Research map / Architectures and efficiency / CNN design
PerforatedCNNs: Acceleration through Elimination of Redundant Convolutions
Mikhail Figurnov, Aizhan Ibraimova, Dmitry P Vetrov, Pushmeet Kohli
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 |
|---|---|---|---|
| fine-tune perforated network weights (ablation) | PerforatedCNN VGG-16 CPU 4.0x · ImageNet | top-5 error increase (%): +71.6 -> +5.5 | Improved |
| fine-tune perforated network weights (ablation) | PerforatedCNN AlexNet GPU 4.1x · ImageNet | top-5 error increase (%): +28.1 -> +6.2 | Improved |
| train from scratch/tune perforated network (ablation) | PerforatedCNN NIN GPU 3.5x · CIFAR-10 | top-1 error increase (%): +37.4 -> +12.4 | Improved |
| combine perforation with linear asymmetric decomposition | Perforation 4x · VGG-16 ImageNet | tuned error increase (%): +5.5 -> +1.6 | Improved |
| combine perforation with linear asymmetric decomposition | Asymmetric decomposition 4x · VGG-16 ImageNet | tuned error increase (%): +2.0 -> +1.6 | Improved |
| combine perforation with asymmetric decomposition | Perforation · VGG-16 ImageNet | memory reduction: 2.5x -> 1.4x | Improved |
| reduce conv2 kernel to 3x3 with impact mask | AlexNet conv2 impact mask · ImageNet | CPU speedup at about 1% top-5 error increase: 4.2x -> 9.1x | Improved |
| use impact mask vs grid mask | Perforation masks · AlexNet conv2 without fine-tuning | speedup-error tradeoff | Improved |
| replace missing values with zero instead of nearest neighbor (ablation) | Perforation interpolation · AlexNet conv3 | error | Hurt |
| use barycentric interpolation instead of nearest neighbor (ablation) | Perforation interpolation · AlexNet conv2/conv3 | error | Mixed |
| accelerate layers followed by max-pooling vs followed by convolution | Perforation · AlexNet conv2 vs conv3 | speedup-error tradeoff | Improved |
| use uniform mask at high speedups | Perforation uniform mask · AlexNet | error standard deviation | Hurt |
Closest papers by meaning
- Run, Don't Walk: Chasing Higher FLOPS for Faster Neural Networks (CVPR 2023 · similarity 0.87)
- Faster CNNs with Direct Sparse Convolutions and Guided Pruning (ICLR 2017 · similarity 0.86)
- ESPACE: Accelerating Convolutional Neural Networks via Eliminating Spatial and Channel Redundancy (AAAI 2017 · similarity 0.85)
- Growing Efficient Deep Networks by Structured Continuous Sparsification (ICLR 2021 · similarity 0.85)
- More Is Less: A More Complicated Network With Less Inference Complexity (CVPR 2017 · similarity 0.85)
- Pruning Filters for Efficient ConvNets (ICLR 2017 · similarity 0.85)
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.
