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

Andrew Millard; Gu{\dh}r\'un Fj\'ola Gu{\dh}mundsd\'ottir; Raghavendra Selvan; Sebastian Mair; Sophia N. Wilson

REVIEW 3 major objections 6 minor 15 references

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

T0 review · 3 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read The paper argues that AI research should stop merely advocating for smaller, selected datasets and start actually using them, showing that coreset-based subset selection can cut training energy by roughly a quarter to a third with little or

desk verdict Useful position paper with a genuinely new aggregate ImageNet energy estimate, but the headline energy-accuracy claim is assembled from two measurements that don't connect. read the letter →

arxiv 2602.19789 v2 pith:5ROP4VNF submitted 2026-02-23 cs.LG cs.CY

Sophia N. Wilson , Andrew Millard , Gu{ḥrún Fjóla Gu{ḥmundsdóttir , Raghavendra Selvan , Sebastian Mair This is my paper

classification cs.LGcs.CY

keywords datafrugalitycoresetselectionsubsetpruningenergyefficiencycarbonemissionsdatasetbiassustainableAI

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

This position paper claims that the machine learning community's commitment to data frugality remains rhetorical: everyone praises smaller, curated datasets, but practice still favors ever-larger ones. The authors try to close that preach-practice gap by quantifying the environmental cost of a canonical vision dataset and by showing that coreset selection makes frugality technically viable today. They estimate that downstream use of ImageNet-1K between 2017 and 2025 consumed about 5.46 GWh of electricity for training and 360 MWh for storage, emitting roughly 2,600 tonnes of CO2e. They report that pruning 25% of the dataset reduces per-epoch training energy by 24-33% and time by 24-40%, while state-of-the-art coreset methods prune 25-35% with no drop in accuracy and can also mitigate known dataset biases. If true, data frugality is not a sacrifice but a concrete, immediately actionable route to responsible AI development.

What carries the argument

The central mechanism is the coreset—a small, weighted subset of training points that approximates the full training objective, so a model trained on the subset reaches accuracy comparable to training on the full dataset at much lower compute. The paper operationalises data frugality as subset selection, measures energy per training epoch with a standard tracking tool, and links dataset size to lifecycle costs by estimating downstream training runs and storage copies of a canonical benchmark. The argument runs through the ratio between pruning fraction and measured energy savings (which is not one-to-one), the accuracy-preservation curves of modern coreset methods, and a bias-mitigation expe

What would settle it

Run the same three architectures on ImageNet-1K with a 25% subset chosen by an accuracy-preserving coreset method, track energy per epoch and total energy including subset construction; if per-epoch savings fall well below 24-33% or total energy does not drop after construction costs, the central practical claim fails. A second check: verify that a 25% coreset preserves top-1 accuracy within a small margin across multiple architectures, not just the two reported.

Watch

Extended reading notes

Core claim

The paper's central claim is that data frugality—selecting small representative subsets via coreset construction—is already technically feasible and environmentally meaningful, and that the ML community's failure to adopt it is a value-action gap rather than a technical barrier. To ground this, the authors estimate that approximately 46,179 training runs on ImageNet-1K from 2017 to 2025 consumed about 5.46 GWh and 2,429 tCO2e, with an additional 360 MWh and 160 tCO2e for storage. They then present per-epoch measurements showing that a random 25% data subset reduces training time by 24-40% and energy by 24-33% across three architectures, and they cite coreset results showing that pruning 25-3

Load-bearing premise

The paper assumes that a 25% subset chosen by an accuracy-preserving coreset method saves about the same energy per epoch as a randomly pruned 25% subset, even though the energy measurements come from random pruning and the accuracy measurements come from different coreset methods and architectures.

Editorial extensions

If this is right

Adopting 25% coreset pruning on large image datasets would cut training energy by roughly a quarter to a third per epoch, with no measured accuracy drop for the cited methods.
Storage and distribution energy would fall proportionally with dataset size, reducing the lifecycle emissions the paper estimates for ImageNet-1K.
Balanced coreset selection gives dataset curators an algorithmic lever to reduce known majority-group bias without retraining on the full skewed data.
Reporting energy and carbon metrics alongside accuracy would close the value-action gap, since the tools for measurement already exist and are lightweight.
Data-Pareto reporting—accuracy per data point rather than raw accuracy—would re-align research incentives away from indiscriminate data accumulation.

Reading between the lines

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

The paper combines evidence from separate experiments: energy savings are measured on a random 25% subset, while accuracy preservation is demonstrated by different coreset methods on different architectures; a direct same-setup measurement would make the claim airtight.
If the energy saved by frugality is reinvested in running more experiments, rebound effects could offset the environmental gains, so frugality needs to be paired with incentives that cap total compute rather than merely making each run cheaper.
The bias-mitigation result is demonstrated on a toy color-MNIST example; applying balanced coreset selection to real imbalanced datasets and measuring downstream fairness metrics would test whether the effect scales.
Coreset construction itself has an upfront energy cost that the paper sets aside as a one-time expense; for one-off training runs this cost may exceed the savings, so the strongest case for frugality is in repeated training, benchmarks, and shared dataset infrastructure.

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.

Desk Editor's Note

read the letter

Useful position paper with a genuinely new aggregate estimate of ImageNet-1K downstream energy and carbon, but the headline claim that coreset-based subset selection cuts energy with little accuracy loss is assembled from two measurements that don't connect.

The new thing here is the back-of-the-envelope number: 46,179 training runs on ImageNet-1K from 2017-2025, 5.46 GWh and 2,429 tCO2e, plus 360 MWh for storage. The method — LLM classification of ICLR papers, linear extrapolation, then applying the fraction to dimensions.ai counts — is rough, but it's transparent and the paper labels it a lower bound. That estimate plus the per-epoch energy measurements on a 25% random pruned subset (24-33% energy savings across three architectures) are real, reproducible data points. The discussion of related position papers (Goel et al., Wang et al., McCoy et al.) is fair, and the critique that coreset papers rarely report energy is well aimed.

Now the soft spots, in proportion. The abstract says coreset-based subset selection can reduce training energy with little accuracy loss. That's not what the experiments show. Table 1 measures energy on subsets pruned uniformly at random; Figure 3's no-loss accuracy comes from Dyn-Unc and InfoMax on Swin-T and ResNet-34, from other papers. Nothing measures energy for the same coresets whose accuracy is cited. So the energy-accuracy claim is a plausible inference, not a result. The paper acknowledges it neglects coreset construction cost — and it criticizes other papers for not reporting energy — so this is a genuine gap, though rectifiable. The ImageNet estimate also depends on an unvalidated LLM classification and the transfer of an ICLR fraction to the whole literature; the authors are upfront about this, but the numbers should be treated as indicative, not precise. The bias mitigation section is a toy example on Colored MNIST, fine for illustrating a point, not evidence about ImageNet-scale bias.

Overall, the central position is reasonable, the evidence is honestly presented, and the limitations are stated in the text — which is refreshing. But the headline overstates what the measurements directly support. For a position paper, this is a moderate fix: either add an experiment that measures energy on Dyn-Unc or InfoMax coresets (or equivalent), or soften the abstract to say "data pruning" rather than "coreset-based subset selection."

For peer review: yes, send it. The estimate and the measurements deserve referee time, and the paper is a useful contribution to the data-frugality conversation. I'd advise the editor to ask for that connecting experiment or a sharper abstract. I'd bring it to a reading group — there's a good discussion to be had about what counts as evidence in position papers.

Would I cite it? Probably yes for the ImageNet estimate, with a caveat about the extrapolation.

Referee Report

3 major / 6 minor

Summary. This position paper argues that the machine-learning community should move from rhetorical advocacy to practiced 'data frugality'—using subset selection to reduce the energy, carbon, and storage costs of large datasets. To ground the position, the authors estimate the aggregate downstream energy and carbon cost of ImageNet-1K: approximately 46,179 training runs between 2017 and 2025, 5.46 GWh for training, and 360 MWh for storage. They then present per-epoch time and energy savings for a 25% uniformly random pruned subset on ResNet-34/50 and Swin-T (Table 1), combine this with coreset accuracy results from He et al. (2024) and Tan et al. (2025) (Figure 3), and report a Colour-MNIST experiment showing that balanced subset construction can mitigate bias (Figure 4). The paper closes with recommendations for individuals, platforms, and policy, and it explicitly acknowledges several limitations of the estimates and of coreset methods.

Significance. If the central empirical claim holds, the paper offers a timely and concrete argument that data frugality is actionable rather than aspirational. Its strengths include transparent energy accounting with an environmental impact statement, use of established measurement tools, and an explicit discussion of rebound effects and lifecycle boundaries. The paper also makes falsifiable quantitative estimates and clearly separates its own measurements from literature-reported numbers. However, the main evidence for 'coreset-based subset selection can substantially reduce training energy with little loss in accuracy' is currently assembled from two disconnected measurements: energy is measured on randomly pruned subsets, while accuracy is taken from coreset methods evaluated by other authors on different architectures. The aggregate ImageNet estimate also rests on unvalidated LLM classification and linear extrapolation. These gaps are load-bearing because the paper's contribution is to support a position with quantitative evidence, but they are local and fixable.

major comments (3)

[Section 4.2, Table 1 vs. Figure 3] The headline claim that 'coreset-based subset selection can substantially reduce training energy consumption with little loss in accuracy' (Abstract) is not directly supported by the experiments. Table 1 measures per-epoch time and energy on a 25% subset sampled uniformly at random, while Figure 3 reports accuracy for Dyn-Unc and InfoMax coresets from He et al. (2024) and Tan et al. (2025), evaluated on Swin-T and ResNet-34, respectively. No experiment measures energy on the coreset-selected subsets whose accuracy is cited. Per-epoch energy is not automatically invariant under the selection method: selection changes data distribution, data-loading characteristics, and training dynamics, so a random 25% subset may have different energy than a 25% Dyn-Unc or InfoMax subset. To support the claim, the paper should either measure energy on the coreset subsets or explicitly reframe the energy
[Section 4.2 and Section 5] The treatment of coreset construction cost is contradictory. Section 4.2 states that the computational burden of constructing the coreset is neglected but calls it 'a one-time cost.' Section 5 later acknowledges that 'this upfront cost is only amortised through reuse, limiting its effectiveness in one-off training scenarios.' These two statements cannot both be used to support the energy-savings claim. For a one-off training run, the net energy saving is unknown and could be small or even negative if coreset construction requires significant full-data computation (e.g., Dyn-Unc's uncertainty scoring). Since the paper itself criticizes other coreset papers for not reporting energy costs, Section 4 should provide at least an order-of-magnitude estimate of construction energy or a sensitivity analysis of net savings under one-time versus repeated use.
[Section 3.1 and Appendix B] The aggregate estimate of 46,179 ImageNet training runs rests on several unvalidated assumptions: (i) LLM-based classification of ICLR papers is reported without any accuracy or human validation; (ii) OLS linear regression on 2017–2022 data is extrapolated to 2023–2025, a period during which the number of accepted papers grows from 1,094 to 3,704, making the projection highly sensitive; and (iii) the ICLR 'trained from random initialization' fraction is assumed representative of all ImageNet-mentioning publications in dimensions.ai. The paper then reports point estimates of 5.46 GWh and 2,429 tCO2e without confidence intervals or sensitivity analyses. Since the goal of Section 3 is to make dataset-level costs explicit, these estimates should be accompanied by an uncertainty range and a discussion of how the total changes under plausible variations in each assumption.

minor comments (6)

[Abstract and Section 1] 'For long, progress has been equated' appears in both places; it should read 'For too long.'
[Section 2.2] The phrase 'definedata frugalityas' is missing spacing; it should read 'define data frugality as.'
[Section 4.2] The sentence 'Note that these substantial reductions are possible without noticeable performance losses using SOTA data reduction techniques' appears immediately after Table 1, which was measured on a random subset. Please make explicit at this point that the energy measurement itself uses random pruning, and that the accuracy claim comes from Figure 3.
[Figure 3 caption] The caption says 'there is no performance loss for 25%-30% of data pruning when using Dyn-Unc/InfoMax,' while the main text states Dyn-Unc prunes 25% without loss and InfoMax prunes up to 35%. The caption and text should be aligned.
[Section 6 and Appendix C] 'Data-Pareto reporting' is mentioned without definition; a one-sentence explanation would help. Also, Appendix C lists bias strengths {0.0, 0.75, 0.95} while the main text and Figure 4 use 0.99; please clarify which values are shown in each figure.
[Conclusion] 'treading data as an abundant input' should be 'treating data as an abundant input.'

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: energy and accuracy evidence are independent, with acknowledged limitations.

full rationale

The paper's derivation chain is not circular. The accuracy evidence in Section 4.1 is explicitly imported from external published work (He et al. 2024; Tan et al. 2025), and the per-epoch energy/time figures in Section 4.2 are new direct measurements on a 25% uniformly random subset; no parameter is fitted to the accuracy results and then renamed as a prediction. The paper explicitly acknowledges the main caveat ('we neglect the computational burden of constructing the coreset') and that 'this upfront cost is only amortised through reuse,' which is a stated limitation rather than a circular reduction. Section 3's estimates use externally cited inputs (e.g., 60 kWh/TB/year from Selvan 2025, carbon intensity from IEA), but these are inputs to an illustrative calculation, not the paper's target claim. Self-citations (Wilson et al. 2026; Chen & Selvan 2025; Anthony et al. 2020) motivate terminology or provide measurement tools, and they do not carry the load of the central empirical claim. No equation equates the claimed result with an input by construction, and the energy-accuracy mismatch identified by the reader is an evidentiary/generality gap, not circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central estimates rest on several unverified extrapolations: ICLR-to-all-literature representativeness, energy-per-epoch constancy, and one-copy-per-run storage. The bias claim relies on a toy experiment. No new entities are introduced. The free parameters are standard model choices and statistical fits.

free parameters (3)

OLS linear trend for ImageNet training ratio = slope/intercept from 2017-2022 ICLR data (not reported explicitly)
Used to project 2023-2025 fractions; fitted to observed ICLR ratios and applied to dimensions.ai counts. The extrapolation is a free choice and not independently validated.
Reference training setup (ResNet-50, A100, 300 epochs) = 0.394 kWh/epoch (measured)
Chosen as representative for ImageNet training; energy per epoch measured on 10-epoch runs then multiplied by 300. Model, hardware, and training regime choices affect the estimate.
Storage annual energy intensity = 60 kWh/TB/year
Taken from Selvan (2025), a co-author's book, without independent verification in the paper. Used to compute storage emissions for ImageNet-1K copies.

assumptions (5)

domain assumption ICLR papers are representative of the broader ML literature for ImageNet training practices
Section 3.1 / Appendix B: the ICLR-derived fraction of papers training on ImageNet from random initialization is applied to all ImageNet-mentioning papers in dimensions.ai.
domain assumption Each training run corresponds to one retained local copy of ImageNet-1K for storage accounting
Section 3.2: storage energy is computed as number of runs × 130 GB × 60 kWh/TB/year. In practice copies may be shared or retained longer, so this is a rough proxy.
domain assumption Energy per epoch is constant over 300 epochs
Section 3.1: 0.394 kWh/epoch × 300 epochs. Measurements are from 10-epoch runs; no evidence is given that per-epoch energy remains flat through training (warmup, checkpointing, etc.).
domain assumption LLM-based classification of ImageNet usage from ICLR PDFs is accurate
Appendix B: ChatGPT was used to infer whether papers trained from random initialisation; no validation set, error rate, or human agreement is reported.
domain assumption The toy Colored-MNIST experiment captures real dataset bias dynamics
Section 4.3: conclusions about bias mitigation are drawn from a synthetic dataset with a single known bias axis (color). Real-world bias is more complex and may not respond to rebalancing in the same way.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Position: Stop Preaching and Start Practising Data Frugality for Responsible Development of AI." pith.science (2026). https://pith.science/paper/5ROP4VNF

@misc{pith2026260219789,
  author       = {Pith},
  title        = {Pith review of: Position: Stop Preaching and Start Practising Data Frugality for Responsible Development of AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5ROP4VNF}},
  note         = {Machine review of arXiv:2602.19789}
}

read the original abstract

This position paper argues that the machine learning community must move from preaching to practising data frugality for responsible artificial intelligence (AI) development. For too long, progress has been equated with ever-larger datasets, driving remarkable advances but now yielding increasingly diminishing performance gains alongside rising energy use and carbon emissions. While awareness of data frugal approaches has grown, their adoption has remained rhetorical, and data scaling continues to dominate development practice. We argue that this gap between preach and practice must be closed, as continued data scaling entails substantial and under-accounted environmental impacts. To ground our position, we provide indicative estimates of the energy use and carbon emissions associated with the downstream use of ImageNet-1K. We then present empirical evidence that data frugality is both practical and beneficial, demonstrating that subset selection methods can substantially reduce training energy consumption with little loss in accuracy, while also mitigating dataset bias. Finally, we outline actionable recommendations for moving data frugality from rhetorical preaching to concrete practice for responsible development of AI.

Figures

Figures reproduced from arXiv: 2602.19789 by the authors.

**Figure 1.** Illustration is inspired by a remark by prominent ML researcher Andrew Ng, comparing an AI model with a rocket ship (quoted in Kelly, 2016). However, progress in rocket science depends less on adding more fuel and more on using it wisely, reflecting the principle of data frugality. Created with Nano Banana (Google, 2026). et al., 2025). Yet, prevailing research incentives continue to reward scale, with reviewers, be… view at source ↗

**Figure 2.** Visualisation of a data lifecycle (purple) and a model lifecycle (blue). Subset selection methods can have positive effects on multiple parts across both lifecycles (grey). A similar dynamic is evident in ML, where the growing creation and use of large-scale data has led to a form of hyper-datafication in which the associated costs are largely ignored (Wilson et al., 2026). Within the ML community, concern on the en… view at source ↗

**Figure 3.** Top-1 accuracy of pruning ImageNet-1K using Dyn-Unc (left), D2 (right), and InfoMax (right). Forgetting, Random, and the performance on the full ImageNet-1K are shown as references. The numbers are taken from He et al. (2024) and Tan et al. (2025). Note that the architectures differ, there is no performance loss for 25%-30% of data pruning when using Dyn-Unc/InfoMax, and that the authors do not compare their methods… view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: shows the influence of different data budgets on classification performance. We report the aligned accuracy which is the best accuracy possible if the models do not focus on colours, whereas the conflicted accuracy is shown for the case where 99% of samples belong to …

**Figure 5.** Figure 5: Left: Observed and projected fraction of papers at ICLR with ImageNet use. Right: Observed and projected growth of papers using ImageNet to train models from random initialization. Extrapolating ImageNet usage beyond ICLR. To estimate total ImageNet usage beyond ICLR, …

**Figure 6.** Figure 6: Total number of publications mentioning ImageNet (blue) retrieved from dimensions.ai and the total number of models trained on ImageNet from random initialisation (green), obtained by applying the annual ICLR-derived training ratios to the total publication counts [PI…

**Figure 7.** Figure 7: Total downloads versus dataset size (MB) for the 876 ImageNet-related datasets hosted on the Hugging Face Hub as of 1st December 2025. Among these, 214 datasets explicitly include imagenet-1k in their dataset identifier [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

**Figure 8.** Figure 8: depicts a visualisation of the data used in the experiment described in Section 4.3 [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]

**Figure 9.** Figure 9: Performance of a classifier trained on the biased Colour-MNIST dataset is reported for the case with no bias (aligned) and with bias (conflicting) for four bias strengths (fraction of samples in the majority group): 0.0%, 75%, 95%, and 99%. Three coreset methods are sh…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 4 linked inside Pith

[9]

Toneva, M., Sordoni, A., des Combes, R

URL https://openreview.net/forum? id=93XT0lKOct. Toneva, M., Sordoni, A., des Combes, R. T., Trischler, A., Bengio, Y ., and Gordon, G. J. An empirical study of example forgetting during deep neural network learning. InInternational Conference on Learning Representations,
[11]

Van Horn, G

URL https://openreview.net/forum? id=xcdlSMYXxD. Van Horn, G. and Perona, P. The devil is in the tails: Fine- grained classification in the wild. arxiv.arXiv preprint arXiv:1709.01450, 2017. Verdecchia, R., Cruz, L., Sallou, J., Lin, M., Wickenden, J., and Hotellier, E. Data-centric green ai an exploratory empirical study. In2022 International Conference ...

arXiv 2017
[13]

13 Stop Preaching and Start Practising Data Frugality for Responsible Development of AI A

URL https://openreview.net/forum? id=FjxyAotxtT. 13 Stop Preaching and Start Practising Data Frugality for Responsible Development of AI A. Overview of Data Reduction Techniques Coreset selection methods.The goal of arepresentative subsetis to approximate a sum over a large number of terms with fewer terms, i.e., f(θ) := Pn i=1 wifi(θ)≈ Pn i∈I ˜wifi(θ) =:...

2025
[14]

A combination of prediction uncertainty and training dynamics when scoring data points is approached by Dyn-Unc (He et al.,

demonstrates speed-ups of up to 2.5× on various vision and NLP benchmarks with minimal performance loss. A combination of prediction uncertainty and training dynamics when scoring data points is approached by Dyn-Unc (He et al.,
[15]

ImageNet

while InfoMax (Tan et al., 2025) selects subsets that maximize representative information content. Both approaches demonstrate that ImageNet-1k can be pruned up to 25% with negligible to no performance degradation. More generally, training-dynamics-based approaches also identify important examples by explicitly monitoring learning behaviour, such as forge...

2025
[577]

acl-long.577/

URL https://aclanthology.org/2022. acl-long.577/. Lei, Y ., Shan, C., and Ge, W. Edgecasednet: An enhanced detection architecture for edge case perception in au- tonomous driving.Plos one, 20(12):e0338638, 2025. Liu, Y .-A., Zhang, R., Guo, J., de Rijke, M., Fan, Y ., and Cheng, X. On the scaling of robustness and effectiveness in dense retrieval. InProce...

arXiv 2022
[2018]

Accessed: 27 January 2026

URL https://www.veritas.com/ content/dam/Veritas/docs/reports/ Veritas-Data-Hoarders-Report-US.pdf . Accessed: 27 January 2026. Violos, J., Diamanti, K.-C., Kompatsiaris, I., and Pa- padopoulos, S. Frugal machine learning for energy- efficient, and resource-aware artificial intelligence.arXiv preprint arXiv:2506.01869, 2025. Wang, Y ., Li, Y ., and Xu, C....

arXiv 2026
[2019]

12 Stop Preaching and Start Practising Data Frugality for Responsible Development of AI Upadhyay, S

URL https://openreview.net/forum? id=BJlxm30cKm. 12 Stop Preaching and Start Practising Data Frugality for Responsible Development of AI Upadhyay, S. K., Quirke, P., Oozeer, N. F., and Bandi, C. Po- sition: Require frontier AI labs to release small “analog” models. InThe Thirty-Ninth Annual Conference on Neu- ral Information Processing Systems Position Pa...

Show all 15 references

[2020]

URL https://openreview.net/forum? id=HJg2b0VYDr. Courty, B., Schmidt, V ., Luccioni, S., Goyal-Kamal, Mari- onCoutarel, Feld, B., Lecourt, J., LiamConnell, Saboni, A., Inimaz, supatomic, Léval, M., Blanche, L., Cru- veiller, A., ouminasara, Zhao, F., Joshi, A., Bogroff, A., de...

2024
[2022]

Lee, K., Ippolito, D., Nystrom, A., Zhang, C., Eck, D., Callison-Burch, C., and Carlini, N

URL https://openreview.net/forum? id=ajH17-Pb43A. Lee, K., Ippolito, D., Nystrom, A., Zhang, C., Eck, D., Callison-Burch, C., and Carlini, N. Deduplicating train- ing data makes language models better. In Muresan, S., Nakov, P., and Villavicencio, A. (eds.),Proceedings of the ...

2022 doi
[2023]

Rather, I

URL https://openreview.net/forum? id=kM5eGcdCzq. Rather, I. H., Kumar, S., and Gandomi, A. H. Breaking the data barrier: a review of deep learning techniques for de- mocratizing ai with small datasets.Artificial Intelligence Review, 57(9):226, 2024. Ritchie, H., Rosado, P., an...

2024
[2024]

Fayyaz, M., Aghazadeh, E., Modarressi, A., Pilehvar, M

URL https://openreview.net/forum? id=aIok3ZD9to. Fayyaz, M., Aghazadeh, E., Modarressi, A., Pilehvar, M. T., Yaghoobzadeh, Y ., and Kahou, S. E. Bert on a data diet: Finding important examples by gradient-based pruning. arXiv preprint arXiv:2211.05610, 2022. Feldman, D. Core-s...

2022 arXiv
[2025]

Scala, F., Flesca, S., and Pontieri, L

URL https://openreview.net/forum? id=4UhTWPwVke. Scala, F., Flesca, S., and Pontieri, L. An efficient model training framework for green ai.Machine Learning, 114 (12):275, 2025. Selvan, R.Sustainable AI: Tools for Moving toward Green AI. O’Reilly Media, 2025. ISBN: 97810981555...

2025 arXiv
[2534]

Kelly, K.The Inevitable: Understanding the 12 Technolog- ical Forces That Will Shape Our Future

PMLR, 2018. Kelly, K.The Inevitable: Understanding the 12 Technolog- ical Forces That Will Shape Our Future. Viking Press, New York, 2016. Quote attributed to Andrew Ng. Killamsetty, K., Durga, S., Ramakrishnan, G., De, A., and Iyer, R. Grad-match: Gradient matching based data...

2018
[6590]

9 Stop Preaching and Start Practising Data Frugality for Responsible Development of AI Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L

PMLR, 2023. 9 Stop Preaching and Start Practising Data Frugality for Responsible Development of AI Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In2009 IEEE Conference on Computer Vision and Pat- tern Re...

2023

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [9]

Toneva, M., Sordoni, A., des Combes, R

URL https://openreview.net/forum? id=93XT0lKOct. Toneva, M., Sordoni, A., des Combes, R. T., Trischler, A., Bengio, Y ., and Gordon, G. J. An empirical study of example forgetting during deep neural network learning. InInternational Conference on Learning Representations,

[2] [11]

Van Horn, G

URL https://openreview.net/forum? id=xcdlSMYXxD. Van Horn, G. and Perona, P. The devil is in the tails: Fine- grained classification in the wild. arxiv.arXiv preprint arXiv:1709.01450, 2017. Verdecchia, R., Cruz, L., Sallou, J., Lin, M., Wickenden, J., and Hotellier, E. Data-centric green ai an exploratory empirical study. In2022 International Conference ...

arXiv 2017

[3] [13]

13 Stop Preaching and Start Practising Data Frugality for Responsible Development of AI A

URL https://openreview.net/forum? id=FjxyAotxtT. 13 Stop Preaching and Start Practising Data Frugality for Responsible Development of AI A. Overview of Data Reduction Techniques Coreset selection methods.The goal of arepresentative subsetis to approximate a sum over a large number of terms with fewer terms, i.e., f(θ) := Pn i=1 wifi(θ)≈ Pn i∈I ˜wifi(θ) =:...

2025

[4] [14]

A combination of prediction uncertainty and training dynamics when scoring data points is approached by Dyn-Unc (He et al.,

demonstrates speed-ups of up to 2.5× on various vision and NLP benchmarks with minimal performance loss. A combination of prediction uncertainty and training dynamics when scoring data points is approached by Dyn-Unc (He et al.,

[5] [15]

ImageNet

while InfoMax (Tan et al., 2025) selects subsets that maximize representative information content. Both approaches demonstrate that ImageNet-1k can be pruned up to 25% with negligible to no performance degradation. More generally, training-dynamics-based approaches also identify important examples by explicitly monitoring learning behaviour, such as forge...

2025

[6] [577]

acl-long.577/

URL https://aclanthology.org/2022. acl-long.577/. Lei, Y ., Shan, C., and Ge, W. Edgecasednet: An enhanced detection architecture for edge case perception in au- tonomous driving.Plos one, 20(12):e0338638, 2025. Liu, Y .-A., Zhang, R., Guo, J., de Rijke, M., Fan, Y ., and Cheng, X. On the scaling of robustness and effectiveness in dense retrieval. InProce...

arXiv 2022

[7] [2018]

Accessed: 27 January 2026

URL https://www.veritas.com/ content/dam/Veritas/docs/reports/ Veritas-Data-Hoarders-Report-US.pdf . Accessed: 27 January 2026. Violos, J., Diamanti, K.-C., Kompatsiaris, I., and Pa- padopoulos, S. Frugal machine learning for energy- efficient, and resource-aware artificial intelligence.arXiv preprint arXiv:2506.01869, 2025. Wang, Y ., Li, Y ., and Xu, C....

arXiv 2026

[8] [2019]

12 Stop Preaching and Start Practising Data Frugality for Responsible Development of AI Upadhyay, S

URL https://openreview.net/forum? id=BJlxm30cKm. 12 Stop Preaching and Start Practising Data Frugality for Responsible Development of AI Upadhyay, S. K., Quirke, P., Oozeer, N. F., and Bandi, C. Po- sition: Require frontier AI labs to release small “analog” models. InThe Thirty-Ninth Annual Conference on Neu- ral Information Processing Systems Position Pa...