Semi-supervised and weakly supervised learning train a model on a small set of expert labels plus a large pool of unlabeled or machine-labeled examples. It is now well documented that this pool has a useful size. Past some point, extra examples stop improving the model, and with noisy machine labels they start to hurt it. The field still reports these methods by their final accuracy alone. The electricity and carbon spent on examples that bought nothing, or bought harm, never appears in any results table. We propose to count that spending, using an accounting tool that environmental economists built for a closely related blind spot: GDP counts output and ignores the pollution produced alongside it.
The plan has two parts. First, we audit many training configurations, each combining a method, an amount of unlabeled data, and a hardware and grid setting. Each is scored against a joint "production frontier" in which per-class accuracy counts as good output and measured energy and carbon count as bad output. The audit splits each configuration's inefficiency into two sources. Scale waste comes from ingesting data past the point where it helps. Technique waste comes from a method that burns more compute than necessary at a given data size. Second, we test whether a cheap probe, run before the full training, can predict the efficient amount of data to ingest. That would prevent the waste instead of reporting it afterwards.
Core Claim
The claim has three parts. Most of the carbon spent on large unlabeled pools in current pipelines is scale waste rather than technique waste. The efficient ingestion volume can be predicted from a few small pilot runs. And that volume moves substantially, and predictably, with the carbon intensity of the local electricity grid.
The new element is a decomposition borrowed from efficiency analysis in economics and applied to data scaling. Nobody has measured how much of semi-supervised learning's footprint comes from data that was past saturation. The file also records no attempt to turn saturation findings into an ingestion budget priced in carbon.
What the Two Literatures Supply
The saturation side motivates the work and constrains it. The first two papers establish that a useful pool size exists:
- How Does Pseudo-Labeling Affect the Generalization Error of the Semi-Supervised Gibbs Algorithm? proves, for an idealized learner, that generalization error reaches a floor as the unlabeled-to-labeled ratio grows. Once the information shared between labeled and pseudo-labeled data is used up, more data does nothing.
- From Theory to Decision Rule shows empirically, on three medical benchmarks, that weak labels from vision-language models help until a small gold-only model matches the labeler's zero-shot accuracy. Past that point they cost up to 0.10 AUC. The same paper shows that noise-robust losses soften the damage but do not move this crossover point.
Two other papers constrain the design, because they show saturation is not universal:
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data introduces UnMixMatch, which adds contrastive and rotation-prediction objectives. It keeps gaining (5.61% over a 10× data increase) on unconstrained data.
- Mixture of Expert/Imitator Networks distills a large expert network into small imitator networks. It reports error falling linearly in the log of corpus size from 5,000 to 6 million sentences, with no plateau.
Our method therefore cannot assume a plateau. When gains continue but shrink, the stopping point has to come from a price on carbon, not from saturation alone.
The carbon side supplies measurement and exposes a gap:
- Quantifying the Carbon Emissions of Machine Learning gives the energy-times-grid-intensity model. It reports intensities of about 20 g CO₂eq/kWh in Quebec and over 736 g in Iowa.
- Hugging Carbon shows that floating-point operation counts translate unevenly into emissions. Vision training averages 0.16 tCO₂e per EFLOP against 0.13 for text, and H100 accelerators emit about 56% less than A100s for the same compute.
- Green AI argues for efficiency as a first-class metric via operation counts.
All three account for emissions after the fact. None connects emissions to the data-scaling curve.
The economics side supplies the method:
- Non-radial slack-based efficiency with undesirable outputs, as in the three-stage undesirable SBM study of China's high-tech industry, scores units that produce both good and bad outputs. It handles several of each without forcing proportional changes.
- The split of overall efficiency into pure technical efficiency and scale efficiency, as in Measuring the Efficiency of Regional Integrated Transport in China, is exactly the scale-versus-technique question we ask. That study found transport inefficiency was mostly technical, not scale.
- The Green GDP deduction logic of Assessing the Climate Impact of Green GDP and the imputed-cost accounting of the Nagasaki regional accounts motivate reporting a "carbon-adjusted" gain alongside the raw one.
We apply these tools at the level of individual training runs rather than provinces. The records say they have only been used at the national, provincial, and industry levels.
Research Program
Each training configuration is treated as a production unit. Its inputs are gold-label count, unlabeled or weak-label count, and accelerator-hours. Its desirable outputs are per-class validation AUC or accuracy, kept as separate outputs rather than one macro average. Its undesirable outputs are measured energy in kWh and carbon in grams CO₂eq. Carbon is computed at two reference intensities, 20 and 736 g/kWh, taken from Lacoste et al.
Energy is measured with hardware power counters and a fixed facility overhead. We do not estimate it from operation counts, because Hugging Carbon shows the two diverge.
We then fit a slack-based frontier twice:
- under variable returns to scale, which yields pure technical efficiency;
- under constant returns to scale, which yields overall efficiency.
Their ratio gives scale efficiency:
$$\text{OE}_k = \text{PTE}_k \times \text{SE}_k,$$
where $k$ indexes a configuration, $\text{OE}_k$ is its overall efficiency, $\text{PTE}_k$ is its efficiency relative to the best practice at its own scale, and $\text{SE}_k$ measures how far its scale is from the most productive scale.
Keeping per-class outputs matters. The Xu et al. file notes that weak data may saturate on common classes while still helping rare ones. A multi-output frontier credits a large pool for rare-class gains that a macro metric would hide.
The dual of the frontier program gives, for each configuration, the implied carbon price at which it would count as efficient. We report this as grams CO₂eq per 0.01 AUC for each increment of data. This is the quantity a practitioner actually needs.
The predictive part works as follows. On each task we run the chosen method at three small fractions of the unlabeled pool and fit a saturating learning curve. We combine the fit with the gold-versus-teacher AUC comparison from Xu et al., which flags pools that will actively hurt. We then choose the ingestion volume at which the predicted marginal gain per gram of carbon falls below a stated price. That price is swept across a range, so no single valuation is imposed. The probe's cost is charged against its own savings.
Baselines
The naive reference is ingesting the full unlabeled or weak pool with the method's default schedule. This is standard practice. All carbon savings and accuracy losses are measured against it.
The strongest existing competitor for weak supervision is the AUC-vs-AUC rule of Xu et al.. It reports 76% accuracy at predicting help versus hurt across 129 configurations. We re-implement it on the same PCAM, ISIC 2019, and NIH-CXR splits and the same vision-language labelers. Where it predicts "help," it ingests the full pool; where it predicts "hurt," it ingests none. It receives the same small validation set as our probe. Beating it means one of two outcomes at equal or better test AUC:
- matching its help/hurt calls while cutting energy on "help" cases by a meaningful margin, which we pre-register as at least 30%;
- losing at most 0.01 AUC at a strictly lower carbon cost.
For semi-supervised learning without weak labels, the strongest competitor is validation-plateau stopping over data increments. Here the unlabeled pool is grown in stages until validation accuracy fails to improve for a fixed patience. It gets the same increment grid and validation data.
The key ablation removes the carbon price. The probe then stops only at predicted saturation, which isolates what pricing contributes beyond saturation detection. A second ablation replaces measured energy with operation counts. It tests whether the Hugging Carbon divergence changes any frontier ranking.
Why not just…?
Why not just use validation early stopping? It stops only when accuracy stops rising. On log-linear curves like those in MEIN and UnMixMatch, accuracy never stops rising, so this rule would ingest everything regardless of cost. On weak labels, Xu et al. report that damage appears on gold-labeled test performance. A validation set drawn from the weakly labeled pool would not reveal it. Plateau stopping is still our main semi-supervised baseline. If it matches our probe at high carbon prices, we will report that.
Why not just set a threshold on accuracy gained per kWh? For a single run with one accuracy number, a directional frontier reduces to exactly this ratio. Our in-loop stopping rule is essentially that threshold, and we say so. The frontier adds value in two places. It supports comparisons across many configurations with several per-class outputs. And it splits waste into scale and technique components, which tells a practitioner whether to use less data or a cheaper method. If the decomposition shows negligible technique waste on every benchmark, the frontier machinery is unnecessary. That would be a result we report, not hide.
Why not just count floating-point operations as Green AI proposes? At a fixed site, carbon is energy times a constant. For choosing among methods on one grid, energy suffices and the grid factor only rescales the price. But operation counts are not energy. Hugging Carbon's modality and hardware gaps suggest that methods with heavy input pipelines, such as UnMixMatch's augmentations, may rank differently under each. Our operation-count ablation tests this directly. Across sites, the 40-fold intensity gap moves the priced stopping point, and operation counts cannot represent that.
Why not just apply Xu et al.'s rule, which already exists? It answers whether to use weak data, not how much. It is right 76% of the time, and it says nothing when gains are positive but tiny. That regime is where our scale-waste hypothesis predicts most carbon goes, so the comparison above settles the question directly.
Proposed Experiments
Weak supervision (medical imaging). We use PCAM, ISIC 2019, and NIH-CXR with BiomedCLIP, QuiltNet, and PubMedCLIP as labelers. The reported crossovers are near 100 gold labels on PCAM, 20–50 on ISIC 2019, and 250–500 on NIH-CXR. We vary gold counts around these points, and weak-pool sizes over two orders of magnitude. Methods are plain cross-entropy, Generalized Cross Entropy, and Co-teaching. Co-teaching trains two networks at once, so it is a natural test case for technique waste.
Unconstrained semi-supervised learning (natural images). We use CIFAR-10, CIFAR-100, and STL-10 with ImageNet-100 as the unconstrained pool. We compare UnMixMatch against hard pseudo-labeling at 1×, 3×, and 10× pool sizes. The reference figures are UnMixMatch's reported 5.61% gain over 10× data and its 4.79% average margin over 13 methods. No reported energy or carbon figure for any of these methods was found on Lacuna.
Text classification. We use Elec and IMDB with MEIN imitator networks against Virtual Adversarial Training, which MEIN reports to be about 8× slower. The pool sweeps from 5,000 toward the millions of sentences MEIN reports.
Decisive checks. The first check is whether scale efficiency explains more of the overall inefficiency than pure technical efficiency. If the reverse holds, as Cao et al. found for transport, the proposal's emphasis on data volume is wrong. The second check is whether the probe's chosen volume falls inside the frontier's efficient region more often than both baselines, at both grid intensities, after charging the probe's own cost. The third check is whether per-class outputs on NIH-CXR make large pools efficient that macro-AUC would score as wasteful. Each check can fail on its own, and a failure in the first would redirect the work toward cheaper methods rather than smaller pools.
References
- How Does Pseudo-Labeling Affect the Generalization Error of the Semi-Supervised Gibbs Algorithm?
- From Theory to Decision Rule: Calibrating the Noisy-Label Crossover for VLM Weak Supervision Across Three Medical-Imaging Benchmarks
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data
- Mixture of Expert/Imitator Networks: Scalable Semi-Supervised Learning Framework
- Quantifying the Carbon Emissions of Machine Learning.
- Hugging Carbon: Quantifying the Training Carbon Emissions of AI Models at Scale
- Green AI
- Assessing green innovation efficiency and spatial characteristics of China’s high tech industry based on the three stage undesirable SBM model
- Measuring the Efficiency of Regional Integrated Transport in China: a Data Envelopment Analysis
- Assessing the Climate Impact of Green GDP: A Model for Estimating Global Climate Mitigation Effects
- System for Integrated Environmental and Economic Accounting at the Regional Level in Nagasaki