Research map / Architectures and efficiency / Neural architecture search
LANA: Latency Aware Network Acceleration
Pavlo Molchanov, Jimmy Hall, Hongxu Yin, Jan Kautz, Nicolo Fusi, Arash Vahdat
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 layer-wise instead of block-wise distillation for pretraining (ablation) | LANA candidate pretraining · EfficientNetV1-B2 teacher mimicking | top-1 accuracy (%): 68.21 -> 76.52 | Improved |
| use layer-wise instead of block-wise distillation (ablation) | LANA candidate pretraining · EfficientNetV2-B3 teacher mimicking | top-1 accuracy (%): 73.53 -> 76.52 | Improved |
| use layer-wise instead of block-wise distillation (ablation) | LANA candidate pretraining · ResNeST-50 teacher mimicking | top-1 accuracy (%): 77.54 -> 80.61 | Improved |
| use ILP search instead of random sampling (ablation) | LANA architecture search · 0.45xEfficientNetV1-B2 | top-1 accuracy (%): 76.44 -> 79.28 | Improved |
| use ILP instead of SNAS differentiable search (ablation) | LANA architecture search · 0.45xEfficientNetV1-B2 | top-1 accuracy (%): 74.20 -> 79.28 | Improved |
| use ILP instead of E-NAS REINFORCE search (ablation) | LANA architecture search · 0.45xEfficientNetV1-B2 | top-1 accuracy (%): 78.85 -> 79.28 | Improved |
| increase candidate pool from 2 to all 197 operations per layer (ablation) | LANA search · 0.55xEfficientNetV1-B2 | top-1 accuracy (%): 79.40 -> 80.00 | Improved |
| increase candidate pool from 5 to 10 operations per layer (ablation) | LANA search · 0.55xEfficientNetV1-B2 | top-1 accuracy (%): 79.52 -> 79.66 | Improved |
| load pretrained candidate weights instead of reinitializing before finetuning (ablation) | LANA 0.45xB2 · ImageNet-1K val | top-1 accuracy (%): 79.42 -> 79.71 | Improved |
| add knowledge distillation during finetuning (ablation) | LANA 0.45xB2 · ImageNet-1K val | top-1 accuracy (%): 0.65% degradation without | Improved |
| use LANA 0.45xB2 instead of distilled vanilla EfficientNetV1-B0 (ablation) | student architecture · ImageNet-1K val | top-1 accuracy (%): 78.72 -> 79.71 | Improved |
| add all operations beyond skip-only single-shot (ablation) | LANA 0.45xEfficientNetV1-B2 · ImageNet-1K val | top-1 accuracy (%): 78.68 -> 79.71 | Improved |
Closest papers by meaning
- Distilling Optimal Neural Networks: Rapid Search in Diverse Spaces (ICCV 2021 · similarity 0.87)
- AOWS: Adaptive and Optimal Network Width Search With Latency Constraints (CVPR 2020 · similarity 0.87)
- Searching for Fast Model Families on Datacenter Accelerators (CVPR 2021 · similarity 0.86)
- Sample-Efficient Neural Architecture Search by Learning Actions for Monte Carlo Tree Search (TPAMI 2021 · similarity 0.86)
- Subnet-Aware Dynamic Supernet Training for Neural Architecture Search (CVPR 2025 · similarity 0.86)
- HEP-NAS: Towards Efficient Few-shot Neural Architecture Search via Hierarchical Edge Partitioning (AAAI 2025 · similarity 0.85)
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.
