Pith. sign in

REVIEW 3 major objections 4 minor 42 references

Energy Considerations for Large Pretrained Neural Networks

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Quantizing pretrained CNN weights to 8–14 bits before fine-tuning cuts training electricity by more than half in six of nine models, while pruning and low-rank factorization do not deliver consistent energy savings.

desk verdict Asking the right energy question with a nine-model benchmark, but the power-measurement chain is missing and baseline inconsistencies undermine the headline numbers; still worth a referee's time as a fixable major revision. read the letter →

arxiv 2506.01311 v1 pith:LCOAJLDE submitted 2025-06-02 cs.LG

Leo Mei , Mark Stamp This is my paper
classification cs.LG
keywords energyconsumptionmodelcompressionsteganographiccapacityquantizationpruninglow-rankfactorizationpretrainedCNNsImageNet
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

The paper tries to establish that the cheapest energy win in fine-tuning large pretrained CNNs comes from bit-level quantization, not from removing parameters. It reports that overwriting the 18 to 24 least significant bits of each 32-bit pretrained weight, then fine-tuning the remaining 8 to 14 bits, cuts total training electricity by more than half for six of the nine models tested, with accuracy loss held under 1%. By contrast, pruning and low-rank factorization, which remove weights rather than reduce their precision, fail to produce consistent energy savings and sometimes increase energy use. If this holds, the practical route to cutting the carbon cost of adapting large vision models is to shrink weight precision before training, not to make the network sparse or low-rank.

What carries the argument

The central object is steganographic capacity: the number of the lowest-order bits of each 32-bit floating-point weight that can be overwritten before validation accuracy drops by more than 1%. The paper measures this capacity for every model, sets each weight's quantization level to the retained high-order bits (8 to 14 bits per weight), and then fine-tunes the quantized model. The comparison techniques are global L1 unstructured pruning and truncated SVD low-rank factorization, with energy computed as $P_{\text{total}} = P_{\text{gpu}} + P_{\text{cpu}} + P_{\text{ram}}$, scaled by a PUE of 1.58. The mechanism that carries the argument is per-epoch energy: stego-compressed models use much less electricity per epoch, so even when they need more epochs than the baseline, total energy still falls.

What would settle it

Re-run the nine fine-tuning experiments on a machine with a calibrated power monitor, such as a wall-plug wattmeter or GPU-sampled energy counters, and compare the ratio of compressed to baseline electricity; if steganographic capacity reduction no longer delivers roughly half the baseline energy in most models, the central claim fails. Alternatively, locating the undisclosed power-sampling code and showing that its readings differ materially from direct measurements would settle it.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a model's steganographic capacity—the number of low-order bits per weight that can be zeroed before accuracy drops by more than 1%—identifies a quantization level that is also an energy lever. Across nine pretrained ImageNet classifiers (AlexNet, ResNet18/34/50/101, InceptionV3, DenseNet121, VGG16, and ConvNeXt), the authors measure total training electricity as the sum of GPU, CPU, and RAM power scaled by the data center PUE of 1.58. They find that steganographic capacity reduction (quantizing to 8–14 bits per weight) reduces energy usage by more than half for ResNet50, ResNet101, InceptionV3, DenseNet121, VGG16, and ConvNeXt, while pruning and low-rank factorization do not achieve consistent reductions and in several cases increase energy use. The paper further reports that stego-compressed models reach comparable accuracy and often train in about half the time, despite sometimes needing more epochs, because per-epoch energy is much lower.

Load-bearing premise

The energy comparison stands on the assumption that the GPU, CPU, and RAM power readings recorded during cloud notebook runs are accurate and consistent, and that scaling them by a fixed PUE of 1.58 yields a fair relative energy value; Section 2.3 says the team could not directly measure energy but does not say which tool or sampling method produced the readings.

Editorial extensions

If this is right

  • Quantizing pretrained CNNs to 8–14 bits per weight before fine-tuning can reduce training electricity by more than half for models like ResNet50, ResNet101, InceptionV3, DenseNet121, VGG16, and ConvNeXt, with accuracy loss below 1%.
  • Pruning and low-rank factorization, at least as applied here, are not reliable energy-saving strategies; they often leave energy use flat or raise it.
  • Because the savings show up as lower energy per epoch rather than fewer epochs, the benefit should persist even when compressed models need additional fine-tuning rounds.
  • Tolerating a slightly larger accuracy drop is likely to unlock further reductions in both training time and electricity, since the paper's 1% threshold is an arbitrary cap.
  • Combining bit-level quantization with pruning or low-rank factorization is a plausible way to compound compression and energy savings, since the methods act at different levels: bits versus weights and matrices.

Reading between the lines

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

  • If the same bit-level slack appears in transformer weights, quantization could lower fine-tuning energy for large language models too; the paper only tests CNNs and leaves this untested.
  • The contrast suggests the energy cost of fine-tuning is dominated by memory traffic and arithmetic precision, not parameter count: removing weights does not shrink per-epoch energy, but reducing bit-width does.
  • A direct measurement of GPU power during fine-tuning, rather than the PUE-scaled estimate, would tell whether the "more than half" reduction is a true hardware effect or partly an artifact of the power sampling method.
  • Varying batch size, optimizer, and early-stopping patience in a follow-up would test whether the energy advantage of quantization is robust across training regimes.
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.

Referee Report

3 major / 4 minor

Summary. The paper investigates whether three model-compression techniques—steganographic capacity reduction (low-order bit overwriting), pruning, and low-rank factorization—reduce the energy consumed when fine-tuning nine pretrained CNNs on ImageNet1K. The authors train each model in baseline and compressed forms on Kaggle, record accuracy, training time, epochs, and energy (GPU+CPU+RAM scaled by PUE 1.58), and report that steganographic capacity reduction consistently reduces energy and training time, while pruning and low-rank factorization do not. The central quantitative claim is that steganographic capacity reduction cuts fine-tuning energy by more than half for six of the nine models.

Significance. The question addressed—whether compression can lower the electricity cost of fine-tuning—is timely and practically relevant. The study is commendable for applying three compression methods under a common 1% accuracy-loss threshold and for attempting to measure full-system energy rather than relying on model size as a proxy. If the energy measurements were trustworthy, the consistent advantage of bit-level quantization over parameter-removal methods would be a useful empirical result for energy-efficient ML. However, the paper does not provide a verifiable power-measurement protocol, contains internal inconsistencies in baseline accuracies, and does not explain how zeroing low-order bits of 32-bit weights reduces per-epoch energy. The significance of the findings is therefore conditional on resolving these issues.

major comments (3)
  1. [2.3, Figures 2, 4, 6] The energy measurements that underpin the central claim are not described. Section 2.3 defines P_total as the sum of P_gpu, P_cpu, and P_ram and multiplies by a fixed PUE of 1.58, but it never states how P_gpu, P_cpu, and P_ram were obtained, which tool or API was used, how the components were sampled, or whether the readings are instantaneous powers or averaged over the run. The authors even write that on Kaggle "we cannot directly measure energy consumption." Without this measurement chain, the kWh values in Figures 2, 4, and 6, and the claim in Section 3.6 that steganographic capacity reduction cuts energy by more than half for six models, are not reproducible or verifiable.
  2. [3.2, Table 6, Figure 5] Baseline accuracies are inconsistent across the paper. For example, AlexNet is reported as 0.6062 in Section 3.2, but as 0.5108 in Table 6 and 0.5108 in Figure 5; VGG16 is 0.6930 in Section 3.2, 0.6550 in Table 6, and 0.6546 in Figure 5; and ConvNeXt is 0.8886 in Section 3.2 but 0.8020 in Table 6. Since each compression threshold is defined as a 1% drop relative to the baseline, this discrepancy changes which bit level or pruning ratio is selected and undermines the compression ratios reported in Table 6.
  3. [3.2, Figure 6] The claimed per-epoch energy reductions from steganographic capacity reduction are not mechanistically explained. The experiment overwrites low-order bits of weights that remain in 32-bit format, so the tensor shapes and the number of floating-point operations are unchanged; unless the authors implemented low-precision storage and kernels, per-epoch energy should not change by the large factors shown in Figure 6 (e.g., ResNet50 from 0.71 to 0.32 kWh/epoch). No such implementation detail or CUDA kernel change is provided, so the energy-per-epoch comparison cannot be interpreted.
minor comments (4)
  1. [General] The experiments appear to be single runs without repeated trials, error bars, or statistical tests; given that the central quantitative claims are ratios of energy measurements, the absence of variance information makes it impossible to assess whether the observed differences are within run-to-run noise.
  2. [2.3] The formula P_total = P_gpu + P_cpu + P_ram lacks units and an explicit time-integration step; please clarify that energy is power multiplied by duration and specify how the duration is measured for each run.
  3. [Various] There are numerous typos and formatting issues, including "DensNet121" in the Figure 1 caption, "arachitectures" in Section 2.1.3, "considere" in Section 3.3, "the line graphs is Figure 6" in Section 3.5, "or our experiments" in Section 3, and "ResNet 34" with a space in Table 7; a careful proofread is needed.
  4. [Abstract / Section 2.4] The abstract says "we first train each model without compression," but the experiments fine-tune pretrained models; please use consistent terminology throughout (e.g., "fine-tune" or "retrain").

Circularity Check

1 steps flagged · score 1.0 of 10

Central energy comparison is not circular; the accuracy-preservation claim is the only definitional statement, and it is guaranteed by the 1% threshold used to define capacity.

  1. self definitional [Section 2.2.1 (capacity definition) and Section 3.6 (Discussion)]
    "We use a threshold of a decrease in accuracy of no more than 1%. ... None of the models suffer a loss in accuracy greater than 1% under steganographic capacity reduction."

    The steganographic capacity of each model is defined in Section 3.2 as the largest number n of low-order bits that can be overwritten while keeping the accuracy drop below the 1% threshold. Therefore the Discussion's statement that no steganographically compressed model loses more than 1% accuracy is true by construction: the compression level was selected to satisfy that exact bound. It is a restatement of the selection rule, not an empirical discovery. This tautology does not extend to the paper's main energy-saving claim, because energy and training time are recorded independently of the capacity selection.

full rationale

The paper's central claim, that steganographic capacity reduction lowers fine-tuning energy while pruning and low-rank factorization do not, rests on measured energy, training time, and epoch counts, not on the compression ratios or the 1% accuracy threshold. The steganographic capacities are re-measured here rather than imported from the cited prior work, which is used only as motivation. Pruning and low-rank results are likewise benchmarked against the same measured baseline. I find no fitted parameter that is later relabeled as a prediction, no uniqueness theorem used to force a choice, and no ansatz smuggled in by citation. The identified self-definitional step is limited to the accuracy-preservation statement, which is guaranteed by the 1% selection threshold. The unstated power-measurement method in Section 2.3 is a reproducibility and correctness concern, not a circularity, since an unverified measurement is not an input from which the conclusion is derived by construction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The energy comparison depends on hand-chosen thresholds (1% accuracy drop, early stopping rule, learning rate, momentum, weight decay), an assumed constant data center PUE, and an unstated power measurement method. Steganographic capacity is an empirical property re-measured here. No new physical entities are introduced.

free parameters (6)
  • Maximum allowed accuracy drop = 1%
    Used to set the compression ratio for all three techniques; a different threshold would change ratios and energy results.
  • Early stopping patience = 3 epochs
    Training stops after three epochs without a 5% validation loss improvement, directly controlling total epochs and energy.
  • Early stopping min delta = 0.05
    Minimum improvement in validation loss required to continue training; part of the stopping rule.
  • Learning rate = 0.01
    Chosen from the VGG paper; affects convergence speed and therefore energy per model.
  • Momentum = 0.9
    Standard SGD momentum value; affects training dynamics and convergence.
  • Weight decay = 0.0001
    Hand-chosen regularization strength; affects training.
assumptions (5)
  • domain assumption Pretrained CNN weights contain redundant low-order bits that can be overwritten without meaningful accuracy loss (steganographic capacity).
    Taken from Agrawal et al. [1] and re-measured in Section 3.2; if false, the quantization step would break the models.
  • domain assumption Total training power is P_gpu + P_cpu + P_ram, and data center overhead is a constant PUE of 1.58 times that total.
    Stated in Section 2.3; the actual method for measuring each power term on Kaggle is not described.
  • domain assumption A fixed hyperparameter set (SGD, lr 0.01, momentum 0.9, weight decay 0.0001) is fair across all compression techniques.
    Stated in Section 3.1; if a technique needs different hyperparameters, the energy comparison may be biased.
  • domain assumption Early stopping with patience 3 and min delta 0.05 reaches a comparable stopping point across conditions.
    Stated in Table 5; this rule determines the number of epochs and hence the energy totals.
  • standard math Validation accuracy on ImageNet is an appropriate criterion for choosing compression ratios.
    Used in Sections 3.2 to 3.4 to apply the 1% threshold.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Energy Considerations for Large Pretrained Neural Networks." pith.science (2026). https://pith.science/paper/LCOAJLDE

@misc{pith2026250601311,
  author       = {Pith},
  title        = {Pith review of: Energy Considerations for Large Pretrained Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LCOAJLDE}},
  note         = {Machine review of arXiv:2506.01311}
}
read the original abstract

Increasingly complex neural network architectures have achieved phenomenal performance. However, these complex models require massive computational resources that consume substantial amounts of electricity, which highlights the potential environmental impact of such models. Previous studies have demonstrated that substantial redundancies exist in large pre-trained models. However, previous work has primarily focused on compressing models while retaining comparable model performance, and the direct impact on electricity consumption appears to have received relatively little attention. By quantifying the energy usage associated with both uncompressed and compressed models, we investigate compression as a means of reducing electricity consumption. We consider nine different pre-trained models, ranging in size from 8M parameters to 138M parameters. To establish a baseline, we first train each model without compression and record the electricity usage and time required during training, along with other relevant statistics. We then apply three compression techniques: Steganographic capacity reduction, pruning, and low-rank factorization. In each of the resulting cases, we again measure the electricity usage, training time, model accuracy, and so on. We find that pruning and low-rank factorization offer no significant improvements with respect to energy usage or other related statistics, while steganographic capacity reduction provides major benefits in almost every case. We discuss the significance of these findings.

Figures

Figures reproduced from arXiv: 2506.01311 by the authors.

Figure 1
Figure 1. Steganographic capacity results ResNet101 — ResNet101 has the highest accuracy of the ResNet models con￾sidered, at 0.8063. This is not surprising, as it is the deepest architecture of our four ResNet models. However, as shown in [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Energy consumption and accuracy perform similarly. On the other hand, the steganographic capacity technique generally reduces the energy usage by a large margin [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Training time and compression rate every model except AlexNet, steganographic capacity reduction has the smallest energy usage per epoch. From this graph and [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Energy usage vs compression rate relative to baseline [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]
Figure 5
Figure 5. Figure 5: Accuracy and number of epochs 21 [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: Energy consumption per epoch Based on our experimental results, it is clear that reducing the size of model weights via quantization can greatly improve model training efficiency. However, reducing the number of model weights via pruning or low-rank factorization does …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 39 canonical work pages

  1. [1]

    On the steganographic capacity of selected learning models

    Rishit Agrawal et al. On the steganographic capacity of selected learning models. In Mark Stamp and Martin Jureˇ cek, editors,Machine Learning, Deep Learning, and AI for Cybersecurity. Springer, 2025

    work page 2025

  2. [2]

    S. B. Atallah et al. How large language models including generative pre- trained transformer (GPT) 3 and 4 will impact medicine and surgery.Tech- niques in Coloproctology, 27(8):609–614, 2023

    work page 2023

  3. [3]

    ASHRAE, Atlanta, GA, 2014

    Victor Avelar, Dan Azevedo, and Alan French.PUE: A Comprehensive Examination of the Metric. ASHRAE, Atlanta, GA, 2014

    work page 2014

  4. [4]

    Springer, 2019

    Valentina Emilia Balas et al., editors.Handbook of Deep Learning Applica- tions. Springer, 2019

    work page 2019

  5. [5]

    Bender, D

    W. Bender, D. Gruhl, N. Morimoto, and A. Lu. Techniques for data hiding. IBM Systems Journal, 35(3.4):313–336, 1996

    work page 1996

  6. [6]

    Bench- mark analysis of representative deep neural network architectures.IEEE Access, 6:64270–64277, 2018

    Simone Bianco, Remi Cadene, Luigi Celona, and Paolo Napoletano. Bench- mark analysis of representative deep neural network architectures.IEEE Access, 6:64270–64277, 2018

    work page 2018

  7. [7]

    Uptime Institute Blog: Global PUEs — Are they going anywhere?https://journal.uptimeinstitute.com/global-pues-are- they-going-anywhere/, 2023

    Daniel Bizo. Uptime Institute Blog: Global PUEs — Are they going anywhere?https://journal.uptimeinstitute.com/global-pues-are- they-going-anywhere/, 2023

    work page 2023

  8. [8]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, and Melanie Subbiah. Language models are few-shot learners. InAdvances in Neural Information Processing Systems, volume 33, pages 1877–1901, 2020. 23

    work page 1901

Show all 42 references
  1. [9]

    Model compression

    Cristian Bucilu˘ a, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. InProceedings of the 12th ACM SIGKDD International Con- ference on Knowledge Discovery and Data Mining, KDD ’06, pages 535–541, 2006

    2006

  2. [10]

    Model compression and acceleration for deep neural networks: The principles, progress, and challenges.IEEE Signal Processing Magazine, 35(1):126–136, 2018

    Yu Cheng, Duo Wang, Pan Zhou, and Tao Zhang. Model compression and acceleration for deep neural networks: The principles, progress, and challenges.IEEE Signal Processing Magazine, 35(1):126–136, 2018

    2018

  3. [11]

    ImageNet: A large-scale hierarchical image database

    Jia Deng et al. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009

    2009

  4. [12]

    Exploiting linear structure within convolutional net- works for efficient evaluation

    Emily Denton et al. Exploiting linear structure within convolutional net- works for efficient evaluation. InProceedings of the 28th International Con- ference on Neural Information Processing Systems — Volume 1, NIPS’14, pages 1269–1277, 2014

    2014

  5. [13]

    The approximation of one matrix by another of lower rank.Psychometrika, 1(3):211–218, 1936

    Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank.Psychometrika, 1(3):211–218, 1936

    1936

  6. [14]

    Peer review of GPT-4 technical report and systems card.PLOS Digital Health, 3(1):e0000417, 2024

    Jack Gallifant et al. Peer review of GPT-4 technical report and systems card.PLOS Digital Health, 3(1):e0000417, 2024

    2024

  7. [15]

    Song Han, Huizi Mao, and William J. Dally. Deep compression: Compress- ing deep neural networks with pruning, trained quantization and Huffman coding.https://arxiv.org/abs/1510.00149, 2016

    2016 arXiv

  8. [16]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR, pages 770–778, 2016

    2016

  9. [17]

    Model complexity of deep learning: A survey.Knowledge and Information Systems, 63(10):2585–2619, 2021

    Xia Hu et al. Model complexity of deep learning: A survey.Knowledge and Information Systems, 63(10):2585–2619, 2021

    2021

  10. [18]

    Wein- berger

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Wein- berger. Densely connected convolutional networks. In2017 IEEE Con- ference on Computer Vision and Pattern Recognition, CVPR, pages 2261– 2269, 2017

    2017

  11. [19]

    Rotational equilib- rium: How weight decay balances learning across neural networks.https: //arxiv.org/abs/2305.17212, 2023

    Atli Kosson, Bettina Messmer, and Martin Jaggi. Rotational equilib- rium: How weight decay balances learning across neural networks.https: //arxiv.org/abs/2305.17212, 2023

    2023 arXiv

  12. [20]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet classi- fication with deep convolutional neural networks.Communications of the ACM, 60:84–90, 2017

    2017

  13. [21]

    Pruning filters with L1-norm and capped L1-norm for CNN compression.Applied Intelligence, 51(2):1152–1160, 2021

    Aakash Kumar et al. Pruning filters with L1-norm and capped L1-norm for CNN compression.Applied Intelligence, 51(2):1152–1160, 2021

    2021

  14. [22]

    Deep learning.Nature, 521(7553):436–444, 2015

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning.Nature, 521(7553):436–444, 2015. 24

    2015

  15. [23]

    On-demand deep model compression for mobile devices: A usage-driven model selection framework

    Sicong Liu et al. On-demand deep model compression for mobile devices: A usage-driven model selection framework. InProceedings of the 16th Annual International Conference on Mobile Systems, Applications, and Services, pages 389–400, Munich Germany, 2018

    2018

  16. [24]

    A ConvNet for the 2020s

    Zhuang Liu et al. A ConvNet for the 2020s. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 11966–11976, 2022

    2022

  17. [25]

    A survey of related research on compression and acceleration of deep neural networks.Journal of Physics: Conference Series, 1213(5):052003, 2019

    Xin Long, Zongcheng Ben, and Yan Liu. A survey of related research on compression and acceleration of deep neural networks.Journal of Physics: Conference Series, 1213(5):052003, 2019

    2019

  18. [26]

    Deep neural networks compression: A com- parative survey and choice recommendations.Neurocomputing, 520:152–170, 2023

    Giosu´ e Cataldo Marin´ o et al. Deep neural networks compression: A com- parative survey and choice recommendations.Neurocomputing, 520:152–170, 2023

    2023

  19. [27]

    Artificial intelligence index report 2023.https:// arxiv.org/abs/2310.03715, 2023

    Nestor Maslej et al. Artificial intelligence index report 2023.https:// arxiv.org/abs/2310.03715, 2023

    2023 arXiv

  20. [28]

    Information hiding: Steganography and watermarking— Attacks and countermeasures.Journal of Electronic Imaging, 10(3):825, 2001

    Nasir Memon. Information hiding: Steganography and watermarking— Attacks and countermeasures.Journal of Electronic Imaging, 10(3):825, 2001

    2001

  21. [29]

    PyTorch: Pruning tutorial.https://docs.pytorch

    Michela Paganini. PyTorch: Pruning tutorial.https://docs.pytorch. org/tutorials/intermediate/pruning_tutorial.html, 2024

    2024

  22. [30]

    To compress, or not to compress: Characterizing deep learn- ing model compression for embedded inference

    Qing Qin et al. To compress, or not to compress: Characterizing deep learn- ing model compression for embedded inference. In2018 IEEE International Conference on Parallel & Distributed Processing with Applications, Ubiq- uitous Computing & Communications, Big Data & Cloud Comp...

    2018

  23. [31]

    Evaluation metrics and statis- tical tests for machine learning.Scientific Reports, 14(1):6086, 2024

    Oona Rainio, Jarmo Teuho, and Riku Kl´ en. Evaluation metrics and statis- tical tests for machine learning.Scientific Reports, 14(1):6086, 2024

    2024

  24. [32]

    R. Reed. Pruning algorithms—A survey.IEEE Transactions on Neural Networks, 4(5):740–747, 1993

    1993

  25. [33]

    Sainath et al

    Tara N. Sainath et al. Low-rank matrix factorization for deep neural network training with high-dimensional output targets. In2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 6655–6659, 2013

    2013

  26. [34]

    From words to watts: Benchmarking the energy costs of large language model inference

    Siddharth Samsi et al. From words to watts: Benchmarking the energy costs of large language model inference. In2023 IEEE High Performance Extreme Computing Conference, HPEC, pages 1–9, 2023

    2023

  27. [35]

    Very deep convolutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In3rd International Conference on Learn- ing Representations, ICLR, pages 1–14, 2015. 25

    2015

  28. [36]

    Petitcolas Stefan Katzenbeisser

    Fabien A.P. Petitcolas Stefan Katzenbeisser. Information hiding techniques for steganography and digital watermarking.The EDP Audit, Control, and Security Newsletter, 28(6):1–2, 2000

    2000

  29. [37]

    Going deeper with convolutions

    Christian Szegedy et al. Going deeper with convolutions. In2015 IEEE Conference on Computer Vision and Pattern Recognition, CVPR, pages 1– 9, 2015

    2015

  30. [38]

    Thompson, Marc Schonwiesner, Yoshua Bengio, and Daniel Willett

    Jessica A.F. Thompson, Marc Schonwiesner, Yoshua Bengio, and Daniel Willett. How transferable are features in convolutional neural network acoustic models across languages? In2019 IEEE International Confer- ence on Acoustics, Speech and Signal Processing, ICASSP, pages 2827–2831, 2019

    2019

  31. [39]

    Thompson, Kristjan Greenewald, Keeheon Lee, and Gabriel F

    Neil C. Thompson, Kristjan Greenewald, Keeheon Lee, and Gabriel F. Manso. Deep learning’s diminishing returns: The cost of improvement is becoming unsustainable.IEEE Spectrum, 58(10):50–55, 2021

    2021

  32. [40]

    The marginal value of adaptive gradient methods in machine learning

    Ashia C Wilson et al. The marginal value of adaptive gradient methods in machine learning. InAdvances in Neural Information Processing Systems, volume 30, 2017

    2017

  33. [41]

    Scaling for edge inference of deep neural networks.Nature Electronics, 1(4):216–222, 2018

    Xiaowei Xu et al. Scaling for edge inference of deep neural networks.Nature Electronics, 1(4):216–222, 2018

    2018

  34. [42]

    Quantization networks

    Jiwei Yang et al. Quantization networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 7308–7316, 2019. 26

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.