Research map / Architectures and efficiency / CNN design
Batch Pruning by Activation Stability
Md Mustakin Alam, Shaker Islam, Aminul Islam
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 |
|---|---|---|---|
| add Batch Normalization to ResNet-18 under B-PAS (ablation) | B-PAS · CIFAR-10 (ResNet-18) | accuracy: 89.87 -> 95.60 | Improved |
| add batch-level instead of sample-level pruning (ablation) | B-PAS · ImageNet-1K (ResNet-50) | top-1 accuracy: 70.87 -> 78.43 | Improved |
| train 200 epochs instead of 90 epochs (ablation) | B-PAS · ImageNet-1K (ResNet-50) | DSI (%): 12 -> 47 | Improved |
| learning rate 0.1 instead of 0.01 (ablation) | B-PAS · ImageNet-1K (ResNet-50) | top-1 accuracy: 74.29 -> 78.43 | Improved |
| use mean std instead of entropy to quantify stability (ablation) | B-PAS · CIFAR-10 (ResNet-18) | accuracy: 89.64 -> 95.60 | Improved |
| use mean std instead of max std to quantify stability (ablation) | B-PAS · CIFAR-10 (ResNet-18) | DSI (%): 13.45 -> 25.06 | Improved |
| use Batch Norm instead of Layer Norm (ablation) | B-PAS · CIFAR-10 (ResNet-18) | accuracy: 93.99 -> 95.60 | Improved |
| increase batch size from 32 to 128 (ablation) | B-PAS · CIFAR-10 (ResNet-18) | DSI (%): 20.25 -> 25.06 | Improved |
| swap B-PAS for random pruning of the same number of batches (ablation) | ResNet-18 · CIFAR-100 (ResNet-18) | accuracy: 78.20 -> 75.36 | Hurt |
| add B-PAS to ResNet-50 full-data training | ResNet-50 · ImageNet-1K (ResNet-50) | top-1 accuracy: 78.07 -> 78.43 | Improved |
| use more aggressive delta schedule (5x10^-5 to 5x10^-4) in B-PAS (ablation) | B-PAS · ImageNet-1K (ResNet-50) | top-1 accuracy: 78.07 -> 76.10 | Hurt |
| add B-PAS to CvT-13 training | CvT-13 · ImageNet-1K (CvT) | top-1 accuracy: 79.65 -> 79.6 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- +InfoBatch (30%) — InfoBatch: Lossless Training Speed Up by Unbiased Dynamic Data Pruning (ICLR 2024) · better on 8 of 10 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Accelerating CNN Training by Pruning Activation Gradients (ECCV 2020 · similarity 0.87)
- Improving the Robustness of Deep Neural Networks via Stability Training (CVPR 2016 · similarity 0.87)
- Variance-Based Pruning for Accelerating and Compressing Trained Networks (ICCV 2025 · similarity 0.87)
- Inducing and Exploiting Activation Sparsity for Fast Inference on Deep Neural Networks (ICML 2020 · similarity 0.86)
- Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty (ICML 2025 · similarity 0.85)
- Visual Prompting Upgrades Neural Network Sparsification: A Data-Model Perspective (AAAI 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.
