Research map / Learning theory and algorithms / Prediction and decision losses
Designing Network Design Spaces
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 |
|---|---|---|---|
| Share the bottleneck ratio across stages (ablation) | AnyNetX · ImageNet, 400MF, 10 epochs | error distribution | No clear effect |
| Share group width across stages (ablation) | AnyNetX with shared bottleneck ratio · ImageNet, 400MF, 10 epochs | error distribution | No clear effect |
| Require stage widths to be non-decreasing (ablation) | AnyNetX · ImageNet, 400MF, 10 epochs | error distribution | Helped |
| Require stage depths to be non-decreasing (ablation) | AnyNetX with non-decreasing widths · ImageNet, 400MF, 10 epochs | error distribution | Helped |
| Constrain widths and depths to a quantized linear structure (ablation) | AnyNetX · ImageNet, low-compute, low-epoch | error distribution | Helped |
| Set width multiplier to 2 (ablation) | RegNetX · ImageNet, low-compute, low-epoch | error distribution | Helped |
| Set initial width equal to linear slope (ablation) | RegNetX · ImageNet, low-compute, low-epoch | error distribution | Helped |
| Add SE operation (ablation) | RegNetX · ImageNet, across flop regimes | error distribution | Helped |
| Use inverted bottleneck with b<1 (ablation) | RegNetX · ImageNet, across flop regimes | error distribution | Hurt |
| Use depthwise convolution with g=1 (ablation) | RegNetX · ImageNet, across flop regimes | error distribution | Hurt |
| Vary input resolution instead of fixing it at 224×224 (ablation) | RegNetX · ImageNet, higher flop regimes | error distribution | Hurt |
| Fix network depth at 20 blocks (ablation) | RegNet · ImageNet, all tested flop regimes | top-1 error | No clear effect |
Closest papers by meaning
- Rewrite the Stars (CVPR 2024 · similarity 0.85)
- Latent Geometry-Driven Network Automata for Complex Network Dismantling (ICLR 2026 · similarity 0.83)
- Unified Scaling Laws for Routed Language Models (ICML 2022 · similarity 0.83)
- Diversity and Depth in Per-Example Routing Models (ICLR 2019 · similarity 0.83)
- Design Space for Graph Neural Networks (NeurIPS 2020 · similarity 0.83)
- AZ-NAS: Assembling Zero-Cost Proxies for Network Architecture Search (CVPR 2024 · 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.