Research map / Architectures and efficiency / Neural architecture search
Auto-DAS: Automated Proxy Discovery for Training-free Distillation-aware Architecture Search
Haosen Sun, Lujun Li, Peijie Dong, Zimian Wei, Shitong Shao
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 |
|---|---|---|---|
| evolve proxies from scratch over instinct and interaction statistics instead of manual DisWOT proxy | Auto-DAS · NAS-Bench-101-KD | Spearman correlation (%): 52.45 -> 85.95 | Improved |
| use DisWOT handcrafted proxy on ViT search space | DisWOT · ViT-Bench PiT CIFAR-100 | Kendall correlation (%): -19.05 | Hurt |
| use Auto-DAS searched student instead of DisWOT student under equal constraints | NDS-ResNet CIFAR-100 · NDS-ResNet CIFAR-100 | accuracy (%): 74.23 -> 74.66 (1M params) | Improved |
| use Auto-DAS searched students on ImageNet instead of DisWOT | ResNet18/ResNet50 · ImageNet-1k | Top-1 accuracy (%): 72.08 -> 72.58 (ResNet18); 78.62 -> 79.25 (ResNet50) | Improved |
| combine instinct and interaction proxy inputs and operations (ablation) | Auto-DAS proxy · PiT CIFAR-100 ranking | Spearman correlation (%): 83.25 -> 90.32 (instinct-only vs full) | Improved |
| use interaction proxy instead of instinct proxy only (ablation) | Auto-DAS proxy · ResNet / NB-101 ranking | Spearman correlation (%): ResNet 80.52 -> 92.82; NB-101 75.86 -> 50.88 | Mixed |
| use gradient-free Auto-DAS (ablation) | Auto-DAS · PiT CIFAR-100 | distilled accuracy (%) / search cost: 77.88 -> 77.72; 1.28 h -> 0.82h | No clear effect |
| use adaptive-elite selection in the evolution (ablation) | Auto-DAS proxy search · ResNet proxy search curve | search result | Improved |
| apply Auto-DAS searched student under other KD methods (FitNets, AT, SP, RKD, CRD) | 1M-parameter students · NDS-ResNet CIFAR-100 | accuracy (%): 74.85 -> 75.28 (FitNets); 75.25 -> 75.86 (CRD) | Improved |
Closest papers by meaning
- Auto-GAS: Automated Proxy Discovery for Training-free Generative Architecture Search (ECCV 2024 · similarity 0.88)
- UniADS: Universal Architecture-Distiller Search for Distillation Gap (AAAI 2024 · similarity 0.84)
- DisWOT: Student Architecture Search for Distillation WithOut Training (CVPR 2023 · similarity 0.80)
- Target-Driven Distillation: Consistency Distillation with Target Timestep Selection and Decoupled Guidance (AAAI 2025 · similarity 0.79)
- Distillation Scaling Laws (ICML 2025 · similarity 0.77)
- Structuring Open-Ended NAS: Semi-Automated Design Knowledge Structuring with LLMs for Efficient Neural Architecture Search (NeurIPS 2026 · similarity 0.77)
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.
