Research map / Architectures and efficiency / RNNs and normalization
DECORE: Deep Compression With Reinforcement Learning
Check 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 misclassification penalty from the low-penalty DECORE setting to DECORE-500 (ablation) | VGG16 on CIFAR-10 · CIFAR-10 | Top-1 accuracy (%): 93.96 -> 94.02 | Helped |
| Use DECORE-200 rather than L1 pruning | VGG16 · CIFAR-10 | Parameter pruning rate: 64.0% -> 89.0% | Helped |
| Use DECORE-200 rather than HRank-1 | VGG16 · CIFAR-10 | Top-1 accuracy (%): 93.43 -> 93.56 | Helped |
| Lower the penalty to DECORE-50 for more aggressive pruning (ablation) | VGG16 · CIFAR-10 | Parameter pruning rate: DECORE-100: 96.6%; DECORE-50: 98.3% | Helped |
| Increase the misclassification penalty to DECORE-500 (ablation) | GoogleLeNet · CIFAR-10 | Top-1 accuracy (%): 95.05 -> 95.20 | Helped |
| Reduce the penalty from DECORE-175 to DECORE-70 for more pruning (ablation) | DenseNet-40 · CIFAR-10 | Parameter pruning rate: 20.7% -> 65.0% | Helped |
| Use DECORE-200 rather than ResNet-56 baseline | ResNet-56 · CIFAR-10 | Parameter pruning rate: 0.0% -> 49.0% | Helped |
| Reduce the penalty from DECORE-300 to DECORE-175 for more aggressive pruning (ablation) | ResNet-110 · CIFAR-10 | Parameter pruning rate: 64.8% -> 79.6% | Helped |
| Use DECORE-8 rather than the uncompressed ResNet-50 | ResNet-50 · ImageNet | Top-1 accuracy (%): 76.15 -> 76.31 | Helped |
| Lower the penalty from DECORE-8 to DECORE-6 for stronger compression (ablation) | ResNet-50 · ImageNet | Parameter count: 22.69M -> 14.10M | Helped |
| Remove the single highest-weight channel (ablation) | VGG16 · CIFAR-10 | Top-1 accuracy (%): 93.96% -> 92.54% | Hurt |
| Remove the lowest-weight channel (ablation) | VGG16 · CIFAR-10 | Top-1 accuracy (%) | No clear effect |
Closest papers by meaning
- N2N learning: Network to Network Compression via Policy Gradient Reinforcement Learning (ICLR 2018 · similarity 0.88)
- Model Preserving Compression for Neural Networks (NeurIPS 2022 · similarity 0.86)
- Auto-Compressing Networks (NeurIPS 2025 · similarity 0.85)
- BAME: Block-Aware Mask Evolution for Efficient N:M Sparse Training (ICML 2025 · similarity 0.84)
- Subspace Networks: Scaling Decentralized Training with Communication-Efficient Model Parallelism (NeurIPS 2025 · similarity 0.84)
- DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence (NeurIPS 2026 · 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.