{"id":"1b65cd0b-d647-4977-b1c8-9aca668860dd","arxiv_id":"2507.07033","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A benchmark of supervised, self-supervised, and semi-supervised contrastive learning on CIFAR-10 and EuroSAT shows labeling energy can dominate training energy, with semi-supervised CCSSL providing a favorable accuracy-energy trade-off.","lead":"This paper measures the energy cost of training supervised, contrastive, and semi-supervised models on CIFAR-10 and EuroSAT, and adds the hidden cost of manually labeling data. It finds that label acquisition can dominate training energy and that semi-supervised contrastive learning offers a better accuracy-per-energy trade-off at the edge.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline energy ratios ('almost 4x less', 'twice') are contradicted by the paper's own Table 1 and Fig. 2; the central quantitative claim lacks internal support.","rationale":"The reader's weakest assumption (the T_label=10 s point estimate) is real and worth flagging, but my concern is more load-bearing: the paper's own numbers contradict its headline claims even under that assumption. A sensitivity analysis of T_label would not resolve the 4x-vs-1.7x discrepancy; the text would need to be revised or the measurements recomputed. This matters because the abstract and introduction sell the paper on specific factors ('up to 4X', 'over twice') that are not reproducible from Table 1 and the Fig. 2 caption. The qualitative insight that label acquisition adds a fixed energy overhead is plausible and consistent with the framework, and I credit the authors for measuring CPU/GPU/RAM with CodeCarbon and for averaging over five seeds. However, no code, data, or per-seed values are provided, so the internal inconsistency cannot be dismissed as a typo without author confirmation. The appropriate disposition is the same conditional one the reader reached: the paper can be corrected by reconciling the reported ratios, adding error bars, and releasing artifacts. My concern reinforces, rather than changes, that verdict.","tokens_in":7622,"tokens_out":7832,"duration_ms":75486,"concrete_test":"Recompute total energy for the Fig. 2 scenario using the paper's own formulas and Table 1. For CIFAR-10 at 50% of the 50,000 training samples (K=25,000), Eq. (3) with P=30 W and T_label=10 s yields 2.08 kWh for full labeling. Add training energies from Table 1: SupCon 1.25 kWh, CCSSL(50) 1.08 kWh, SimCLR 1.26 kWh, with CCSSL labeling 50% of its subset (1.04 kWh). Verify the three ratios reported in Section 4.3: SupCon/SimCLR = 4.0? CCSSL/SimCLR = 2.4? SupCon/CCSSL ≈ 4? If the recomputed ratios are about 2.6, 1.7, and 1.7 respectively, the headline 'almost 4× less energy' is contradicted by the paper's own data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The flagship quantitative claims do not survive recomputation from the paper's own tables. Section 2 states that CCSSL requires 'almost 4× less energy' than SupCon to reach similar accuracies, and the Abstract says 'reducing resource consumption by up to 4X.' Yet Section 4.3 reports that 'CCSSL 50 (50)... consuming 2.4× the energy of SimCLR (for reference, SupCon consumes 4× more SimCLR's energy).' These two ratios imply SupCon/CCSSL = 4/2.4 ≈ 1.67, i.e., CCSSL uses about 60% of SupCon's energy, not 25%. The 'almost 4× less' claim is therefore a factor of roughly 2.4 too strong. The 'label acquisition may consume twice the amount of energy consumed for training' claim is also inconsistent with the authors' own numbers: for CIFAR-10 at 50% data, Eq. (3) with T_label=10 s, P=30 W gives 25,000 samples × 300 J = 2.08 kWh. SupCon training costs 1.25 kWh (Table 1), so labeling/training = 1.67, not 'over twice.' (The cross-entropy baseline does exceed twice, at 3.3, but Section 2 attributes the over-twice statement to SupCon specifically.) Thus, even if one accepts the labeling-time model the reader flagged as weakest, the headline energy ratios do not follow from the reported data. This is an internal inconsistency, not merely a disagreement with prior work, and it directly undermines the paper's central contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an empirical comparison of supervised, self-supervised, and semi-supervised contrastive learning methods (cross-entropy baseline, SimCLR, SupCon, CCSSL) on CIFAR-10 and EuroSAT, measuring training energy with CodeCarbon and adding an estimated labeling energy cost from Eq. (3). The authors claim that label acquisition can consume over twice the training energy for supervised methods and that semi-supervised CCSSL achieves near-supervised accuracy with up to 4x less total energy, motivating energy-aware benchmarking for edge learning.","tokens_in":7938,"tokens_out":6144,"duration_ms":57165,"significance":"If the numbers were correct, the paper would address a genuinely underexplored component of ML energy footprints (labeling) and provide a useful framework for comparing supervision regimes. The experimental setup is transparent (ResNet-18, fixed epochs, CodeCarbon tracking), and the inclusion of both accuracy and energy is valuable. However, the headline quantitative claims are not supported by the paper's own reported data, which weakens the contribution as written. The qualitative insight—that labeling energy can dominate training energy under certain assumptions and that semi-supervised methods may reduce total energy—is plausible and worth publishing, but only after the numerical claims are corrected and made reproducible from the reported tables and equations.","major_comments":[{"comment":"The statement that for SupCon 'label acquisition may consume twice the amount of energy consumed for training' is not supported by the paper's own numbers. Using Eq. (3) with T_label=10 s and P_energy=30 W, labeling 25,000 CIFAR-10 samples (the 50% subset) costs 2.08 kWh, while Table 1 reports SupCon training energy of 1.25 kWh at 50% data and 2.51 kWh at 100% data. The labeling/training ratio is 1.66 in both cases, not 'over twice'. The ratio exceeds two only for the cross-entropy baseline at 50% data (2.08/0.63 = 3.3), which is not the claim made in Section 2.","section":"Abstract & Section 2"},{"comment":"The claim that CCSSL requires 'almost 4× less energy' than SupCon to reach similar accuracies is contradicted by the paper's own Section 4.3, which states that CCSSL(50)(50) consumes 2.4× the energy of SimCLR while SupCon consumes 4× SimCLR's energy. These ratios imply CCSSL/SupCon ≈ 0.6, i.e., about 40% less energy, not 4x less (which would be 25% of SupCon's energy). Recomputing from Table 1 and Eq. (3) gives CCSSL(50)(50) total energy ≈ 1.08 + 1.04 = 2.12 kWh and SupCon total energy ≈ 1.25 + 2.08 = 3.33 kWh, again a ratio of 0.64. The 'up to 4X' claim in the Abstract is therefore unsupported by any reported comparison.","section":"Section 2 & Section 4.3"},{"comment":"The energy ratios quoted in Section 4.3 (2.4× and 4× relative to SimCLR) are not traceable to any reported table or figure values. From Table 1 and the labeling model, CCSSL(50)(50)/SimCLR ≈ 2.12/1.26 ≈ 1.68, and SupCon/SimCLR ≈ 3.33/1.26 ≈ 2.64, not 2.4 and 4. Either Fig. 2 contains additional energy components not described in the text, or the numbers in the text are misreported. The authors must provide the actual measured totals (e.g., in a table) and reconcile the text with those totals, because the central energy-efficiency comparison is based on these ratios.","section":"Section 4.3 & Fig. 2"},{"comment":"The labeling energy model depends critically on the assumed T_label = 10 s per sample and P_energy = 30 W, both taken from a single citation for natural images and applied uniformly to both datasets. The paper's qualitative conclusion that labeling dominates supervised training energy is load-bearing on these parameters. Since the authors aim to guide edge deployments, they should provide a sensitivity analysis over T_label (e.g., 1–30 s) and P_energy (e.g., 5–100 W), and discuss how the trade-off shifts when labeling is faster, automated, or crowdsourced. Without such an analysis, the robustness of the central claim is unclear.","section":"Section 3.1, Eq. (3)"}],"minor_comments":[{"comment":"The title has a spacing issue: 'SELF-SUPERVISED LEARNING A T THE EDGE' should read 'AT THE EDGE'.","section":"Title"},{"comment":"The header 'EuroSA T' contains an unnecessary space; it should be 'EuroSAT'.","section":"Table 1"},{"comment":"The notation 'Penergy(Watts) × KT label' is ambiguous; please write it as 'K · T_label' to clarify the multiplication.","section":"Equation (3)"},{"comment":"The caption says 'labeling cost of ∼ 300 Joules per sample' (consistent with 10 s × 30 W), but the text in Section 4.3 refers to 'the cost of labeling a single sample' without stating the per-sample energy; consider stating it explicitly in the figure caption or main text.","section":"Figure 2 caption"},{"comment":"The notation 'CCSSL(20)(50)' is confusing: it means CCSSL using 50% of the training set with 20% of that subset labeled, but this should be stated more clearly, e.g., 'CCSSL with 20% of the selected subset labeled'.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a candidate for acceptance after a thorough numerical correction. The main issue is that the Abstract and Section 2 make strong quantitative claims ('over twice', 'up to 4X') that are not supported by the paper's own data; the reported ratios in Section 4.3 also disagree with recomputations from Table 1 and Eq. (3). The authors should correct the numbers, add a sensitivity analysis for the labeling parameters, and ideally provide the raw values behind Fig. 2. The qualitative finding—labeling cost matters and semi-supervised learning can reduce total energy—is plausible and worth preserving."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Off the record: the paper's framing is useful and the measurements look honestly done, but the two headline numbers—'over twice' for SupCon labeling and 'almost 4x less' for CCSSL—contradict the paper's own tables. The stress-test note is right, and this is a load-bearing inconsistency, not a cosmetic issue.\n\nWhat's new: nobody else, as far as the citations show, has added labeling energy to the evaluation of SimCLR, SupCon, and CCSSL on CIFAR-10 and EuroSAT across three data fractions. The training energy is measured with CodeCarbon over five seeds, and the labeling model is a simple formula (power × time × samples) that is transparent and externally grounded. That gives practitioners a concrete way to reason about total energy at the edge.\n\nThe soft spots: from Table 1 and Eq. (3), for CIFAR-10 at 50% data, SupCon training is 1.25 kWh and labeling 25k samples at 10 s/sample and 30 W adds ~2.08 kWh, a 1.67x ratio, not the claimed 'over twice.' For CCSSL vs SupCon, Section 4.3 says CCSSL 50 uses 2.4x SimCLR and SupCon uses 4x SimCLR, which puts CCSSL at 60% of SupCon's total, not 25%. The '4x less' claim is off by a factor of ~2.4. The 'SupCon consumes 4x more than SimCLR' reference in Fig. 2 also doesn't match the labeling model, giving ~2.6x instead of 4x.\n\nThe qualitative conclusion—that labeling costs can flip the energy ranking in favor of semi-supervised learning—is plausible and likely correct. But the paper as written cannot support the magnitude it claims. Missing code and accuracy error bars are smaller issues that could be fixed with a release.\n\nWho it's for: practitioners choosing between supervised, self-supervised, and semi-supervised on edge hardware. They'd get the framework, but they should not quote the specific ratios until the authors correct them.\n\nRecommendation: send it to peer review with a major-revision request. The question is valid, the data appears measured, and the errors are fixable. Do not desk-reject; do not accept as is. I would not cite the numbers until corrected.","headline":"Useful framing with honest measurements, but the headline energy ratios don't survive a recomputation from the paper's own tables.","tokens_in":8490,"tokens_out":4474,"would_cite":false,"duration_ms":43423,"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":"Including label acquisition in the energy budget, semi-supervised contrastive learning matches supervised accuracy on roughly a quarter of the energy.","keywords":["self-supervised learning","contrastive learning","semi-supervised learning","energy efficiency","label acquisition cost","edge computing","CIFAR-10","sustainable AI"],"falsifier":"A controlled annotation-time study for CIFAR-10 or EuroSAT reporting mean seconds per label at measured desktop power would settle the core claim: if the true mean is below roughly 2-3 seconds per image, the label bill no longer dominates training energy and the paper's reported 2x and 4x energy ratios collapse.","tokens_in":7403,"feed_emoji":"⚡","tokens_out":7810,"duration_ms":79875,"temperature":0.7,"pith_summary":"The paper asks how much energy a machine-learning model really costs when the energy spent by humans labeling data is included, not just the electricity used in training. It argues that on resource-limited edge hardware, supervised contrastive learning's accuracy advantage disappears under this accounting because label acquisition can consume more than twice the energy of training. The central quantitative claim is that a semi-supervised contrastive method, CCSSL, reaches accuracy nearly equal to the supervised contrastive model while consuming roughly a quarter of the total energy, with the gap widest in low-data regimes. This matters because edge devices operate under tight energy budgets, so choosing a learning method by accuracy alone misprices the real cost of deployment.","feed_headline":"Semi-supervised learning hits supervised accuracy on 4x less energy","feed_subtitle":"Including label acquisition, supervised models can spend more energy annotating than training; CCSSL avoids the bill.","key_machinery":"The load-bearing object is a two-term energy ledger, $E_{\\text{total}} = E_{\\text{train}} + E_{\\text{labeling}}$, with labeling cost computed from Eq. (3), $E_{\\text{labeling}} = P_{\\text{energy}} \\times K T_{\\text{label}} / 3600$ kWh. The argument hinges on inserting the per-sample annotation time $T_{\\text{label}}=10$ s at $P_{\\text{energy}}=30$ W and comparing four supervision regimes (supervised cross-entropy, SimCLR, SupCon, CCSSL) under the same training procedures. This ledger converts accuracy differences into energy differences, and it is what makes the quantitative claim about CCSSL's roughly 4x advantage legible.","core_discovery":"On its own terms, the paper establishes that a fair energy comparison of learning methods at the edge must add label-acquisition energy to training energy, and that once this is done the ranking changes. Using a ResNet-18 backbone on CIFAR-10 (with EuroSAT as a second dataset), the authors measure GPU, CPU, and RAM energy during 1,000-epoch training, then add a per-sample labeling cost of 10 seconds at 30 W. In this accounting, fully supervised approaches (a cross-entropy baseline and supervised contrastive learning, SupCon) reach the highest accuracies but carry a labeling bill that can outweigh their training energy, while self-supervised SimCLR trains cheaply but visibly loses accuracy. The central result is that the semi-supervised contrastive method CCSSL, trained with a labeled subset and unlabeled data, reaches accuracy within a point or less of SupCon while consuming roughly a quarter of SupCon's total energy; the paper reports the advantage persists and sharpens in low-data regimes.","pith_inferences":["Editorial inference: the 10 s/30 W labeling estimate is for natural images; if annotation time varies by domain, the crossover point between supervised and semi-supervised energy moves accordingly.","Editorial inference: the paper's ledger excludes the energy of transmitting data to a human annotator; for remote or federated edge settings, communication energy would further favor locally trained SSL.","Editorial inference: a testable extension is to run the same comparison with active learning, where labels are requested only for the most informative samples; this could lower the label bill of CCSSL below what the paper reports.","Editorial inference: if annotation becomes automated or near-free, supervised methods regain competitiveness, so the paper's conclusion is conditioned on the scarcity and cost of human labeling that motivates SSL in the first place."],"forward_implications":["Energy benchmarking of learning systems should include labeling as a first-class cost; otherwise supervised methods appear cheaper than they are in edge deployments.","Semi-supervised contrastive learning becomes the default energy-accuracy tradeoff point for resource-constrained devices, since it matches supervised accuracy at a fraction of the total energy.","The advantage of partial labeling grows as data becomes scarcer: in the 20% and 50% data regimes, the accuracy gap between CCSSL and SupCon stays small while the energy gap widens.","In dynamic edge environments where models must be retrained on freshly observed data, the per-sample labeling bill recurs, making label-efficient SSL increasingly attractive relative to one-time supervised training."],"supporting_citations":[{"why":"Provides the 10 s per-image annotation time used in Eq. (3) to compute labeling energy.","marker":"[13]"},{"why":"Defines the CCSSL class-aware semi-supervised contrastive method that delivers the headline energy-accuracy tradeoff.","marker":"[6]"},{"why":"Defines SupCon, the supervised contrastive baseline whose accuracy and labeling burden anchor the comparison.","marker":"[9]"},{"why":"Defines SimCLR, the self-supervised baseline representing the zero-label endpoint of the energy spectrum.","marker":"[2]"},{"why":"Supplies the power-tracking tool used to estimate CPU/GPU/RAM energy during training.","marker":"[14]"},{"why":"Provides the RAM power-per-GB constant used in the memory energy term.","marker":"[16]"},{"why":"Defines the kNN evaluation protocol used to report representation accuracy.","marker":"[17]"}],"fun_headline_variants":["Semi-supervised beats supervised when labeling energy counts","Labels cost 4x more than training — semi-supervised wins","CCSSL: near-supervised accuracy with 4x less total energy","Semi-supervised: supervised accuracy at a quarter the energy","Counting labeling energy flips the ranking to semi-supervised"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole quantitative comparison rests on assuming every labeled image costs 10 seconds of human attention at 30 W of desktop power, regardless of dataset, labeler, or interface.","fun_headline_variants_meta":{"raw":{"variants":["Semi-supervised beats supervised when labeling energy counts","Labels cost 4x more than training — semi-supervised wins","CCSSL: near-supervised accuracy with 4x less total energy","Semi-supervised: supervised accuracy at a quarter the energy","Counting labeling energy flips the ranking to semi-supervised"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000348,"raw_usage":{"total_tokens":1893,"prompt_tokens":923,"completion_tokens":970,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":884}},"tokens_in":539,"tokens_out":970,"duration_ms":10304,"temperature":1.0,"reasoning_tokens":884,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:48:29.529583+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled annotation-time study for CIFAR-10 or EuroSAT reporting mean seconds per label at measured desktop power would settle the core claim: if the true mean is below roughly 2-3 seconds per image, the label bill no longer dominates training energy and the paper's reported 2x and 4x energy ratios collapse.","supporting_citations":[{"cited_title":"Convergence of edge computing and deep learning: A comprehensive survey,","cited_arxiv_id":null,"evidence_quote":"Defines the CCSSL class-aware semi-supervised contrastive method that delivers the headline energy-accuracy tradeoff."},{"cited_title":"Exploring simple siamese representation learning,","cited_arxiv_id":null,"evidence_quote":"Defines SupCon, the supervised contrastive baseline whose accuracy and labeling burden anchor the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines SimCLR, the self-supervised baseline representing the zero-label endpoint of the energy spectrum."},{"cited_title":"Supervised Contrastive Learn- ing,","cited_arxiv_id":null,"evidence_quote":"Supplies the power-tracking tool used to estimate CPU/GPU/RAM energy during training."},{"cited_title":"Dissecting supervised contrastive learn- ing,","cited_arxiv_id":null,"evidence_quote":"Provides the RAM power-per-GB constant used in the memory energy term."}],"review_version":1}