Research map / Architectures and efficiency / Pruning and efficient tuning
Discrimination-aware Channel Pruning for Deep Neural Networks
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 |
|---|---|---|---|
| Use discrimination-aware channel selection with reconstruction loss | ResNet-50 · ILSVRC-12 | Top-1 error gap (%): +1.87 -> +1.06 | Helped |
| Use discrimination-aware channel selection with reconstruction loss | ResNet-50 · ILSVRC-12 | Top-5 error gap (%): +1.12 -> +0.61 | Helped |
| Use DCP instead of width-multiplier training | MobileNet v1 and MobileNet v2 · CIFAR-10 | Err. gap (%): MobileNet v1: +0.48 -> -0.41; MobileNet v2: +0.45 -> -0.22 | Helped |
| Use gradient-based channel selection rather than random channel selection (ablation) | DCP on VGGNet · CIFAR-10 | Err. gap (%): +0.14 -> -0.17 | Helped |
| Add intermediate discrimination-aware losses to width-multiplier training (ablation) | WM on VGGNet · CIFAR-10 | Err. gap (%): +0.38 -> +0.11 | Helped |
| Add intermediate discrimination-aware losses to width-multiplier training (ablation) | WM on MobileNet · CIFAR-10 | Err. gap (%): 0.26% | Helped |
| Increase the pruning rate from 30% to 70% (ablation) | DCP on ResNet-50 · ILSVRC-12 | Top-1 error: 23.60 -> 27.25 | Hurt |
| Add discrimination-aware loss alongside reconstruction loss (λ=0 to λ=1.0) (ablation) | DCP on ResNet-56 · CIFAR-10 | Testing error: 12.24 -> 7.84 | Helped |
| Use both reconstruction and discrimination-aware losses rather than discrimination loss alone (ablation) | DCP on ResNet-56 · CIFAR-10 | Testing error: 8.28 -> 7.84 | Helped |
| Tighten stopping tolerance from ε=0.1 to ε=0.001 (ablation) | DCP-Adapt on VGGNet · CIFAR-10 | Testing error (%): 12.68 -> 5.43 | Helped |
| Increase the number of intermediate discrimination-aware losses from 3 to 9 (ablation) | DCP on ResNet-56 · CIFAR-10 | Error gap (%): +0.31 -> +0.20 | Helped |
| Use more samples for channel selection (ablation) | DCP on ResNet-18 · ILSVRC-12 | Testing error | Helped |
Closest papers by meaning
- Rethinking the Smaller-Norm-Less-Informative Assumption in Channel Pruning of Convolution Layers (ICLR 2018 · similarity 0.92)
- Network Pruning via Performance Maximization (CVPR 2021 · similarity 0.91)
- Operation-Aware Soft Channel Pruning using Differentiable Masks (ICML 2020 · similarity 0.91)
- Revisiting Random Channel Pruning for Neural Network Compression (CVPR 2022 · similarity 0.91)
- REPrune: Channel Pruning via Kernel Representative Selection (AAAI 2024 · similarity 0.88)
- BilevelPruning: Unified Dynamic and Static Channel Pruning for Convolutional Neural Networks (CVPR 2024 · 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.