Research map / Architectures and efficiency / Pruning and efficient tuning
DMCP: Differentiable Markov Channel Pruning for Neural Networks
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 |
|---|---|---|---|
| Use differentiable Markov channel pruning | MobileNetV2 and ResNet on ImageNet · ImageNet classification at matched FLOPs | Top-1 accuracy | Helped |
| Use Expected Sampling instead of Direct Sampling (ablation) | DMCP · MobileNetV2-210M and ResNet50-1.1G | Top-1 accuracy: 72.2 vs 72.4; 74.1 vs 74.0 | Mixed |
| Use MobileNetV2 1.5x rather than 1.0x as the unpruned network (ablation) | DMCP · ImageNet, MobileNetV2 59M and 210M | Top-1 accuracy: 62.6 -> 62.7; 71.8 -> 72.2 | Helped |
| Add the warmup phase (ablation) | DMCP · ImageNet, MobileNetV2-210M | Top-1 accuracy: 71.4 -> 72.2 | Helped |
| Replace the variant sandwich rule with the original sandwich rule (ablation) | DMCP · ImageNet, MobileNetV2-210M | Top-1 accuracy: 72.2 -> 71.5 | Hurt |
| Add task loss to architecture-parameter updates (ablation) | DMCP · ImageNet, MobileNetV2-59M | Top-1 accuracy: 52.1 -> 61.5 | Helped |
| Update network weights along with architecture parameters (ablation) | DMCP · ImageNet, MobileNetV2-59M | Top-1 accuracy: 61.5 -> 62.7 | Helped |
| Replace Markov modeling with independent Bernoulli architecture variables (ablation) | DMCP · ImageNet, MobileNetV2-210M | Top-1 accuracy: 2.3% lower than DMCP | Hurt |
| Replace warmup with a pre-trained model (ablation) | DMCP · ImageNet, MobileNetV2-210M | Top-1 accuracy: 0.6% accuracy drop | Hurt |
| Use slimmable training for the pruned model (ablation) | DMCP · ImageNet, MobileNetV2 300M and 211M | Top-1 accuracy: 73.5 -> 74.6; 72.2 -> 73.5 | Helped |
Closest papers by meaning
- Revisiting Random Channel Pruning for Neural Network Compression (CVPR 2022 · similarity 0.90)
- Rethinking the Smaller-Norm-Less-Informative Assumption in Channel Pruning of Convolution Layers (ICLR 2018 · similarity 0.89)
- Pruning from Scratch (AAAI 2020 · similarity 0.89)
- Network Pruning via Performance Maximization (CVPR 2021 · similarity 0.89)
- BilevelPruning: Unified Dynamic and Static Channel Pruning for Convolutional Neural Networks (CVPR 2024 · similarity 0.88)
- MDP: Multidimensional Vision Model Pruning with Latency Constraint (CVPR 2025 · similarity 0.87)
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.