Research map / Label-efficient and robust learning / Recognition and long-tail
The Power of Ensembles for Active Learning in Image Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use ensemble variation-ratio acquisition instead of MC Dropout acquisition | Active learning for CIFAR-10 · CIFAR-10, 14,500 labeled images | accuracy: 88.4% -> 91.5% | Helped |
| Use ensemble variation-ratio acquisition instead of Core-Set | Active learning for CIFAR-10 · CIFAR-10, 14,500 labeled images | accuracy: 88.2% -> 91.5% | Helped |
| Use ensemble variation-ratio acquisition instead of random acquisition | Active learning for diabetic retinopathy · Diabetic retinopathy, 21,000 labeled images | AUC: 0.965 -> 0.983 | Helped |
| Use ensemble variation-ratio acquisition instead of random acquisition | Active learning for ImageNet · ImageNet, 280k labeled images | top-1 accuracy: 0.470 -> 0.494 | Helped |
| Use ensemble variation-ratio acquisition instead of single-network entropy acquisition | Active learning for CIFAR-10 · CIFAR-10, DenseNet, reaching 85% accuracy | labeled images needed: 9959 (13300) -> 7053 (9613) | Helped |
| Increase the ensemble from three networks to five or seven (ablation) | Ensemble-based active learning · CIFAR-10 | accuracy | No clear effect |
| Fix the training-example order within mini-batches across ensemble members (ablation) | Ensemble-based active learning · CIFAR-10 | accuracy | No clear effect |
| Use the same weight initialization for all ensemble members (ablation) | Ensemble-based active learning · CIFAR-10 | accuracy | Hurt |
| Reduce ensemble network capacity to match the average active units in MC Dropout (ablation) | Ensemble-based active learning · CIFAR-10 | accuracy | Hurt |
| Increase MC Dropout network capacity to compensate for dropped units (ablation) | MC Dropout active learning · CIFAR-10 | accuracy | No clear effect |
| Train MC Dropout using images selected by the ensemble (ablation) | MC Dropout active learning · CIFAR-10 | accuracy | Helped |
| Use ensemble-selected rather than MC-selected images to train the ensemble (ablation) | Ensemble-based active learning · CIFAR-10 | accuracy | Helped |
Closest papers by meaning
- A Simple Yet Powerful Deep Active Learning With Snapshots Ensembles (ICLR 2023 · similarity 0.87)
- Deep Bayesian Active Learning with Image Data (ICML 2017 · similarity 0.86)
- Active Learning for Convolutional Neural Networks: A Core-Set Approach (ICLR 2018 · similarity 0.85)
- Active Bayesian Assessment of Black-Box Classifiers (AAAI 2021 · similarity 0.85)
- More or Less: When and How to Build Convolutional Neural Network Ensembles (ICLR 2021 · similarity 0.84)
- Improved Algorithm for Deep Active Learning under Imbalance via Optimal Separation (ICML 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.