Research map / Architectures and efficiency / Pruning and efficient tuning
The Unreasonable Effectiveness of Random Pruning: Return of the Most Naive Baseline for Sparse Training
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 |
|---|---|---|---|
| Increase the base network size | Randomly pruned networks · CIFAR-10 and ImageNet | test accuracy | Helped |
| Use ERK layer-wise sparsity ratios rather than uniform ratios (ablation) | Randomly pruned ResNets · CIFAR-10 | test accuracy | Helped |
| Use SNIP layer-wise sparsity ratios rather than ERK+ (ablation) | Randomly pruned Wide ResNet-50 · ImageNet | test accuracy | Helped |
| Force the final fully connected layer to be dense while preserving overall parameter count (ablation) | ERK-pruned Wide ResNet-50 · ImageNet | test accuracy | Helped |
| Reduce final fully connected layer sparsity (ablation) | ERK-pruned Wide ResNet-50 · ImageNet | test accuracy | Helped |
| Use GraSP ratios instead of ERK or SNIP ratios (ablation) | Randomly pruned ResNets · CIFAR-10 | test accuracy | Hurt |
| Increase sparsity in large ResNets | Randomly pruned ResNet-56 and ResNet-110 · CIFAR-10 | adversarial robustness | Helped |
| Use random pruning instead of a dense network | ResNet-20 · CIFAR-10 | ECE | Helped |
| Use SNIP or ERK ratios rather than uniform ratios (ablation) | Randomly pruned networks · CIFAR-10 trained, CIFAR-100 tested | ROC-AUC | Helped |
| Use random pruning rather than a dense model | Large networks · CIFAR-10 trained, CIFAR-100 tested | ROC-AUC | Helped |
| Use SNIP ratios rather than ERK+ ratios (ablation) | Randomly pruned Wide ResNet-50 · ImageNet | test accuracy: +2% accuracy | Helped |
| Increase the final fully connected layer's parameter allocation (ablation) | ERK-pruned Wide ResNet-50 · ImageNet | test accuracy | Helped |
Closest papers by meaning
- Training Your Sparse Neural Network Better with Any Mask (ICML 2022 · similarity 0.94)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.92)
- SAFE: Finding Sparse and Flat Minima to Improve Pruning (ICML 2025 · similarity 0.91)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · similarity 0.91)
- No Free Prune: Information-Theoretic Barriers to Pruning at Initialization (ICML 2024 · similarity 0.91)
- Targeted Low-rank Refinement: Enhancing Sparse Language Models with Precision (ICML 2025 · similarity 0.90)
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.