Research map / Architectures and efficiency / Pruning and efficient tuning
Towards Optimal Structured CNN Pruning via Generative Adversarial 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 |
|---|---|---|---|
| jointly prune channels and blocks rather than blocks alone (ablation) | GAL on ResNet-50 · ImageNet ILSVRC 2012 | Top-5 error (%): 9.06 -> 9.18; FLOPs 2.33B -> 1.84B | Mixed |
| add adversarial regularization to the discriminator (ablation) | GAL · ResNet-56 and DenseNet-40 on CIFAR-10 | generator output similarity to baseline | Helped |
| add FISTA instead of SGD with hard thresholding (ablation) | GAL · ResNet-56 on CIFAR-10 | error after fine-tuning (%): 9.65 -> 8.42 | Helped |
| add FISTA instead of SGD with hard thresholding (ablation) | GAL · DenseNet-40 on CIFAR-10 | error after fine-tuning (%): 7.64 -> 5.50 | Helped |
| remove the GAN objective (ablation) | GAL with FISTA on ResNet-56 and DenseNet-40 · CIFAR-10 | classification error (%) | Hurt |
| replace the GAN with label-dependent CGAN (ablation) | GAL with FISTA on ResNet-56 · CIFAR-10 | error (%): 8.42 -> 9.56 | Hurt |
| fine-tune the pruned LeNet (ablation) | GAL-0.1 · MNIST | error (%): 1.03 -> 1.01 | Hurt |
| fine-tune the pruned LeNet (ablation) | GAL · MNIST | error (%): GAL-0.01: 0.95 -> 0.86; GAL-0.05: 1.05 -> 0.90; GAL-0.1: 1.03 -> 1.01 | Mixed |
| increase sparsity from GAL-0.6 to GAL-0.8 (ablation) | GAL on ResNet-56 · CIFAR-10 | error after fine-tuning (%): 6.62 -> 8.42 | Hurt |
| increase sparsity from GAL-0.5 to GAL-1 (ablation) | GAL on ResNet-50 · ImageNet ILSVRC 2012 | Top-5 error (%): 9.06 -> 10.25 | Hurt |
| increase sparsity from GAL-0.5-joint to GAL-1-joint (ablation) | GAL on ResNet-50 · ImageNet ILSVRC 2012 | Top-5 error (%): 9.18 -> 10.88 | Hurt |
Closest papers by meaning
- One Less Reason for Filter Pruning: Gaining Free Adversarial Robustness with Structured Grouped Kernel Pruning (NeurIPS 2023 · similarity 0.92)
- DPFPS: Dynamic and Progressive Filter Pruning for Compressing Convolutional Neural Networks from Scratch (AAAI 2021 · similarity 0.91)
- Provable Filter Pruning for Efficient Neural Networks (ICLR 2020 · similarity 0.91)
- Dynamic Structure Pruning for Compressing CNNs (AAAI 2023 · similarity 0.91)
- Learning Filter Pruning Criteria for Deep Convolutional Neural Networks Acceleration (CVPR 2020 · similarity 0.91)
- Convolutional Neural Network Pruning With Structural Redundancy Reduction (CVPR 2021 · similarity 0.91)
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.