Research map / Image generation / Super-resolution
Universally Slimmable Networks and Improved Training Techniques
Jiahui Yu, Thomas S. Huang
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 |
|---|---|---|---|
| add both smallest and largest width to the sampled widths (sandwich rule) instead of 3 random widths (ablation) | US-MobileNet v1 · ImageNet val (AVG over widths) | top-1 error (%): 38.20 -> 36.45 | Improved |
| always train the smallest width in addition to 2 random widths (ablation) | US-MobileNet v1 · ImageNet val (AVG over widths) | top-1 error (%): 38.20 -> 36.73 | Improved |
| always train the largest width in addition to 2 random widths (ablation) | US-MobileNet v1 · ImageNet val (AVG over widths) | top-1 error (%): 38.20 -> 38.70 | Mixed |
| add inplace distillation from the largest width (ablation) | US-MobileNet v1 · ImageNet val (FLOPs-accuracy spectrum) | top-1 error (%) | Improved |
| add inplace distillation from the largest width (ablation) | US-WDSR · DIV2K x2 super-resolution | PSNR | Improved |
| replace individually trained networks with a single US-Net | MobileNet v1 · ImageNet val (AVG over widths) | top-1 error (%): 36.9 -> 34.7 | Improved |
| replace 4-switch slimmable network with a US-Net | MobileNet v2 · ImageNet val (AVG over widths) | top-1 error (%): 34.1 -> 32.9 | Improved |
| replace individually trained models with a US-WDSR | WDSR · DIV2K x2 super-resolution | PSNR: 0.01 lower PSNR | Hurt |
| replace individually trained networks with a US-Net | Atari Breakout PPO · BreakoutNoFrameskip-v4 | mean episode reward | Improved |
| use exact average instead of moving average for BN post-statistics (ablation) | US-MobileNet v1 · ImageNet val (0.25x) | top-1 error (%): 44.4 -> 44.3 | No clear effect |
| compute BN post-statistics on 1k samples instead of 1.28M (ablation) | US-MobileNet v1 · ImageNet val (0.25x) | top-1 error (%): 44.3 -> 44.4 | No clear effect |
| add output averaging by number of input channels (ablation) | US-MobileNet v1 · ImageNet val (AVG over widths) | top-1 error (%): 34.7 -> 34.3 | Improved |
Closest papers by meaning
- Slimmable Neural Networks (ICLR 2019 · similarity 0.93)
- Three Guidelines You Should Know for Universally Slimmable Self-Supervised Learning (CVPR 2023 · similarity 0.88)
- Dynamic Slimmable Network (CVPR 2021 · similarity 0.88)
- Network Sparsity Unlocks the Scaling Potential of Deep Reinforcement Learning (ICML 2025 · similarity 0.82)
- $\mu$PC: Scaling Predictive Coding to 100+ Layer Networks (NeurIPS 2025 · similarity 0.82)
- Training a Predictive Coding Network on ImageNet using Equilibrium Propagation (NeurIPS 2026 · similarity 0.82)
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.
