Research map / Architectures and efficiency / RNNs and normalization
Instance-Dependent Early Stopping for Adaptive Data Pruning
Suqin Yuan, Runqi Lin, Felix Azian, Lei Feng, Bo Han, Gang Niu, Masashi Sugiyama, Tongliang Liu
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 |
|---|---|---|---|
| stop backprop on mastered instances (IES) | ResNet-18 training · CIFAR-10 | test accuracy: 92.9% -> 92.9% with 54.6% mini-batches saved | No clear effect |
| stop backprop on mastered instances (IES) | SGD(F) ResNet-18 · CIFAR-10 | test accuracy: 92.1% -> 92.4% | No clear effect |
| pretrain with IES before head fine-tuning | ResNet-101 · ImageNet-1k to CIFAR-10 | transfer accuracy: 80.3% -> 81.2% | Improved |
| pretrain with IES before head fine-tuning | ResNet-101 · ImageNet-1k to CIFAR-100 | transfer accuracy: 55.6% -> 57.5% | Improved |
| pretrain with IES before head fine-tuning | ResNet-101 · ImageNet-1k to Caltech-101 | transfer accuracy: 57.4% -> 59.9% | Improved |
| use IES instead of conventional early stopping at ~2x speedup | ResNet · CIFAR-10 | test accuracy: 90.4% -> 93.7% | Improved |
| use IES instead of conventional early stopping at ~3x speedup | ResNet · CIFAR-100 | test accuracy: 63.9% -> 73.0% | Improved |
| use IES instead of EfficientTrain at ~2x speedup | ResNet · CIFAR-100 | test accuracy: 75.0% -> 74.9% | No clear effect |
| use second-order loss difference criterion instead of loss value or other orders (ablation) | IES · CIFAR-10 and CIFAR-100 | test accuracy | Improved |
| select instances by IES rather than random removal (ablation) | backprop sample selection · CIFAR-10 and CIFAR-100 | test accuracy | Improved |
| vary delta over four orders of magnitude (ablation) | IES · CIFAR | test accuracy: within approximately 2% of baseline | No clear effect |
| apply IES under label noise (20% and 40%) | IES · CIFAR-10 symmetric 20% noise | best accuracy: 87.81% -> 87.81%, 0% mini-batch saved | No clear effect |
Closest papers by meaning
- Instance-dependent Early Stopping (ICLR 2025 · similarity 0.98)
- Instance-Conditional Timescales of Decay for Non-Stationary Learning (AAAI 2024 · similarity 0.84)
- A Statistical Theory of Regularization-Based Continual Learning (ICML 2024 · similarity 0.84)
- Repeated Random Sampling for Minimizing the Time-to-Accuracy of Learning (ICLR 2024 · similarity 0.84)
- ConsistentEE: A Consistent and Hardness-Guided Early Exiting Method for Accelerating Language Models Inference (AAAI 2024 · similarity 0.84)
- Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty (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.
