Skip to content

Position: Stop Preaching and Start Practising Data Frugality for Responsible Development of AI

Conference: ICML 2026
arXiv: 2602.19789
Code: https://github.com/saintslab/data-frugality
Area: AI Safety / Sustainable AI / Data-efficient
Keywords: data frugality, coreset, carbon emissions, responsible AI, subset selection

TL;DR

This position paper points out a "value-action gap" in the ML community regarding "data frugality." While researchers acknowledge that coresets save energy, few actually report energy consumption or carbon emissions. Using ImageNet-1K as a case study, the authors calculate a conservative lower bound of ~5.82 GWh / 2589 tCO2e for downstream training and storage, calling for data frugality to transition from a slogan to a measurable, actionable, and rewardable engineering practice.

Background & Motivation

Background: Current mainstream AI training follows the scaling laws of Kaplan and Hoffmann, equating "larger datasets" with "better models," while benchmarks and leaderboards reward experiments on massive corpora. Simultaneously, a parallel movement advocates for "scaling down" through data subset methods like coreset selection and dataset condensation.

Limitations of Prior Work: The authors conducted a systematic survey of 10 representative coreset papers. They found that 8/10 cited "computational efficiency" as a motivation and 3/10 mentioned "energy efficiency," but only 1/10 actually reported energy savings, while 6/10 reported time savings. Publicity for data frugality far outpaces practice.

Key Challenge: Carbon emissions do not scale linearly with compute or data size—a 25% data pruning achieves 24%–40% time savings and 24%–33% energy savings across different architectures. Without measuring actual energy and carbon, "energy saving" claims remain rhetorical and useless for responsible AI decision-making. Furthermore, carbon estimates depend heavily on grid carbon intensity, which varies by over 60x globally (e.g., Turkmenistan vs. Lesotho).

Goal: (i) Quantify the downstream training and storage carbon costs for "national-level" datasets like ImageNet-1K to provide a citable lower bound; (ii) Empirically demonstrate that coresets save power without compromising accuracy; (iii) Translate these needs into concrete recommendations across three layers: People, Platforms, and Policies.

Key Insight: By combining the SAINTS Lab's Carbontracker measurements, OpenReview metadata, and Hugging Face download statistics, the authors construct a "conservative lower bound." They aim to show that even under the most minimal assumptions, the "opportunity cost" of ignoring data frugality is staggering.

Core Idea: Transition data frugality from an activist issue of a "value-action gap" into a reporting and peer-review norm aligned with the ML community—if a method claims energy efficiency, it must measure energy.

Method

As a position paper, the "method" consists of three supporting arguments: carbon cost accounting, empirical coreset evidence, and an actionable framework.

Overall Architecture

The paper is organized through a lifecycle perspective: the data lifecycle (collection, cleaning, storage, distribution) and the model lifecycle (training, selection, deployment) run in parallel, with coreset/subset selection impacting multiple stages. Based on this:

  1. Quantifying the Baseline: Using OpenReview metadata (ICLR 2017–2022) to estimate the number of papers training ImageNet-1K from scratch, then extrapolating to the full ML literature using dimensions.ai keyword indexing and projecting to 2023–2025.
  2. Empirical Gains: Measuring energy and time for full vs. 25% pruned training on A100 GPUs using ResNet-34 / ResNet-50 / Swin-T with Carbontracker. Referencing SOTA subset selection curves (Dyn-Unc, InfoMax) to show no accuracy loss.
  3. Fairness Side Effects: Comparing random, reweighted, and balanced sampling on Coloured-MNIST (99% majority bias) to prove coresets can simultaneously assist in debiasing.
  4. Action Framework: Consolidating evidence into appeals for People, Platforms, and Policies.

Key Designs

  1. Reproducible Lower Bound Estimation of ImageNet-1K Downstream Carbon Costs:

    • Function: Converts "environmental cost of datasets" from an abstract slogan into a citable figure—5.82 GWh energy and 2589 tCO2e (global average).
    • Mechanism: Training cost = estimated training count \(N \approx 46{,}179 \pm 1{,}154\) × ResNet-50 single epoch energy \(\approx 0.394\) kWh × 300 epochs, resulting in \(5.46 \pm 0.14\) GWh. Storage cost = \(N \times 130\) GB × 60 kWh/TB/yr \(\approx 360 \pm 9\) MWh. \(N\) is derived from the proportion of "train from scratch" papers in OpenReview, extrapolated via indexing.
    • Design Motivation: The authors deliberately used public metadata to avoid dependence on proprietary vendor data. They acknowledge this as a "lower bound"—Hugging Face Hub counts 214 ImageNet-1K derivatives with over 2.5 million downloads (55x the paper count), suggesting the real cost is much higher.

  2. Empirical Dual-Axis Core-set Gains: Accuracy Curves × Energy Tables:

    • Function: Answers "how much is saved vs. how much is lost" using accuracy-pruning ratio curves and energy tables.
    • Mechanism: Accuracy is demonstrated using SOTA curves from Dyn-Unc (Swin-T) and InfoMax (ResNet-34), showing 25%–35% of ImageNet-1K can be pruned without Top-1 drop. Energy is measured for ResNet-34/50 and Swin-T over 10 epochs (3 runs on A100), recording both CPU and GPU usage via Carbontracker. Results show 25% pruning yields savings of 32% time / 29% energy for ResNet-34, 40% / 33% for ResNet-50, and 24% / 24% for Swin-T.
    • Design Motivation: It highlights that 25% data \(\neq\) 25% energy, debunking the common assumption of data size as a perfect proxy for energy. It also honestly includes the one-time cost of coreset construction, noting it amortizes within 3–4 training runs.

  3. The Framework: Actionable Proposals for People, Platforms, and Policies:

    • Function: Translates ethical appeals into specific actions for conference calls, leaderboard rules, and funding terms.
    • Mechanism: (People) Use "Data-Pareto" plots (Accuracy vs. Data Volume) instead of single accuracy points; measure what you motivate. (Platforms) Mandate compute reporting (e.g., CVPR 2026) and introduce data-efficient challenges like BabyLM. (Policies) Standardize reporting, promote shared data infrastructure (e.g., Berzelius) to reduce local redundancy, and propose "data sunset laws" requiring approval for large-scale data use.
    • Design Motivation: The authors argue that the value-action gap persists because "thinking is easier than doing"; frugality must move from "personal virtue" to an "institutional default."

Loss & Training

This is a position paper; no main model training loss is proposed. The measurement stack uses Carbontracker and CodeCarbon for training, and a storage intensity coefficient of 60 kWh/TB/yr (Selvan, 2025).

Key Experimental Results

Main Results: ImageNet-1K Downstream Environmental Cost Estimates

Dimension Estimated Value Equivalent Carbon (445 g/kWh) Equiv. Annual Per Capita Carbon Footprint
Training (46.2k × 300 ep × 0.394 kWh) 5.46 ± 0.14 GWh 2429 ± 61 tCO2e ~514 ± 13 people
Storage (46.2k copies × 130 GB × 60 kWh/TB/yr) 360 ± 9 MWh 160 ± 4 tCO2e ~34 ± 1 people
Total 5.82 ± 0.15 GWh 2589 ± 65 tCO2e ~548 people
Same Energy @ Turkmenistan (1310 g/kWh) — 7624 ± 191 tCO2e —
Same Energy @ Lesotho (21 g/kWh) — 122 ± 3 tCO2e —

Ablation Study: Impact of 25% Data Pruning on Training Time/Energy (A100, Single GPU)

Model Params min/epoch (full → 25%) Time Savings kWh/epoch (full → 25%) Energy Savings
ResNet-34 21.8M 35.2 → 23.8 32% 0.2798 → 0.1989 29%
ResNet-50 25.6M 40.7 → 24.3 40% 0.3940 → 0.2645 33%
Swin-T 28.3M 58.7 → 44.6 24% 0.7002 → 0.5300 24%

Key Findings

  • "25% Data \(\neq\) 25% Energy": Energy reduction ratios vary from 24% to 33% across architectures, proving that data size is not a perfect proxy for energy.
  • Dyn-Unc and InfoMax can prune 25% / 35% of ImageNet-1K respectively without accuracy loss; this represents a potential saving of 621–854 tCO2e.
  • Coreset construction costs amortize after 3–4 training runs, even for "expensive" methods like Dyn-Unc.
  • Coloured-MNIST experiments show that balanced/reweighted sampling significantly improves conflicted accuracy under 99% bias; data frugality can inherently assist in debiasing.
  • Iceberg Effect: Downloads of ImageNet derivatives on Hugging Face are 55x higher than paper-based counts, indicating that the true environmental cost is far higher than the lower bound provided.

Highlights & Insights

  • "Quantify lower bounds before calling for action": Unlike many sustainability pieces, the authors use public metadata to establish a solid lower bound, minimizing room for counter-argument.
  • The rule "measure what you motivate" is a concise action that can be directly implemented in conference review forms.
  • Data-Pareto reporting paradigm: Plotting accuracy against data volume is transferable to other fields (e.g., model compression).
  • Shared dataset infrastructure: A neglected perspective—pointing out that every lab storing its own ImageNet copy constitutes dozens of MWh of wasteful redundancy.

Limitations & Future Work

  • The estimate only covers ImageNet-1K and training/storage costs, excluding embodied costs of collection, cleaning, and transmission.
  • Empirical evidence is limited to image classification; generative models are more sensitive to long-tail data, and coreset migration remains an open problem.
  • Rebound effect: Increased efficiency might lead to "running more experiments" rather than net energy reduction.
  • Policy feasibility: High-level suggestions like carbon caps or data sunset laws require international regulation and are harder to implement than reporting norms.
  • vs. Kandpal & Raffel (2025): They argue the "human labor value" of training data is undervalued; this paper argues the environmental "price" is undervalued.
  • vs. Wang et al. (2025) / Goel et al. (2025): While they call for LLM "scaling down," this paper brings the argument to specific subset selection methods and empirical reporting.
  • vs. Strubell et al. (2019) & Luccioni et al. (2023): These focused on single-model training costs; this paper scales the perspective to dataset-level aggregate downstream costs.

Rating

  • Novelty: ⭐⭐⭐⭐ (Combination of lower-bound accounting + Data-Pareto paradigm + 3-layer framework).
  • Experimental Thoroughness: ⭐⭐⭐⭐ (Conservative, reproducible numbers; multi-architecture evidence).
  • Writing Quality: ⭐⭐⭐⭐⭐ (Clean structure, restrained argumentation, proactive regarding limitations).
  • Value: ⭐⭐⭐⭐ (Provides a template for integrating data frugality into conference and funding standards).