Research map / Architectures and efficiency / RNNs and normalization
Prioritized Subnet Sampling for Resource-Adaptive Supernet Training
Bohong Chen, Mingbao Lin, Rongrong Ji, Liujuan Cao
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 |
|---|---|---|---|
| use prioritized subnet pools instead of uniform sampling (vs USNet, MobileNet-V1 ~110M FLOPs) | USNet · ImageNet val | top-1 accuracy (%): 61.9 -> 67.9 | Improved |
| use prioritized subnet sampling instead of MutualNet (MobileNet-V1 ~110M FLOPs) | MutualNet · ImageNet val | top-1 accuracy (%): 63.5 -> 67.9 | Improved |
| train under multiple constraints (PSS-Net_M) instead of FLOPs only | PSS-Net · ImageNet val (MobileNet-V1 569M) | top-1 accuracy (%): 74.2 -> 74.5 | No clear effect |
| add prioritized subnet sampling vs random search at 300M FLOPs (ablation) | PSS-Net · ImageNet val (MobileNet-V1) | top-1 accuracy (%): 71.2 -> 72.3 (+1.1) | Improved |
| add prioritized subnet sampling vs random search at 400M FLOPs (ablation) | PSS-Net · ImageNet val (MobileNet-V1) | top-1 accuracy (%): 72.5 -> 72.9 (+0.4) | Improved |
| add prioritized subnet sampling vs random search at 500M FLOPs (ablation) | PSS-Net · ImageNet val (MobileNet-V1) | top-1 accuracy (%): 73.7 -> 73.8 (+0.1) | No clear effect |
| calibrate top-5 subnets instead of top-1 for BN (ablation) | PSS-Net · ImageNet val (MobileNet-V1) 300M | top-1 accuracy (%): 71.4 -> 72.3 (+0.9) | Improved |
| increase calibrated subnets from top-5 to top-20 (ablation) | PSS-Net · ImageNet val (MobileNet-V1) 300M | top-1 accuracy (%): 72.3 -> 72.4 | No clear effect |
| increase pool size M from 10 to 500 (baseline 50) (ablation) | PSS-Net · ImageNet val (MobileNet-V1) | top-1 accuracy (%): avg boost -0.3 (M=10) vs +0.1 (M=500) | Improved |
| vary p_end (0.1, 0.001) from 0.01 (ablation) | PSS-Net · ImageNet val (MobileNet-V1) | top-1 accuracy (%): avg boost -0.1 and +0.0 | No clear effect |
| vary eta_end (0.1, 0.001) from 0.01 (ablation) | PSS-Net · ImageNet val (MobileNet-V1) | top-1 accuracy (%): avg boost -0.2 and -0.1 | No clear effect |
| use channel divisor 4 instead of 8 / 16 hurts (ablation) | PSS-Net · ImageNet val (MobileNet-V1) 300M | top-1 accuracy (%): divisor 16: 71.2; divisor 8: 72.3; divisor 4: 72.5 | Improved |
Closest papers by meaning
- Subnet-Aware Dynamic Supernet Training for Neural Architecture Search (CVPR 2025 · similarity 0.87)
- GreedyNAS: Towards Fast One-Shot NAS With Greedy Supernet (CVPR 2020 · similarity 0.85)
- ElasticViT: Conflict-aware Supernet Training for Deploying Fast Vision Transformer on Diverse Mobile Devices (ICCV 2023 · similarity 0.85)
- AlphaNet: Improved Training of Supernets with Alpha-Divergence (ICML 2021 · similarity 0.85)
- Resolution Adaptive Networks for Efficient Inference (CVPR 2020 · similarity 0.85)
- Adaptive Depth Networks with Skippable Sub-Paths (NeurIPS 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.
