Research map / Architectures and efficiency / CNN design
Training CNNs with Selective Allocation of Channels
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 |
|---|---|---|---|
| Train convolutional layers with channel de-allocation and re-allocation (selective convolution) | CNNs, including DenseNet, ResNet, and ResNeXt · CIFAR-10, CIFAR-100, Fashion-MNIST, Tiny-ImageNet, and ImageNet | classification error rate: CIFAR-10 DenseNet-40: 6.62 -> 6.09; ImageNet ResNet-50: 23.9 -> 23.4 | Helped |
| Use selective convolution for compression, applying de-allocation to DenseNet-BC-190 | DenseNet-BC-190 · CIFAR-10 | FLOPs: 9,388M -> 4,287M | Helped |
| Replace CondenseNet LGC layers with selective convolution using de-allocation | CondenseNet-182 · CIFAR-10 | test error rate: 3.76 -> 3.50 | Helped |
| Add spatial shifting to re-allocated channels with zero-initialized weights (ablation) | DenseNet-40 SelectConv · CIFAR-10 | test error rate | Helped |
| Remove spatial shifting from zero-initialized re-allocation (ablation) | DenseNet-40 SelectConv · CIFAR-10 | test error rate | Hurt |
| Use random weight re-initialization instead of zero initialization and spatial shifting (ablation) | DenseNet-40 SelectConv · CIFAR-10 | test error rate | Hurt |
| Remove re-allocation and use de-allocation only (ablation) | DenseNet-40 SelectConv · CIFAR-10 | test error rate | Hurt |
| Apply learned spatial shifts to all channels without de/re-allocation (ablation) | DenseNet-40 · CIFAR-10 | test error rate | Helped |
Closest papers by meaning
- ChannelNets: Compact and Efficient Convolutional Neural Networks via Channel-Wise Convolutions (NeurIPS 2018 · similarity 0.88)
- Rethinking the Smaller-Norm-Less-Informative Assumption in Channel Pruning of Convolution Layers (ICLR 2018 · similarity 0.87)
- Refining Architectures of Deep Convolutional Neural Networks (CVPR 2016 · similarity 0.87)
- CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters (CVPR 2022 · similarity 0.86)
- The Unreasonable Effectiveness of Fully-Connected Layers for Low-Data Regimes (NeurIPS 2022 · similarity 0.86)
- Pick-or-Mix: Dynamic Channel Sampling for ConvNets (CVPR 2024 · similarity 0.84)
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.