Research map / Architectures and efficiency / Pruning and efficient tuning
Lookahead: A Far-sighted Alternative of Magnitude-based Pruning
Sejun Park*, Jaeho Lee*, Sangwoo Mo, Jinwoo Shin
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 |
|---|---|---|---|
| Replace magnitude-based scores with adjacent-layer lookahead distortion scores | FCN · MNIST | test error rate at 0.12% weights remaining: 67.62 -> 16.45 | Improved |
| Replace magnitude-based scores with adjacent-layer lookahead distortion scores | Conv-6 · CIFAR-10 | test error rate at 3.62% weights remaining: 28.60 -> 24.52 | Improved |
| Replace magnitude-based scores with adjacent-layer lookahead distortion scores | VGG-19 · CIFAR-10 | test error rate at 1.24% weights remaining: 77.90{ $\pm$0.12} -> 22.82{ $\pm$0.81} | Improved |
| Replace magnitude-based scores with adjacent-layer lookahead distortion scores | ResNet-18 · CIFAR-10 | test error rate at 0.36% weights remaining: 26.77{ $\pm$1.04} -> 22.45{ $\pm$0.64} | Improved |
| Replace magnitude-based scores with adjacent-layer lookahead distortion scores | ResNet-50 · Tiny-ImageNet | top-5 test error rate at 0.72% weights remaining: 59.18 $\pm16.31$ -> 46.16 $\pm1.04$ | Improved |
| Use only succeeding-layer lookahead (ablation) | LAP · MNIST | test error rate at 0.12% weights remaining: 16.45 -> 50.11 | Hurt |
| Use only preceding-layer lookahead (ablation) | LAP · MNIST | test error rate at 0.12% weights remaining: 16.45 -> 28.51 | Hurt |
| Prune layers in forward order using already-pruned preceding layers (ablation) | LAP · MNIST | test error rate at 0.12% weights remaining: 16.45 -> 12.00 | Improved |
| Prune layers in backward order using already-pruned succeeding layers (ablation) | LAP · MNIST | test error rate at 0.12% weights remaining: 16.45 -> 13.00 | Improved |
| Apply forward-order pruning gradually over five steps (ablation) | LAP · MNIST | test error rate at 0.12% weights remaining: 16.45 -> 10.98 | Improved |
| Apply backward-order pruning gradually over five steps (ablation) | LAP · MNIST | test error rate at 0.12% weights remaining: 16.45 -> 11.92 | Improved |
| Use estimated activation probabilities to rescale lookahead scores (ablation) | LAP · MNIST | test error rate at 0.12% weights remaining: 11.01 -> 7.99 | Improved |
Closest papers by meaning
- “Learning-Compression” Algorithms for Neural Net Pruning (CVPR 2018 · similarity 0.90)
- Pruning Neural Networks at Initialization: Why Are We Missing the Mark? (ICLR 2021 · similarity 0.90)
- An Operator Theoretic View On Pruning Deep Neural Networks (ICLR 2022 · similarity 0.90)
- Robust Pruning at Initialization (ICLR 2021 · similarity 0.89)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · similarity 0.88)
- Targeted Low-rank Refinement: Enhancing Sparse Language Models with Precision (ICML 2025 · similarity 0.88)
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.
