Pith. sign in

REVIEW 4 major objections 7 minor 17 references

GreenAuto: An Automated Platform for Sustainable AI Model Design on Edge Devices

T0 review · 4 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims GreenAuto, an automated pipeline, can find image classifiers for edge phones that match or beat MobileNet-V2's accuracy while using a fraction of the energy, without human intervention.

desk verdict A concrete automated energy-aware NAS demo with real measured gains on one phone; the gradient-guided search step is underspecified, so the measured results are worth more than the stated mechanism. read the letter →

arxiv 2501.14995 v1 pith:3NS6LKOI submitted 2025-01-25 cs.LG

Xiaolong Tu , Dawei Chen , Kyungtae Han , Onur Altintas , Haoxin Wang This is my paper · ORCID
classification cs.LG
keywords sustainableAIenergy-efficientdeeplearningneuralarchitecturesearchParetofrontedgedeviceson-deviceenergymeasurementcarbonfootprintgradientdescentsampling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

GreenAuto is an end-to-end automated platform for designing image-classification models on edge devices, optimizing for both accuracy and measured inference energy. The paper claims the pipeline explores a search space of over 959,000 model variants and finds models that outperform a standard mobile baseline while using a fraction of its energy, with no human in the loop. If this holds, energy-efficient on-device AI can be treated as an automated search problem rather than a manual tuning exercise, and the approach could scale to billions of edge devices.

What carries the argument

The load-bearing mechanism is a three-stage loop: an expanded NAS search space varying output channels, kernel sizes, and strides (959,417 models); a Pareto-front multi-objective search that uses kernel-level energy-predictor estimates and NASWOT accuracy scores, guided by a gradient descent direction g*(x) (a weighted combination of per-objective gradients) to sample the next model candidates; and an automated on-device energy measurement using an external power monitor synchronized by inference start and stop triggers. The iteratively updated Pareto front and a final gradient-magnitude-based selection procedure pick the best model.

What would settle it

Run the same experiment with the same 170-model training budget on the same edge device using randomly sampled architectures; if random sampling reaches comparable accuracy-energy Pareto points or finds a model with similar accuracy and energy, the claimed gradient-guided sampling efficiency is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that an automated pipeline combining an expanded neural architecture search space, pre-trained kernel-level energy predictors, a zero-cost accuracy proxy, Pareto-front search with gradient-guided sampling, and automated on-device power measurement can efficiently identify sustainable DNN models. On CIFAR-10 with a smartphone edge device, GreenAuto reports best models of 0.93 accuracy at 4.27 mJ and 0.88 accuracy at 0.2 mJ, both outperforming MobileNet-V2's 0.88 accuracy and 7.53 mJ on the same device. The search trained only 170 models on a single GPU over roughly one week, with a per-model search carbon footprint of 0.013 kgCO2 versus 0.231 kgCO2 for the NASNet-A baseline. The authors conclude that this pipeline demonstrates efficient identification of sustainable models without human intervention.

Load-bearing premise

The search loop assumes that 'gradient descent' over a discrete set of architectural choices—output channel counts, kernel sizes, and strides—produces meaningful guidance, even though the paper does not define a differentiable parameterization of the search space.

Editorial extensions

If this is right

  • Sustainable model search can run without human intervention, automating the measurement, training, and refinement loop on a target edge device.
  • The search cost is drastically lower than conventional NAS: 170 trained models versus 20,000, with per-model search carbon emissions of 0.013 kgCO2 compared with 0.231 kgCO2 for NASNet-A.
  • Models found by the pipeline can beat a common mobile baseline: 0.93 accuracy at 4.27 mJ and 0.88 accuracy at 0.2 mJ versus MobileNet-V2's 0.88 accuracy at 7.53 mJ.
  • Because the platform is modular and reconfigurable, the same pipeline can be redirected to other objectives, hardware, or search-space definitions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: if the kernel-level energy predictors transfer to new devices without retraining, the same global-view search could be reused across hardware; the paper does not establish this transferability.
  • Editorial extension: the paper never defines a differentiable relaxation of the discrete search space, so the gradient-guided sampling may be equivalent to a heuristic; a random-search or evolutionary baseline with the same measurement budget could test whether the gradients are doing real work.
  • Editorial extension: the carbon comparison is per-model, not total search cost; the reported ratio depends on the number and type of GPUs used, which differ between the two searches.
  • Editorial extension: the same Pareto-front pipeline could be extended to transformer-based or LLM-style models on edge devices, as the authors note in their future-work section, though energy predictors for those architectures would be needed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. GreenAuto is an end-to-end automated platform for neural architecture search targeting energy-efficient image classifiers on edge devices. The paper expands the NAS-Bench-201 cell space by varying output channels, kernel sizes, and strides (Table 1), yielding 959,417 candidate models. A kernel-level energy predictor from the authors' prior work and the NASWOT zero-cost accuracy proxy provide initial estimates; a Pareto-front search with two proposed 'gradient descent' algorithms (Algorithms 1 and 2) iteratively selects candidates for physical training and on-device energy measurement using a Monsoon power monitor. On CIFAR-10, the platform reports a best model at 0.88 accuracy and 0.2 mJ inference energy and another at 0.93 accuracy and 4.27 mJ, both compared favorably to MobileNet-V2 (0.88, 7.53 mJ), and a carbon-footprint comparison against NASNet-A showing 0.0478% of the per-model search emissions. The paper claims the pipeline operates without human intervention.

Significance. If the results hold, GreenAuto would be a useful demonstration that an automated, hardware-aware NAS pipeline can find image classifiers with substantially lower inference energy than a common mobile baseline, and the reported physical measurements on a real device are a strength. The platform's modular design and the use of a pre-trained energy predictor with a reported Kendall's tau of 0.526 on 1,179 physically measured models provide a credible basis for further study. However, the core gradient-guided sampling mechanism is not mathematically specified for the discrete architecture space, and the external baseline comparison is not controlled; these issues currently prevent the efficiency and sustainability claims from being fully supported.

major comments (4)
  1. [Section 2.2, Eq. (4) and Algorithm 1] The search selects models by comparing each candidate's gradient g_i(x) with the optimal gradient direction g*(x), but the search space (Table 1) is discrete: output channels, kernel sizes, and strides take categorical values. The paper never defines a differentiable parameterization, continuous relaxation, embedding, or finite-difference rule that would make g_i(x) meaningful for an unevaluated architecture. Consequently, the inner-product similarity in line 9 is formally unspecified, and the claimed efficiency gain from gradient-guided sampling is not established. Please provide an explicit definition (e.g., a straight-through estimator over a continuous relaxation, or finite differences over the discrete configuration grid) or revise the claims to present the sampler as a heuristic.
  2. [Section 4, Table 2 and the carbon comparison] The comparison with NASNet-A uses different hardware (450 K40 GPUs over seven days versus a single RTX 4090 over roughly one week), different search algorithms, and different training pipelines, so the reported 0.0478% carbon-emission ratio conflates methodology with hardware generation and is not a controlled result. A fair comparison would require re-running NASNet-A (or an equivalent evolutionary baseline) on the same GPU and measurement setup, or clearly labeling the numbers as an illustrative order-of-magnitude estimate rather than a benchmark result.
  3. [Section 4, Fig. 3 and Table 2] The evaluation reports a single run without error bars, confidence intervals, or repeated-seed results. Since the initial k models and per-iteration samples are drawn from distributions (Algorithm 1, lines 3 and 9), the observed Pareto front and best models could be substantially affected by stochasticity. Please report variance over multiple independent runs, or at minimum a sensitivity analysis of the reported accuracy and energy values.
  4. [Sections 2.1 and 2.2] The energy predictor from the authors' prior work is used both as the search objective and as the guide for candidate sampling, making the search self-referential. While the final candidates are physically measured and the Kendall's tau of 0.526 on 1,179 models is an external check, the paper should clarify whether these 1,179 models are independent of the search process and discuss how predictor bias (e.g., systematic underestimation of energy for certain configurations) could affect which regions of the search space are explored. This is important because the stopping criterion and the reported best models depend on the predictor's ranking during search.
minor comments (7)
  1. [Fig. 3 and Section 4 text] The text states that the best model in Fig. 3d achieves 0.88 accuracy and 0.2 mJ, but the caption for (d) lists 0.88 accuracy and 1.00 mJ; additionally, the progression from iteration 5 to iteration 7 appears to increase energy, which would contradict the definition of a Pareto front. Please correct the caption or the text.
  2. [Section 2.2 heading] The heading contains a typo: 'Pareto fornt-Based Model Search' should read 'Pareto front-Based Model Search'.
  3. [Section 4, Fig. 3b] The text says 'In the first iteration, neither the Pareto front nor the best model changes,' but Fig. 3b is labeled 'Pareto Frontier for iteration 1' and 'Best Model of iteration 1'; please clarify whether these labels refer to the previous front or a new front.
  4. [Section 4, Table 2] The carbon-footprint notation '0.013/model1' contains a stray '1' and the metric is not defined; please clarify whether the reported value includes training energy, measurement energy, and any idle overhead.
  5. [Algorithm 1 and Section 2.2] The notation is inconsistent: the algorithm uses 'k' for the initial sample count and 'm' for the per-iteration sample count, while Section 4 states the opposite (initial model number m = 100, per-iteration k = 10). Please standardize the notation.
  6. [Section 2.2] The stopping criterion (accuracy > 0.9 and energy < 7 mJ) is introduced only in Section 4, but Algorithm 1 refers to 'predefined constraints'; the criterion should be stated in Section 2.2 where the algorithm is defined.
  7. [Section 2.1] The 'average prediction accuracy of 86.2%' for the energy predictor is cited from prior work but not defined; please state the metric (e.g., percentage of predictions within a tolerance band) so that the Kendall's tau value can be interpreted alongside it.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: final models are physically measured and the energy predictor is validated against measured data, so the reported results do not reduce to the search inputs.

full rationale

GreenAuto's derivation chain is: (1) a pre-trained kernel-level energy predictor from prior work [10] estimates energy and NASWOT [11] estimates accuracy; (2) these estimates guide Pareto-based sampling; (3) sampled models are trained and their real energy is measured on-device (§2.3); (4) the Pareto front is updated from these real measurements and the final best models in Table 2 are selected from that measured front. None of the reported final accuracies or energies is defined as the output of the predictor: the predictor is used only as a search heuristic, and the final candidates are physically measured. The paper also checks the predictor against real measurements (Kendall's Tau = 0.526 over 1,179 models), so the self-citation to [10] is not the sole support for the predictor's validity. The main specification gap is that Algorithm 1 uses gradients g_i(x) over a discrete architecture space without defining a relaxation or finite-difference rule, so the sampling step is underspecified; this is a correctness/completeness risk, not a circularity, because no equation or fitted parameter makes the claimed result true by construction. The stated limitation that the search space has a fixed skeleton also concerns scope, not self-reference. Therefore no circular step rises to the level of the defined patterns.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central demonstration rests on a fitted energy predictor from prior work, hand-set search weights and stopping thresholds, and an implicitly differentiable treatment of a discrete search space. The paper provides moderate correlational evidence for the predictor and physical measurements for final models, but the reader must accept several unstated modeling choices.

free parameters (6)
  • Kernel-level energy predictor parameters = not reported here; from prior work [10]
    The estimated energy E_p for all 959,417 candidates is produced by this fitted predictor, and its reliability in this paper is summarized only by Kendall's Tau of 0.526 on 1,179 models.
  • Search weights w_s = w_s_acc=1, w_s_ene=3 in evaluation
    Chosen by the authors to bias the search toward energy efficiency; different weights would change the reported best model and stopping behavior.
  • Model selection weights w_d = configurations (10,1) and (1,10) for GD, (1,1) for WS
    Hand-set weights in Algorithm 2 determine which Pareto-front point is called 'best' in the evaluation.
  • Stopping criterion thresholds = accuracy > 0.9, energy < 7 mJ
    Chosen from MobileNet-V2's measured performance on the same platform; the threshold directly determines when the search stops and which results are reported.
  • Initial and per-iteration sample counts = k=100 initial, m=10 per iteration
    Search hyperparameters set in the evaluation, with no sensitivity analysis across different values.
  • Expanded search space ranges = C_out in {1,2,...,10,16,32,64,128,256}, kernel sizes {1,3,5,7}, strides {1,2}
    Manual expansion of NAS-Bench-201 to create 959,417 models; different ranges would change all reported results.
assumptions (6)
  • standard math Pareto dominance and multiple gradient descent formulations in Eqs. (1) to (4)
    Used as stated for multi-objective optimization; these are standard mathematical tools.
  • domain assumption Energy consumption is a practical and quantifiable proxy for sustainability
    Introduced in Section 1; reasonable for electrically powered edge devices, but it is a modeling assumption, not a physical law.
  • domain assumption NASWOT score correlates with validation accuracy in the expanded search space
    Used as the accuracy proxy in Eq. (3) and throughout Algorithm 1, but no evidence is given that NASWOT remains predictive in this expanded configuration space.
  • domain assumption The kernel-level energy predictor generalizes to the expanded space and target device
    The only evidence is Kendall's Tau of 0.526 on 1,179 models, which is moderate and does not guarantee reliable ranking across the full 959,417 model space.
  • ad hoc to paper The discrete architecture space is differentiable enough for gradient descent
    Algorithm 1 assumes gradients g_i(x) can be computed over discrete choices of output channels, kernel size, and stride, but no differentiable relaxation or surrogate is defined.
  • domain assumption Monsoon power monitor measurements correctly capture inference energy on the Android device
    The setup in Section 3 uses an external power monitor with BMS isolation, but no calibration or validation data is shown in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GreenAuto: An Automated Platform for Sustainable AI Model Design on Edge Devices." pith.science (2026). https://pith.science/paper/3NS6LKOI

@misc{pith2026250114995,
  author       = {Pith},
  title        = {Pith review of: GreenAuto: An Automated Platform for Sustainable AI Model Design on Edge Devices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3NS6LKOI}},
  note         = {Machine review of arXiv:2501.14995}
}
read the original abstract

We present GreenAuto, an end-to-end automated platform designed for sustainable AI model exploration, generation, deployment, and evaluation. GreenAuto employs a Pareto front-based search method within an expanded neural architecture search (NAS) space, guided by gradient descent to optimize model exploration. Pre-trained kernel-level energy predictors estimate energy consumption across all models, providing a global view that directs the search toward more sustainable solutions. By automating performance measurements and iteratively refining the search process, GreenAuto demonstrates the efficient identification of sustainable AI models without the need for human intervention.

Figures

Figures reproduced from arXiv: 2501.14995 by the authors.

Figure 1
Figure 1. System overview of GreenAuto, operating autonomously without human intervention. 2.2 Pareto fornt-Based Model Search Building on our energy efficiency-driven search space and the corre￾sponding energy consumption and accuracy estimates, GreenAuto explores DNN models that achieve better energy efficiency while maintaining or exceeding current levels of inference accuracy. This search process can be formulated as a mu… view at source ↗
Figure 2
Figure 2. Timing sync between edge device and external [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Pareto front update and the best model selection in each iteration. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 14 canonical work pages

  1. [1]

    Sustain- able AI: Environmental implications, challenges and opportunities

    Carole-Jean Wu, Ramya Raghavendra, Udit Gupta, Bilge Acun, Newsha Ardalani, Kiwan Maeng, Gloria Chang, Fiona Aga, Jinshi Huang, Charles Bai, et al. Sustain- able AI: Environmental implications, challenges and opportunities. Proc. MLSys, 4:795–813, 2022

    work page 2022

  2. [2]

    DeepEn2023: Energy Datasets for Edge Artificial Intelligence

    Xiaolong Tu, Anik Mallik, Haoxin Wang, and Jiang Xie. Deepen2023: Energy datasets for edge artificial intelligence. arXiv preprint arXiv:2312.00103, 2023

    work page Pith review arXiv 2023

  3. [3]

    Sustainable AI: AI for sustainability and the sustain- ability of AI

    Aimee Van Wynsberghe. Sustainable AI: AI for sustainability and the sustain- ability of AI. AI and Ethics, 1(3):213–218, 2021

    work page 2021

  4. [4]

    The role of artificial intelligence in achiev- ing the sustainable development goals

    Ricardo Vinuesa, Hossein Azizpour, Iolanda Leite, Madeline Balaam, Virginia Dignum, Sami Domisch, Anna Felländer, Simone Daniela Langhans, Max Tegmark, and Francesco Fuso Nerini. The role of artificial intelligence in achiev- ing the sustainable development goals. Nature Communications, 11(1):1–10, 2020

    work page 2020

  5. [5]

    https://www.eia.gov/energyexplained/ energy-and-the-environment/where-greenhouse-gases-come-from.php?utm_ source=chatgpt.com

    Energy and the environment explained. https://www.eia.gov/energyexplained/ energy-and-the-environment/where-greenhouse-gases-come-from.php?utm_ source=chatgpt.com. Accessed on Oct. 2024

    work page 2024

  6. [6]

    https://iot-analytics.com/number-connected-iot-devices/

    State of IoT 2024. https://iot-analytics.com/number-connected-iot-devices/. Ac- cessed on Oct. 2024

    work page 2024

  7. [7]

    Energy drain of the object detection processing pipeline for mobile devices: Analysis and implications

    Haoxin Wang, BaekGyu Kim, Jiang Xie, and Zhu Han. Energy drain of the object detection processing pipeline for mobile devices: Analysis and implications. IEEE Transactions on Green Communications and Networking , 5(1):41–60, 2020

    work page 2020

  8. [8]

    LEAF+AIO: Edge-assisted energy-aware object detection for mobile augmented reality

    Haoxin Wang, BaekGyu Kim, Jiang Xie, and Zhu Han. LEAF+AIO: Edge-assisted energy-aware object detection for mobile augmented reality. IEEE Transactions on Mobile Computing, 22(10):5933–5948, 2022

    work page 2022

Show all 17 references
  1. [9]

    NAS-Bench-201: Extending the scope of reproducible neural architecture search

    Xuanyi Dong and Yi Yang. NAS-Bench-201: Extending the scope of reproducible neural architecture search. In Proc. ICLR, 2020

    2020

  2. [10]

    Unveiling energy efficiency in deep learning: Measurement, prediction, and scoring across edge devices

    Xiaolong Tu, Anik Mallik, Dawei Chen, Kyungtae Han, Onur Altintas, Haoxin Wang, and Jiang Xie. Unveiling energy efficiency in deep learning: Measurement, prediction, and scoring across edge devices. In Proc. IEEE/ACM SEC, pages 80–93, 2023

    2023

  3. [11]

    Neural architecture search without training

    Joe Mellor, Jack Turner, Amos Storkey, and Elliot J Crowley. Neural architecture search without training. In Proc. International conference on machine learning , pages 7588–7598. PMLR, 2021

    2021

  4. [12]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    2009

  5. [13]

    https://ai.google.dev/edge/litert/models/measurement

    TFLite Benchmark. https://ai.google.dev/edge/litert/models/measurement. Ac- cessed on Oct. 2024

    2024

  6. [14]

    https://www.msoon.com/high-voltage-power-monitor

    Monsoon. https://www.msoon.com/high-voltage-power-monitor. Accessed on Oct. 2024

    2024

  7. [15]

    Regularized evolution for image classifier architecture search

    Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In Proc. AAAI Conference on GreenAuto: An Automated Platform for Sustainable AI Model Design on Edge Devices HOTMOBILE ’25, February 26–27, 2025, La Quint...

    2025

  8. [16]

    Transformer models: an introduction and catalog

    Xavier Amatriain, Ananth Sankar, Jie Bing, Praveen Kumar Bodigutla, Timothy J Hazen, and Michaeel Kazi. Transformer models: an introduction and catalog. arXiv preprint arXiv:2302.07730, 2023

    2023 arXiv

  9. [17]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.