Research map / Architectures and efficiency / Pruning and efficient tuning
Dynamic Slimmable Network
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 |
|---|---|---|---|
| Use dynamic channel slicing with DS-Net instead of the strongest reported static or dynamic compression baseline | DS-ResNet and DS-MBNet · ImageNet val | Top-1 Acc.: 76.6 vs 76.2; 76.1 vs 75.6; 74.6 vs 73.8; 74.5 vs 73.0; 72.8 vs 71.5; 70.1 vs 67.9 | Helped |
| Transfer the ImageNet-trained dynamic gate | DS-ResNet · CIFAR-10 | Top-1 Acc.: 97.4 -> 97.8 | Helped |
| Use the EMA target rather than the online widest subnet as distillation teacher (ablation) | Slimmable MobileNetV1 supernet · ImageNet val, slimmest subnet | Top-1 accuracy: 66.5 -> 68.1 | Helped |
| Add an ensemble of target subnets to EMA-based in-place distillation (ablation) | Slimmable MobileNetV1 supernet · ImageNet val, slimmest subnet | Top-1 accuracy: 68.1 -> 68.3 | Helped |
| Use IEB (EMA target and ensemble supervision) instead of original in-place distillation (ablation) | Slimmable MobileNetV1 supernet · ImageNet val, widest subnet | Top-1 accuracy: 74.0 -> 74.6 | Helped |
| Add SGS with Try Best targets rather than target loss alone (ablation) | DS-ResNet gate · ImageNet val | Top-1 Acc.: 76.8 -> 76.6; MAdds 3.6B -> 3.1B | Mixed |
| Use complexity loss alone (ablation) | DS-ResNet gate · ImageNet val | Top-1 Acc.: 76.8 -> 66.2 | Hurt |
| Set hard samples to the slimmest path (Give Up) instead of the widest path (Try Best) (ablation) | DS-ResNet gate · ImageNet val | Top-1 Acc.: 76.6 -> 73.7 | Hurt |
| Share the first fully connected layer between attention and slimming gate heads (ablation) | DS-ResNet gate · ImageNet val | Top-1 Acc.: 72.7 -> 76.6 | Helped |
| Use a one-hot slimming head instead of a scalar slimming head (ablation) | DS-ResNet gate · ImageNet val | Top-1 Acc.: 73.6 -> 76.6 | Helped |
| Enable dynamic slimming gate routing rather than using a similarly sized fixed supernet subnet (ablation) | DS-MBNet and DS-ResNet · ImageNet val | Top-1 Acc.: 69.3 -> 70.1; 73.4 -> 74.6 | Helped |
| Raise the distillation temperature from τ=1 to τ=4 (ablation) | DS-MBNet supernet · ImageNet val, slimmest subnet | Top-1 accuracy: 59.2 -> 49.0 | Hurt |
Closest papers by meaning
- Dynamic Sparse Graph for Efficient Deep Learning (ICLR 2019 · similarity 0.88)
- Dynamic Sparse Training: Find Efficient Sparse Network From Scratch With Trainable Masked Layers (ICLR 2020 · similarity 0.87)
- Slimmable Neural Networks (ICLR 2019 · similarity 0.87)
- Dynamic Filter Networks (NeurIPS 2016 · similarity 0.87)
- Advancing Dynamic Sparse Training by Exploring Optimization Opportunities (ICML 2024 · similarity 0.87)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · 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.