aortic stenosis CT, 4-fold CV average | accuracy (%): 67.27 -> 78.57 | Improved |
Closest papers by meaning
- Auto-GAN: Self-Supervised Collaborative Learning for Medical Image Synthesis (AAAI 2020 · similarity 0.83)
- VALIANT: Prompt Instability for Active Learning in Black-Box Medical Imaging (AAAI 2026 · similarity 0.82)
- Deep Learning in Medical Image Registration: Magic or Mirage? (NeurIPS 2024 · similarity 0.82)
- Unlocking Pre-trained Image Backbones for Semantic Image Synthesis (CVPR 2024 · similarity 0.81)
- Improving Deep Learning for Accelerated MRI With Data Filtering (NeurIPS 2025 · similarity 0.81)
- SAFLEX: Self-Adaptive Augmentation via Feature Label Extrapolation (ICLR 2024 · similarity 0.81)
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.
