Research map / Architectures and efficiency / Neural architecture search
TF-NAS: Rethinking Three Search Freedoms of Latency-Constrained Differentiable Neural Architecture Search
Yibo Hu, Xiang Wu, Ran He
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 random sampling for a second path (bi-sampling) in addition to Gumbel (ablation) | TF-NAS · ImageNet (15ms target, second-path sampling ablation) | top-1 accuracy: 75.8 -> 76.3 | Improved |
| add a second Gumbel-sampled path (ablation) | TF-NAS · ImageNet (15ms target, second-path sampling ablation) | top-1 accuracy: 75.8 -> 75.7 | Hurt |
| add a second path chosen by max architecture parameter (ablation) | TF-NAS · ImageNet (15ms target, second-path sampling ablation) | top-1 accuracy: 75.8 -> 75.5 | Hurt |
| replace skip-as-candidate with sink-connecting depth search (ablation) | TF-NAS · ImageNet (15ms target, depth-level search space) | top-1 accuracy: 75.6 -> 76.3 | Improved |
| take skip out of the candidate set (mutual exclusion) (ablation) | TF-NAS · ImageNet (15ms target, depth-level search space) | top-1 accuracy: 75.6 -> 76.1 | Improved |
| eliminate architecture redundancy via sink-connecting (ablation) | TF-NAS · ImageNet (15ms target, depth-level search space) | top-1 accuracy: 76.1 -> 76.3 | Improved |
| add elasticity-scaling to the C1 latency objective (ablation) | C1 objective · ImageNet (15ms target, latency objective) | top-1 accuracy: 75.6 -> 75.9 | Improved |
| add elasticity-scaling to the C2 latency objective (ablation) | C2 objective · ImageNet (15ms target, latency objective) | top-1 accuracy: 76.2 -> 76.1 | No clear effect |
| add elasticity-scaling to the proposed target-latency objective (ablation) | TF-NAS · ImageNet (15ms target, latency objective) | top-1 accuracy: 76.3 -> 76.3 | No clear effect |
| use the proposed target-latency objective instead of C1 (no elasticity-scaling) (ablation) | latency objective · ImageNet (15ms target, latency objective) | top-1 accuracy: 75.6 -> 76.3 | Improved |
| set lambda to 0.1 instead of 0.01 (ablation) | TF-NAS · ImageNet (15ms target, lambda sweep) | top-1 accuracy: 75.7 -> 76.3 | Improved |
| add SE module to candidate operations (TF-NAS-A) (ablation) | TF-NAS-A · ImageNet (mobile setting) | top-1 accuracy: 76.5 -> 76.9 | Improved |
Closest papers by meaning
- Progressive Differentiable Architecture Search: Bridging the Depth Gap Between Search and Evaluation (ICCV 2019 · similarity 0.90)
- Improving Differentiable Neural Architecture Search by Encouraging Transferability (ICLR 2023 · similarity 0.89)
- Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective (ICLR 2021 · similarity 0.89)
- Zero-Cost Operation Scoring in Differentiable Architecture Search (AAAI 2023 · similarity 0.89)
- BaLeNAS: Differentiable Architecture Search via the Bayesian Learning Rule (CVPR 2022 · similarity 0.89)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 2025 · similarity 0.88)
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.
