{"id":"d07ec5b7-9b6e-4aed-a108-dd7f3b84b1bb","arxiv_id":"2608.09998","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"On a CPU-based CIFAR-10 benchmark, training accounts for nearly all carbon emissions, and more complex models do not yield proportionally higher accuracy.","lead":"This paper reviews Green AI and compares six deep learning models by their carbon emissions on a CPU, finding that training dominates the footprint and that more complex models are not proportionally more accurate. It is a useful benchmark for practitioners who want to weigh accuracy against environmental cost.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The complexity-vs-accuracy claim rests on single runs with no reported training protocol; missing hyperparameters, not just run-to-run noise, could invert the EAP ranking.","rationale":"The reader identified single-run measurements as the weakest assumption; I agree that this is the central fragile point, but I would sharpen it: the absence of any reported training protocol means the experiment does not currently separate architectural behavior from training-effort artifacts. The training-dominates-emissions conclusion is robust because the per-phase gaps in Fig. 5 are orders of magnitude and would survive reasonable reruns. The 'increased complexity does not systematically improve accuracy' conclusion, however, depends on accuracies from one run per model with unstated epochs and optimization settings; a few accuracy points separate several models, and EAP rankings are directly driven by those points. The paper's own Section 7 acknowledges the single-run limitation and promises repeated experiments, which supports a conditional rather than a reject verdict. I also note the multi-label/multi-class inconsistency as a scope error that should be corrected, but it is secondary to the empirical support for the central claim. The concrete test above would settle whether the claimed accuracy-emissions trade-off is a property of the architectures or an artifact of the unreported training setup.","tokens_in":17097,"tokens_out":4433,"duration_ms":53374,"concrete_test":"Retrieve the released Figshare code and train all six models under one fully specified protocol (fixed epoch count, batch size, optimizer, learning rate, and random seed), repeating each run at least 10 times on the same CPU machine while recording CodeCarbon emissions and test accuracy. If the 95% confidence intervals for VGG16/ResNet/EfficientNet accuracies overlap, or if the EAP ordering changes across repetitions, the complexity-efficiency claim is not supported. Separately, inspect the code to confirm whether the target labels are multi-hot (multi-label) or one-hot single-class; if one-hot, the 'multi-label' framing should be corrected to multi-class.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 4.2 reports one execution per model, and Section 7 concedes measurements came from single runs, but the paper never reports epochs, batch size, optimizer, learning rate, early stopping, or validation-based model selection. Training emissions scale linearly with the number of steps, and test accuracy depends strongly on the optimization schedule; therefore the second central claim—that added architectural complexity does not buy proportional accuracy—is not distinguished from the possibility that the larger models (e.g., VGG19 at 78%) were undertrained or stopped at an unlucky seed. The 'training phase dominates' finding is safe because Fig. 5 shows order-of-magnitude gaps, but the EAP ranking in Table 7 and the associated model-selection advice are load-bearing on the unstable single-run accuracy measurements. A separate label inconsistency compounds this: 'multi-label' is claimed in the abstract and introduction, yet CIFAR-10 with one-hot categorical encoding is single-label multi-class, so the stated scope is also not exactly what was run.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper combines a PRISMA-based systematic literature review of Green AI / Green DL and a comparative analysis of seven carbon-footprint measurement tools with an empirical study that measures the carbon emissions of six CNN architectures (Simple CNN, U-Net, ResNet, VGG16, VGG19, EfficientNet) on CIFAR-10 using CodeCarbon on a CPU. It reports total and per-phase emissions (preprocessing, training, testing), defines an Emissions per Accuracy Point (EAP) metric, and concludes that the training phase dominates emissions and that increased architectural complexity does not proportionally improve accuracy, motivating carbon-aware model selection.","tokens_in":17231,"tokens_out":4144,"duration_ms":44829,"significance":"If the quantitative claims hold, the paper makes a useful, reproducible contribution: the experimental setup, CodeCarbon measurements, per-phase breakdown, and EAP rankings provide a concrete benchmark for CPU-based green DL evaluation, and the literature review with tool comparison is a helpful synthesis. The central qualitative finding that training dominates per-run emissions in this setup is strongly supported by orders-of-magnitude differences, and the paper is unusually transparent in its code/data availability and in its stated limitations. However, the accuracy-complexity claim and the EAP ranking are more fragile because they rest on single runs with unreported training hyperparameters, and several terminological and unit errors weaken the presentation.","major_comments":[{"comment":"The entire quantitative comparison rests on a single execution per model, and the training protocol is not reported: epochs, batch size, optimizer, learning rate, early stopping, and validation-based model selection are all absent. Because training emissions scale with the number of optimization steps and accuracy depends strongly on the optimization schedule, the central claim that 'increased architectural complexity does not systematically translate into proportional accuracy gains' is not distinguished from the possibility that the larger models (e.g., VGG19 at 78%) were undertrained or affected by seed variability. Please report the full hyperparameters, repeat the runs, and provide variance estimates or at least a sensitivity discussion; without this, the EAP ranking in Table 7 and the associated model-selection advice are not load-bearing.","section":"Section 4.2, Section 7 (Internal Validity), Table 7"},{"comment":"EAP is defined as C_i / A_i, but A_i is reported as a fraction (e.g., 0.84 for U-Net), so the values in Table 7 (e.g., 2.10) are grams of CO2e per unit accuracy, not 'grams per accuracy point.' If 'accuracy point' means a percentage point, the denominator should be 84, giving 0.02095 g/point. The normalized EAP ranking is invariant to this uniform scaling, but the units and the text's interpretation of EAP as 'grams of CO2e emitted per accuracy point' are incorrect and should be corrected.","section":"Section 5, Eq. (1), Table 7"},{"comment":"The empirical task is repeatedly described as 'multi-label classification,' but CIFAR-10 with one-hot categorical encoding is a single-label, multi-class classification problem. The paper itself states that each image has a single class and uses 'to_categorical' for one-hot encoding. The terminology should be corrected throughout, or the study should actually include a multi-label setting; as written, the stated experimental scope does not match the data used.","section":"Abstract, Section 4.3.2, Section 4.1"},{"comment":"The conclusion that 'the training phase is the primary source of emissions' is presented as a lifecycle finding, but the measurements cover only one training run and one test pass with no deployment or inference-lifetime analysis. The paper's own literature review (Section 2.3.1) cites Schwartz et al. and Wu et al., who report that inference can dominate in deployed systems (Wu et al. report ~65% inference versus 35% training). The claim should be explicitly restricted to 'per-run emissions in this CPU-based experimental setting' in the abstract and conclusion, not stated as a general lifecycle result.","section":"Section 5, Section 8, Abstract"}],"minor_comments":[{"comment":"The PRISMA flow diagram numbers are internally consistent, but the text could state the final numbers of included studies from database search and snowballing separately; currently they must be inferred from the diagram.","section":"Section 2.1, Fig. 2"},{"comment":"The comparison criteria and table are useful, but the abbreviations 'C1'–'C6' are reintroduced in the text without being explicitly tied to the table rows for readers skimming the narrative.","section":"Section 3.2, Table 5"},{"comment":"EfficientNet is cited to a rice-leaf-disease paper rather than the original EfficientNet paper (Tan and Le, 2019), and VGG16/VGG19 are cited to secondary sources rather than the original Simonyan and Zisserman work; please cite the primary architecture papers.","section":"References [59], [60], [61]"},{"comment":"The data preparation section reports that blurry and dark images were removed but gives no counts or thresholds; please state how many images were removed and how 'blurry' and 'dark' were defined.","section":"Section 4.3.2"},{"comment":"The hardware description should specify whether CPU frequency scaling, turbo boost, and background processes were controlled beyond 'unnecessary background processes were minimized,' since these can materially affect energy measurements on a local machine.","section":"Section 4.2"},{"comment":"There are several typographical artifacts in the tables and captions, such as 'T able 1,' 'T racker,' and 'F unctionality'; these should be cleaned up in the final version.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is the accepted author version of a paper already published in Applied Intelligence (2026). The report is written for a journal-review context; if this is a re-review of the published version, the authors should be asked to clarify whether revision is feasible for the published record or whether an erratum/corrigendum would be more appropriate for the unit and terminology issues."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is one of those papers where the whole is slightly less than the sum of its parts. The systematic review is competent and the tool comparison is handy, but the only genuinely new piece is a small CPU-based benchmark of six standard CNNs on CIFAR-10 using CodeCarbon. The authors deserve credit for putting the data on Figshare, using a widely accepted measurement tool, and for stating plainly in Section 7 that the measurements come from single runs. The 'training dominates' result is on solid ground within this setup: the gaps are orders of magnitude and no one will be surprised.\n\nThe soft spot is the second headline claim, that architectural complexity does not buy proportional accuracy. That conclusion rests on one execution per model, with no reported epochs, batch size, optimizer, learning rate, or early stopping. Training emissions scale with the number of steps, and test accuracy depends strongly on the optimization schedule. VGG19 at 78% could be undertrained or unlucky. So the EAP ranking in Table 7 is not yet a reliable model-selection guide. It is also mislabeled: EAP = Ci/Ai with Ai as a fraction gives grams per unit accuracy, not grams per accuracy point. And calling CIFAR-10 a 'multi-label' task is simply wrong; it is single-label, multi-class. Fixing these is easy, but they do cap the quantitative value.\n\nThe reader's report understates the novelty floor: 'training dominates' and 'bigger is not always better' are already in the cited literature, so the benchmark is mostly confirmatory. That is fine, but it means the paper's value is as a reproducible data point for CPU-scale Green AI, not as a conceptual advance.\n\nIf this came to me as a new submission, I'd send it to referees. It is clearly written, honest about limitations, and the empirical core is transparent enough that a good referee can request the missing training configuration and repeated runs. For the version in front of us, I'd treat the qualitative conclusion as safe and the quantitative ranking as provisional.\n\nRecommendation: engage with it — it is a useful baseline for small-scale CPU experiments, and worth citing once the reporting issues are acknowledged.","headline":"A competent but narrow Green AI survey with a small CPU benchmark; the training-dominates finding is solid in context, but the complexity-vs-accuracy claim is undercut by single runs and missing training details.","tokens_in":17791,"tokens_out":2433,"would_cite":true,"duration_ms":26363,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"In CPU benchmarks, training dwarfs other phases in AI models' carbon cost","keywords":["Green Artificial Intelligence","Green Deep Learning","Carbon footprint assessment tools","Sustainable development","Model optimization","Emissions per Accuracy Point","CodeCarbon","Deep learning energy efficiency"],"falsifier":"Run each of the six models multiple times (for example, ten runs) under the same CPU conditions and compare run-to-run variation in emissions and accuracy. If the spread within a model matches or exceeds the differences between models, such as EfficientNet's 4.67 g versus Simple CNN's 6.02 g, the EAP ranking is not stable. A second decisive test is repeating the benchmark on a GPU or cloud instance; if training no longer dominates the lifecycle, the paper's central claim is specific to CPU settings.","tokens_in":16859,"feed_emoji":"🌱","tokens_out":4380,"duration_ms":43321,"temperature":0.7,"pith_summary":"This paper argues that the carbon footprint of deep learning models should be a first-class selection criterion, alongside accuracy. It backs the argument with a controlled CPU-based benchmark of six convolutional architectures on the CIFAR-10 image set, measured with CodeCarbon at a fixed regional carbon intensity. The results show that the training phase produces the overwhelming share of emissions, and that heavier architectures do not buy proportionally better accuracy. The paper introduces Emissions per Accuracy Point (EAP), grams of CO2e per accuracy point, as a way to compare models on sustainability, and ranks the six models by it. If the finding holds, model choice and optimization effort should be guided by carbon efficiency, not accuracy alone.","feed_headline":"Training dominates AI models' carbon cost in CPU tests","feed_subtitle":"Six convolutional networks measured by emissions per accuracy point; lighter models match heavier ones.","key_machinery":"The carrying mechanism is a pair of ratios. CodeCarbon estimates emissions with $C = E \\times CI$, where $E$ is energy consumed and $CI$ is regional carbon intensity (39.49 gCO2e/kWh for Ontario, Canada); the paper uses this to attribute emissions per phase. The comparative metric is EAP, $EAP_i = C_i / A_i$, with a min-max normalized version $EAP^{Norm}_i = (EAP_i - \\min(EAP)) / (\\max(EAP) - \\min(EAP))$. EAP does the argumentative work: it turns raw emissions into a performance-adjusted quantity, so a high-accuracy model with high emissions (ResNet) can be ranked below a lighter model (U-Net) once carbon cost per accuracy point is considered.","core_discovery":"On the authors' CPU-based experimental setup, measured carbon emissions ranged from 1.76 gCO2e for U-Net to 15.64 gCO2e for VGG19 across the full pipeline, with training contributing orders of magnitude more than preprocessing and testing. Accuracy did not track emissions: ResNet reached the highest accuracy (90%) at 12.34 g, while VGG19 combined the highest emissions with the lowest accuracy (78%), and U-Net obtained 84% accuracy for less than one-seventh of VGG19's emissions. The paper's central claim is that training is the primary source of emissions in this setting and that increased architectural complexity does not systematically translate into proportional accuracy gains. It formalizes the accuracy–emissions trade-off through EAP, defined as total emissions divided by accuracy, with lower values indicating better carbon efficiency; normalized EAP scores place U-Net at 0.00 and VGG19 at 1.00.","pith_inferences":["If repeated runs confirm the single-run measurements, the EAP ranking is robust enough to inform model selection in resource-constrained settings; a natural next step is computing confidence intervals from several runs per model.","The pattern may shift at deployment scale: literature cited by the paper reports inference as roughly 65% of lifecycle emissions in production, so a CPU-training benchmark likely underestimates the relative importance of inference for always-on services.","EAP generalizes beyond the six architectures tested; applying it to GPU or cloud benchmarks would test whether training-dominance is a property of CPUs or of deep learning generally.","The measurement pipeline of CodeCarbon plus EAP could be extended to pruning, quantization, and mixed-precision training, with the testable prediction that these techniques reduce EAP mainly by cutting training-phase energy."],"forward_implications":["Optimization effort in CPU-bound training should concentrate on the training phase, since preprocessing and testing contribute emissions on the order of $10^{-3}$ to $10^{-1}$ g.","Model selection should report EAP alongside accuracy; a lighter architecture can deliver competitive accuracy at a fraction of the carbon cost.","Architectural complexity is not a reliable proxy for predictive gain: the highest-emission model in the benchmark had the lowest accuracy.","Carbon-aware benchmarking, where emissions are reported with performance metrics, becomes a practical complement to accuracy-only reporting.","The quantitative rankings are directly applicable to small-to-moderate scale CPU training; the authors caution that GPU, cloud, and distributed settings may differ qualitatively."],"supporting_citations":[{"why":"Supplies the measurement tool and the emission formula $C = E \\times CI$ used in the benchmark.","marker":"[53]"},{"why":"Provides the CIFAR-10 dataset on which all six models were trained and tested.","marker":"[63]"},{"why":"Defines the U-Net architecture that achieved the lowest EAP in the benchmark.","marker":"[57]"},{"why":"Defines ResNet, the highest-accuracy model in the comparison.","marker":"[58]"},{"why":"Provides the EfficientNet architecture and its compound scaling rationale.","marker":"[59]"},{"why":"Defines the VGG16 architecture included in the comparison.","marker":"[60]"},{"why":"Defines the VGG19 architecture, the highest-emission and lowest-accuracy model.","marker":"[61]"},{"why":"Supplies the prior definition and use of emissions per unit of performance, the basis for EAP.","marker":"[64]"},{"why":"Establishes energy and policy considerations for deep learning that motivate reporting carbon cost.","marker":"[3]"},{"why":"Provides the reporting framework and efficiency metric that EAP extends.","marker":"[4]"}],"fun_headline_variants":["Training dominates AI carbon cost; simpler models match accuracy","U-Net: 84% accuracy at 1.76g CO2e vs VGG19's 15.64g","AI's carbon footprint: training phase is the top emitter","Study: more complex AI models don't guarantee higher accuracy","Lighter AI models rival heavy ones in accuracy with far less carbon"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark's conclusions rest on the assumption that one execution of each model, on one machine and one dataset, gives a representative picture of both accuracy and emissions; the paper itself notes that measurements came from single experimental runs.","fun_headline_variants_meta":{"raw":{"variants":["Training dominates AI carbon cost; simpler models match accuracy","U-Net: 84% accuracy at 1.76g CO2e vs VGG19's 15.64g","AI's carbon footprint: training phase is the top emitter","Study: more complex AI models don't guarantee higher accuracy","Lighter AI models rival heavy ones in accuracy with far less carbon"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001244,"raw_usage":{"total_tokens":5111,"prompt_tokens":959,"completion_tokens":4152,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":4068}},"tokens_in":575,"tokens_out":4152,"duration_ms":30332,"temperature":1.0,"reasoning_tokens":4068,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:49:06.959157+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run each of the six models multiple times (for example, ten runs) under the same CPU conditions and compare run-to-run variation in emissions and accuracy. If the spread within a model matches or exceeds the differences between models, such as EfficientNet's 4.67 g versus Simple CNN's 6.02 g, the EAP ranking is not stable. A second decisive test is repeating the benchmark on a GPU or cloud instance; if training no longer dominates the lifecycle, the paper's central claim is specific to CPU settings.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition, pp","cited_arxiv_id":null,"evidence_quote":"Defines ResNet, the highest-accuracy model in the comparison."},{"cited_title":"Accessed: 2024-06-26 (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the measurement tool and the emission formula $C = E \\times CI$ used in the benchmark."},{"cited_title":"University of Toronto18268744(2009)","cited_arxiv_id":null,"evidence_quote":"Provides the CIFAR-10 dataset on which all six models were trained and tested."},{"cited_title":"In: 2021 Interna- tional Conference on Disruptive Technologies for Multi-disciplinary Research and Applications (CENTCON), vol","cited_arxiv_id":null,"evidence_quote":"Defines the VGG19 architecture, the highest-emission and lowest-accuracy model."}],"review_version":1}