Research map / Architectures and efficiency / Neural architecture search
CE-NAS: An End-to-End Carbon-Efficient Neural Architecture Search Framework
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CE-NAS, a carbon-aware neural architecture search framework that uses an RL agent to allocate GPUs between one/few-shot and vanilla evaluation based on forecasted grid carbon intensity from a time-series transformer, integrated with LaMOO and qNEHVI.
Problem
Neural architecture search has a very high carbon footprint, and existing efficient NAS methods ignore temporal and geographic variation in grid carbon intensity.
Approach
Decouple NAS into energy-efficient sampling (supernet-based one/few-shot evaluation) and energy-intensive training, and use an RL agent that shifts GPU resources between them according to predicted carbon intensity, with LaMOO partitioning the search space.
Key results
- On HW-NASBench, CE-NAS reduces carbon emissions by up to 7.22X relative to vanilla LaMOO while staying close in hypervolume (about 4100 vs 4117).
- On CIFAR-10, CE-Net-P2 reaches 97.84% top-1 accuracy (text also cites 97.35% for CE-Net-P1 with 1.68M params) at 38.53 lbs CO2.
- On ImageNet, CE-Net-G2 achieves 80.6% top-1 (table lists 19.4% error) at 0.78 ms TensorRT FP16 latency on V100, with 909.86 lbs CO2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine carbon-aware GPU allocation with one-shot sampling and true architecture evaluation | NAS on HW-NASBench · HW-NASBench | relative carbon emission: 3.24X-7.22X carbon compared to other approaches | Helped |
| combine carbon-aware GPU allocation with one-shot sampling and true architecture evaluation | NAS on HW-NASBench · HW-NASBench | hypervolume: 4100 and 4117, differing only by 0.034% | Helped |
| use CE-NAS instead of vanilla or one-shot NAS under a carbon budget | NAS on HW-NASBench · HW-NASBench | search efficiency: up to 5X | Helped |
| use CE-NAS rather than the heuristic GPU allocation strategy | NAS on NasBench301 · NasBench301 | hypervolume | Helped |
| search a LaMOO-selected region rather than the entire search space (ablation) | random architecture sampling · HW-NASBench | average accuracy: 68.28 -> 72.12 | Helped |
| search a LaMOO-selected region rather than the entire search space (ablation) | random architecture sampling · HW-NASBench | average edge GPU energy: 22.84 mJ -> 16.59 mJ | Helped |
| use predicted rather than actual carbon intensity (ablation) | CE-NAS · HW-NASBench, carbon budgets 5000g–50000g | hypervolume: 3724.84 vs 3742.36; 3774.21 vs 3796.93; 3891.70 vs 3898.43; 3966.70 vs 3971.16 | Hurt |
| use continuous rather than discrete RL actions (ablation) | CE-NAS · HW-NASBench, carbon budgets 5000g–50000g | hypervolume: 3724.84 vs 3645.21; 3774.21 vs 3748.83; 3891.70 vs 3853.86; 3966.70 vs 3920.47 | Hurt |
Surprising findings
- Continuous RL action space performs slightly worse than the discrete action space, suggesting fine-grained allocation is unnecessary.
Limitations
- Carbon cost of RL agent and time-series transformer training is not counted, by the authors' own assumption of negligible cost.
- Results on open-domain tasks are reported with inconsistent numbers (97.35% vs 97.84% for CIFAR-10; 80.6% accuracy vs 19.4% error table entry).
Experimental setup
Datasets: ElectricityMap carbon intensity traces HW-NASBench NasBench201 NasBench301 CIFAR-10 ImageNet
Benchmarks: HW-NASBench NasBench301 CIFAR-10 ImageNet
Metrics: hypervolume log hypervolume difference carbon emission (CO2 lbs) top-1 error TensorRT latency MAPE
Baselines: Vanilla LaMOO One-shot LaMOO heuristic GPU allocation strategy STCF DACF CC PNAS NASNet-A DARTS OFA-Net PC-DARTS
Builds on: LaMOO qNEHVI one-shot NAS few-shot NAS actor-critic
Closest papers by meaning
- Searching for a Robust Neural Architecture in Four GPU Hours (CVPR 2019 · similarity 0.88)
- NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search (ICLR 2020 · similarity 0.88)
- Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective (ICLR 2021 · similarity 0.88)
- NAS-Bench-101: Towards Reproducible Neural Architecture Search (ICML 2019 · similarity 0.87)
- MOTE-NAS: Multi-Objective Training-based Estimate for Efficient Neural Architecture Search (NeurIPS 2024 · similarity 0.86)
- EG-NAS: Neural Architecture Search with Fast Evolutionary Exploration (AAAI 2024 · similarity 0.86)
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.