Research map / Architectures and efficiency / RNNs and normalization
Selection via Proxy: Efficient Data Selection for Deep Learning
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 ResNet20 instead of ResNet164 for least-confidence selection | Active learning on CIFAR10 · CIFAR10, 50% labeling budget | data selection speed-up: 1.0× -> 7.2× | Helped |
| Use ResNet20 instead of ResNet164 for least-confidence selection | Active learning on CIFAR100 · CIFAR100, 50% labeling budget | top-1 error: 26.9±0.21 -> 26.6±0.14 | No clear effect |
| Use ResNet18 instead of ResNet50 for least-confidence selection | Active learning on ImageNet · ImageNet, 50% labeling budget | data selection speed-up: 1.0× -> 1.6× | Helped |
| Use fastText instead of VDCNN29 for least-confidence selection | Active learning on Amazon Review Polarity · Amazon Review Polarity, 40% labeling budget | data selection speed-up: 1.0× -> 41.9× | Helped |
| Use least-confidence active learning rather than random sampling | Active learning on Amazon Review Full · Amazon Review Full | top-1 error | No clear effect |
| Use ResNet20 as the forgetting-events proxy to select half the CIFAR10 training data | Core-set training of ResNet164 · CIFAR10, 50% data retained | end-to-end training speed-up: 1.6 × | Helped |
| Train the ResNet56 proxy for 50 epochs rather than 181 epochs (ablation) | Forgetting-events core-set selection on CIFAR10 · CIFAR10, 50% data retained | end-to-end training speed-up: 1.8 × | Helped |
| Train the ResNet56 proxy for 50 epochs rather than 181 epochs (ablation) | Forgetting-events core-set selection on CIFAR10 · CIFAR10, 50% data retained | top-1 error: 5.1% -> 5.4% | Hurt |
| Train the proxy for fewer epochs (ablation) | Active learning on CIFAR10 · CIFAR10, least-confidence selection, 50% budget | data selection speed-up: 7.2× -> 25.6× | Helped |
| Train the proxy for fewer epochs (ablation) | Active learning on CIFAR100 · CIFAR100, least-confidence selection, 50% budget | top-1 error: 26.6±0.14 -> 27.9±0.64 | Hurt |
| Train the proxy for fewer epochs (ablation) | Active learning on ImageNet · ImageNet, least-confidence selection, 50% budget | data selection speed-up: 1.6× -> 3.1× | Helped |
| Use ResNet20 rather than ResNet164 for greedy k-centers selection | Active learning on CIFAR10 · CIFAR10, 50% labeling budget | top-1 error: 5.6±0.04 -> 6.1±0.18 | Hurt |
Closest papers by meaning
- PEAKS: Selecting Key Training Examples Incrementally via Prediction Error Anchored by Kernel Similarity (ICML 2025 · similarity 0.85)
- Compute-Constrained Data Selection (ICLR 2025 · similarity 0.85)
- A CLIP-Powered Framework for Robust and Generalizable Data Selection (ICLR 2025 · similarity 0.84)
- DsDm: Model-Aware Dataset Selection with Datamodels (ICML 2024 · similarity 0.84)
- Data-Efficient Learning via Clustering-Based Sensitivity Sampling: Foundation Models and Beyond (ICML 2024 · similarity 0.84)
- NEAR: A Training-Free Pre-Estimator of Machine Learning Model Performance (ICLR 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.