Research map / Architectures and efficiency / CNN design
PENNI: Pruned Kernel Sharing for Efficient CNN Inference
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 |
|---|---|---|---|
| Add coefficient pruning and model shrinking after decomposition and retraining (ablation) | PENNI on CIFAR10 VGG16 · CIFAR10, VGG16 | FLOPs reduction: 41.65% -> 93.26% | Helped |
| Add coefficient pruning and model shrinking after decomposition and retraining (ablation) | PENNI on CIFAR10 ResNet18 · CIFAR10, ResNet18 | Pruned accuracy: 93.89% -> 94.01% | Helped |
| Add coefficient pruning and model shrinking after decomposition and retraining (ablation) | PENNI on CIFAR10 ResNet56 · CIFAR10, ResNet56 | Pruned accuracy: 94.00% -> 93.38% | Hurt |
| Add coefficient pruning and model shrinking after decomposition and retraining (ablation) | PENNI on ImageNet AlexNet · ImageNet, AlexNet | FLOPs reduction: 25.88% -> 70.04% | Helped |
| Add coefficient pruning and model shrinking after decomposition and retraining (ablation) | PENNI on ImageNet AlexNet · ImageNet, AlexNet | Top-1 accuracy: 55.41% -> 55.57% | Helped |
| Add coefficient pruning and model shrinking after decomposition and retraining (ablation) | PENNI on ImageNet ResNet50, R1 · ImageNet, ResNet50 | FLOPs reduction: 21.10% -> 94.73% | Helped |
| Use a larger subspace dimension, d=6, rather than d=5 (ablation) | PENNI on ImageNet ResNet50 · ImageNet, ResNet50 | Top-1 accuracy | Mixed |
| Compress VGG16 with PENNI rather than the baseline | VGG16 · VGG16-CIFAR10, CPU | Latency: 12.9 -> 5.96 | Helped |
| Compress VGG16 with PENNI rather than the baseline | VGG16 · VGG16-CIFAR10, CPU | Memory: 137 -> 77.6 | Helped |
| Compress VGG16 with PENNI rather than the baseline | VGG16 · VGG16-CIFAR10, GPU | Latency: 10.8 -> 7.26 | Helped |
| Compress VGG16 with PENNI rather than the baseline | VGG16 · VGG16-CIFAR10, GPU | Memory: 487 -> 424 | Helped |
| Increase the number of basis kernels beyond d=4 (ablation) | PENNI on VGG16 · CIFAR10, VGG16 | Test accuracy | No clear effect |
Closest papers by meaning
- Faster CNNs with Direct Sparse Convolutions and Guided Pruning (ICLR 2017 · similarity 0.89)
- Towards Compact CNNs via Collaborative Compression (CVPR 2021 · similarity 0.89)
- Exploiting Kernel Sparsity and Entropy for Interpretable CNN Compression (CVPR 2019 · similarity 0.88)
- KernelDNA: Cross-Layer Kernel Sharing via Decoupled Neural Adapters (NeurIPS 2026 · similarity 0.88)
- SPDY: Accurate Pruning with Speedup Guarantees (ICML 2022 · similarity 0.87)
- REPrune: Channel Pruning via Kernel Representative Selection (AAAI 2024 · similarity 0.86)
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.