Pith. sign in

REVIEW 3 major objections 6 minor 45 references

Adapting only the normalization layers of a frozen backbone, plus a margin-weighted head on precomputed features, matches or beats fine-tuning on most benchmarks at a fraction of the training cost.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 11:20 UTC pith:4ALBNUHZ

load-bearing objection A clearly specified, genuinely efficient decoupled transfer-learning pipeline; the speed and CO2 numbers are credible, but the accuracy comparison uses a linear-probing baseline labeled 'fine-tuning', so the central 'marginal trade-off' claim is not yet supported. the 3 major comments →

arxiv 2607.13043 v1 pith:4ALBNUHZ submitted 2026-06-13 cs.LG cs.AIcs.CV

Beyond Backbone Backpropagation: A Decoupled Strategy for Efficient Transfer Learning

Daniel Vila-Cruz , Laura Mor\'an-Fern\'andez , Ver\'onica Bol\'on-Canedo This is my paper
classification cs.LG cs.AIcs.CV
keywords transfer learningnormalization adaptationfrozen backbonedecoupled classifiermargin-based reweightingenergy-efficient trainingmedical image classificationvision transformers
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper claims that the expensive step in transfer learning—backpropagating through a pretrained backbone—is often unnecessary. For CNN and Transformer backbones frozen after pretraining, the authors adapt only the normalization layers (BatchNorm running statistics for CNNs, LayerNorm bias shifts for Transformers), precompute features once, and train a redesigned margin-weighted classifier head on those stored features. Across seven architectures and three medical imaging datasets, they report accuracy competitive with fine-tuning or low-rank adaptation, often matching or surpassing the baseline, while cutting training time and CO2 emissions by roughly an order of magnitude or more, frequently 10–20x. In most tested cases the frozen-backbone pipeline finishes faster on a CPU than the baseline does on a GPU. The broader insight the authors draw is that under domain shift, performance loss stems more from statistical misalignment in normalization layers than from representational inadequacy of the pretrained backbone.

Core claim

On the paper's own terms, the central discovery is that normalization-layer statistics carry a large share of the domain-shift signal. By forwarding target-domain samples through a frozen backbone in training mode (CNNs) and monitoring the normalized change in BatchNorm running means until it falls below a threshold, or by iteratively shifting each LayerNorm bias by the negative mean activation with a layer-wise early exit (Transformers), the backbone's feature distribution is re-centered on the target domain without any gradient update to its weights. The precomputed features are then fed to a deeper head (Linear→BatchNorm→ReLU→Linear) trained with a cross-entropy loss that amplifies ambigu

What carries the argument

The load-bearing mechanism is the decoupled two-stage pipeline: a single forward pass through the frozen backbone (after normalization-statistics adaptation) stores features once, and the classifier is trained independently on those stored features. The two adaptation rules are (1) for CNNs, an early-stopped update of BatchNorm running means using the normalized change delta = ||mu_t - mu_{t-1}||_2 / (||mu_{t-1}||_2 + epsilon) below a threshold tau_BN, and (2) for Transformers, a gradient-free LayerNorm bias update b(t) = b(t-1) - eta * E[x], with a layer-wise early exit based on the update magnitude. The redesigned head and margin-weighted loss (amplification factor alpha for samples with m

Load-bearing premise

The load-bearing premise is that re-centering normalization statistics (BatchNorm running means or LayerNorm biases) closes most of the domain gap; if the mismatch actually lives in the backbone's weights or attention structure, the no-backprop advantage collapses—and the paper itself shows the premise fails for DeiT, where bias shifting disrupts the class/distillation token balance.

What would settle it

Measure the normalized BatchNorm running-mean change delta from Eq. (1) on a domain with a known representational gap (e.g., a completely different imaging modality). If delta falls below tau_BN while accuracy remains far below fine-tuning, then delta is not a sufficient proxy for domain alignment and the central claim fails. Conversely, on a dual-token distilled transformer, disabling the LayerNorm bias update should restore the accuracy lost by the full pipeline, isolating the bias-shift as the causal mechanism.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the claim holds, any transfer-learning task can be adapted by precomputing features once, so repeated hyperparameter sweeps or long training runs cost almost nothing extra—the paper measures a speedup of up to 177x per additional configuration for DenseNet121.
  • The hardware barrier to entry drops dramatically: for all tested models except Swin, CPU training with this pipeline was faster than GPU fine-tuning or low-rank adaptation, so clinical or prototyping settings without dedicated GPUs remain viable.
  • CO2 per training run falls by an order of magnitude or more, making large-scale experimentation (many architectures, many hyperparameter sets) environmentally affordable.
  • The method is architecture-agnostic in principle, but the DeiT result shows a boundary condition: if pretraining tuned normalization layers around a delicate balance between class and distillation tokens, re-centering biases can break that balance and hurt accuracy.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable consequence the authors do not spell out: the gap between this pipeline's accuracy and full fine-tuning's accuracy can be read as an estimate of the portion of domain shift that is representational rather than statistical, turning the method into a diagnostic probe.
  • The same decoupled structure should transfer to other vision tasks—detection or segmentation heads trained on cached features—and the authors' federated-learning suggestion would let institutions share only head parameters or feature statistics, an extension they mention but do not implement.
  • The fixed amplification factor alpha and percentile threshold gamma are likely not optimal across datasets; an adaptive schedule that sets weights from the observed margin distribution is a cheap experiment that could remove manual tuning.
  • The DeiT failure suggests a targeted fix: apply the LayerNorm bias update separately to class and distillation tokens, or skip it in layers where the two-token balance is fragile—a concrete variation not tested in the paper.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper presents a transfer-learning pipeline for medical image classification that freezes the pretrained backbone and performs adaptation only on normalization layers: thresholded batch-statistics updates for CNNs (Eq. 1) and iterative bias shifting for layer-normalized Transformers (Eq. 2). Features are precomputed once and used to train a two-layer head with a margin-weighted cross-entropy loss (Eq. 5). Experiments on Brain Cancer MRI, BreakHis, and PatchCamelyon across seven architectures report accuracy that is 'competitive' with baselines while cutting training time by roughly 10-20x and CO2 emissions substantially; the authors also report that CPU training can beat GPU fine-tuning/LoRA in most cases. The broader claim is that domain-gap degradation stems more from normalization misalignment than from representational inadequacy of the pretrained backbone.

Significance. If the accuracy evidence survives scrutiny, the pipeline is a useful low-resource transfer-learning baseline, and the scalability/CO2 tables are detailed and informative. The paper also includes an ablation separating the contributions of normalization adaptation and margin weighting. However, the current evidence does not support the headline claim that full backbone backpropagation yields only marginal gains, because the CNN baseline is not full fine-tuning. The absence of repeated runs and the undescribed BreakHis split further weaken the accuracy evidence. The contribution is still meaningful as an efficient linear-probing-style adaptation method, but the central accuracy/cost claim must be re-established against the appropriate baseline.

major comments (3)
  1. [§4.2, Tables 2-4] The CNN 'traditional fine-tuning' baseline is defined as training only the last layer while the backbone is frozen, i.e., linear probing. Section 1, however, claims that 'the marginal gains of full backbone backpropagation are outweighed by its significant computational overhead,' and Tables 2-4 label this baseline 'Fine-Tuning.' No full-backbone fine-tuning experiment is reported, so the accuracy/cost trade-off is established only against a linear probe, not against the method the text criticizes. On PatchCamelyon, standard full fine-tuning of these backbones commonly reports markedly higher accuracy than the 79-84% shown in Table 4, so the 'marginal accuracy trade-off' is unsupported. Please add at least one full-fine-tuning baseline per CNN architecture (or per representative model) and compare both accuracy and training time, or revise the claims to be explicitly relative to linear p
  2. [Tables 2-4] All accuracies are single runs with no error bars, seeds, or statistical tests. Several conclusions rest on differences below one percentage point (e.g., ResNet18 on BreakHis 99.33 vs 98.50; MobileNetV3 on PatchCamelyon 82.96 vs 82.64), and the claim of 'averaging 1.6% lower' on PatchCamelyon is based on one run per model. On small datasets such as BreakHis and Brain Cancer MRI, random split variation can easily exceed these differences. Please report mean and standard deviation over at least 3-5 seeds and indicate whether observed differences are statistically significant; if this is not possible, the accuracy claims must be correspondingly softened.
  3. [§4.1.2, BreakHis] The train/validation/test split for BreakHis is not described. BreakHis contains 82 patients, and the standard protocol is a patient-exclusive split; an image-level random split can place images from the same patient in both training and test, inflating accuracy through patient leakage. Given that most models exceed 98% on this dataset, please specify the split procedure and use patient-level partition (or justify why image-level partition is appropriate for the clinical claim).
minor comments (6)
  1. [§5.1, MobileNetV3 on BreakHis] The sentence about following the protocol of [45] reports 40x accuracy of 95.59% for MobileNetV3, which conflicts with the 98.67% for MobileNetV3 in Table 2. Clarify whether this is due to a different image resolution (672x448 vs 256x256), split, or experimental setting.
  2. [§5.2, Table 6] The 'speedup of 177x' is based on linear extrapolation of regression fits to training times. State explicitly that this assumes the linear model holds beyond the measured hyperparameter range, or add measurements for P=160 rather than extrapolation.
  3. [Figure 6] The figure caption is too sparse: it should state what is plotted (wall-clock time, total training time, or per-epoch time), which hardware is used for each curve, and whether the baseline includes repeated backbone forward passes. This is essential for interpreting the 'CPU faster than GPU' claim.
  4. [General] There is no code or data availability statement. Given the empirical nature of the paper and the sensitivity of the accuracy claims to split and hyperparameters, a public implementation would greatly aid reproducibility.
  5. [Abstract and Tables] The abstract says the method 'often matches or surpasses baseline performance,' but in Table 4 the CNN results are consistently 0.7-3.5 points below the baseline. Please make the abstract and conclusions more precise about the accuracy trade-off.
  6. [§4.1.3] Typo: 'histopatology' should be 'histopathology.'

Circularity Check

0 steps flagged

No significant circularity: all components are a priori procedures and results come from held-out benchmarks.

full rationale

The proposed pipeline consists of fixed adaptation procedures (Eq. 1 for BN statistics, Eq. 2 for LN bias, Eqs. 4-5 for margin weighting) that are not fitted to the test labels and are not derived from the accuracies they are later compared with. Hyperparameters (alpha=10, gamma=20th percentile, tau_BN=0.01, eta=0.1, tau_LN=0.0001) are fixed across all runs and datasets. The ablation study (Table 7) tests each component in isolation, and the DeiT failure (Section 5.1) shows the method can be falsified, so the normalization-centric claim is not protected from data. The only significant weakness is experimental: the CNN baseline is last-layer-only adaptation, explicitly stated in Section 4.2, so the claim that full backbone backpropagation yields marginal gains is not fully supported. That is a comparison validity issue rather than circularity: no equation reduces to an input, and no load-bearing self-citation chain was found.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The paper adds no fundamentally new physical or mathematical entity. Its central claim rests on unproven statistical assumptions: normalization statistics dominate domain shift; updating BN running means without labels is safe; LN bias shifts by negative activation mean are safe; small-margin samples are the right ones to up-weight. These are empirical domain assumptions, not derived results. The fixed hyperparameters (α, γ, τ_BN, η, t_max, τ_LN, 512) are free choices that affect all reported numbers.

free parameters (7)
  • amplification factor α = 10.0
    Weight assigned to ambiguous samples in Eq. (5); set by hand and used in all runs.
  • margin threshold γ = 20th percentile of batch margins
    Defines which samples are 'ambiguous'; chosen heuristic.
  • BN convergence threshold τ_BN = 0.01
    Early-stopping rule for BN statistics adaptation, Eq. (1).
  • LN adaptation rate η = 0.1
    Step size in bias update, Eq. (2).
  • max LN adaptation steps t_max = 20
    Stopping mechanism for LN bias adaptation.
  • LN layer convergence threshold τ_LN = 0.0001
    Layer-wise early exit for LN adaptation, Eq. (3).
  • head hidden size = 512
    Width of the first Linear layer in the redesigned classifier head (Section 3.3); affects capacity and compute.
axioms (4)
  • domain assumption Domain shift is primarily statistical misalignment in normalization layers, not representational inadequacy.
    Central motivation, Section 1; if false, frozen-backbone + normalization adaptation cannot close the gap.
  • domain assumption BN running statistics can be updated without labels or gradients during forward passes and this improves domain alignment.
    Backbone adaptation for CNNs, Section 3.2.1, relies on this.
  • ad hoc to paper The bias update b(t) = b(t-1) - η E[x] shifts LN outputs toward target distribution without hurting attention.
    Eq. (2), Section 3.2.2; no theoretical derivation; DeiT failure (Section 5.1) shows it is not universally safe.
  • domain assumption Samples with small prediction margin are the most instructive; reweighting them by α improves generalization.
    Margin-based weighted loss, Eq. (5); plausible but no proof or comparison to other reweighting schemes.

pith-pipeline@v1.3.0-alltime-deepseek · 14865 in / 15932 out tokens · 163880 ms · 2026-08-02T11:20:14.044747+00:00 · methodology

0 comments
read the original abstract

Deep learning models achieve state-of-the-art image classification but face deployment challenges due to computational costs and energy demands. We propose a lightweight training strategy that adapts normalization layers of the model to the new domain and decouples feature extraction from classifier optimization, reducing overhead by precomputing features only once. A redesigned classifier head with margin-based weighted loss further minimizes ambiguity without end-to-end backpropagation. Evaluated across four CNN architectures (ResNet18, ResNet50, MobileNet, DenseNet121), three Transformer models (ViT, Swin and DeiT) and three medical datasets (Brain Cancer MRI, BreakHis and PatchCamelyon), our approach significantly reduces the required training time with only a marginal accuracy trade-off, often matching or surpassing baseline performance. This efficiency translates to reducing CO2 by orders of magnitude, offering a practical and environmentally sustainable solution for resource-constrained clinical or prototyping environments.

Figures

Figures reproduced from arXiv: 2607.13043 by Daniel Vila-Cruz, Laura Mor\'an-Fern\'andez, Ver\'onica Bol\'on-Canedo.

Figure 1
Figure 1. Figure 1: Proposed classifier head, where C denotes the number of output classes. The inclusion of BN and ReLU improves feature conditioning and encourages non-linearity, which can enhance generalization under domain shift. Despite the 13 [PITH_FULL_IMAGE:figures/full_fig_p014_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Samples from the Brain Cancer MRI dataset [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Samples from the BreakHis dataset 4.1.3. PCAM The PatchCamelyon (PCAM) dataset [44] is a large-scale histopatology bench￾mark consisting of 327,670 images (originally 96 × 96 and resized to 224 × 224). Representative samples from this dataset are illustrated in [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Samples from the PatchCamelyon dataset 4.2. Models We evaluate our proposed framework across two dominant architectural para￾digms to assess the efficiency-to-performance ratio of our approach. To provide a rigorous benchmark, we compare our method against the standard adaptation techniques for each architecture type: 18 [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Estimated time required (in seconds) to perform hyper parameter tuning using di [PITH_FULL_IMAGE:figures/full_fig_p029_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Training comparison using CPU and GPU. 28 [PITH_FULL_IMAGE:figures/full_fig_p029_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 7 linked inside Pith

  1. [1]

    Schwartz, J

    R. Schwartz, J. Dodge, N. A. Smith, O. Etzioni, Green ai, Communications of the ACM 63 (12) (2020) 54–63

    2020

  2. [2]

    Osonuga, A

    A. Osonuga, A. A. Osonuga, S. C. Fidelis, G. C. Osonuga, J. Juckes, D. B. Olawade, Bridging the digital divide: artificial intelligence as a catalyst for health equity in primary care settings, International Journal of Medical In- formatics (2025) 106051. 32

    2025

  3. [3]

    R. Poh, B. W. Tan, H. W. Lee, K. L. Tan, Bridging the digital divide: Socioe- conomic equity in ai-assisted hybrid learning, SN Computer Science 7 (1) (2025) 50

    2025

  4. [4]

    Liang, J

    T. Liang, J. Glossner, L. Wang, S. Shi, X. Zhang, Pruning and quantization for deep neural network acceleration: A survey, Neurocomputing 461 (2021) 370–403

    2021

  5. [5]

    Gholami, S

    A. Gholami, S. Kim, Z. Dong, Z. Yao, M. W. Mahoney, K. Keutzer, A sur- vey of quantization methods for efficient neural network inference, in: Low- power computer vision, Chapman and Hall/CRC, 2022, pp. 291–326

    2022

  6. [6]

    J. Gou, B. Yu, S. J. Maybank, D. Tao, Knowledge distillation: A survey, International journal of computer vision 129 (6) (2021) 1789–1819

    2021

  7. [7]

    D. C. Lepcha, B. Goyal, A. Dogra, A. Alkhayyat, P. K. Sahu, A. Ali, V . Kukreja, Deep learning in medical image analysis: A comprehensive re- view of algorithms, trends, applications, and challenges, Computer Model- ing in Engineering & Sciences 145 (2) (2025) 1487

    2025

  8. [8]

    J. R. Zech, M. A. Badgeley, M. Liu, A. B. Costa, J. J. Titano, E. K. Oermann, Variable generalization performance of a deep learning model to detect pneu- monia in chest radiographs: a cross-sectional study, PLoS medicine 15 (11) (2018) e1002683

    2018

  9. [9]

    Oakden-Rayner, J

    L. Oakden-Rayner, J. Dunnmon, G. Carneiro, C. Ré, Hidden stratification causes clinically meaningful failures in machine learning for medical imag- ing, in: Proceedings of the ACM conference on health, inference, and learn- ing, 2020, pp. 151–159. 33

    2020

  10. [10]

    Strubell, A

    E. Strubell, A. Ganesh, A. McCallum, Energy and policy considerations for modern deep learning research, in: Proceedings of the AAAI conference on artificial intelligence, V ol. 34, 2020, pp. 13693–13696

    2020

  11. [11]

    Verdecchia, J

    R. Verdecchia, J. Sallou, L. Cruz, A systematic review of green ai, Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 13 (4) (2023) e1507

    2023

  12. [12]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recogni- tion, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  13. [13]

    Huang, Z

    G. Huang, Z. Liu, L. Van Der Maaten, K. Q. Weinberger, Densely connected convolutional networks, in: Proceedings of the IEEE conference on com- puter vision and pattern recognition, 2017, pp. 4700–4708

    2017

  14. [14]

    A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, H. Adam, Mobilenets: Efficient convolutional neural net- works for mobile vision applications, arXiv preprint arXiv:1704.04861 (2017)

    Pith/arXiv arXiv 2017

  15. [15]

    M. Tan, Q. Le, Efficientnet: Rethinking model scaling for convolutional neural networks, in: International conference on machine learning, PMLR, 2019, pp. 6105–6114

    2019

  16. [16]

    Simonyan, A

    K. Simonyan, A. Zisserman, Very deep convolutional networks for large- scale image recognition, arXiv preprint arXiv:1409.1556 (2014)

    Pith/arXiv arXiv 2014

  17. [17]

    Ronneberger, P

    O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, in: International Conference on Medical 34 image computing and computer-assisted intervention, Springer, 2015, pp. 234–241

    2015

  18. [18]

    H. Laçi, K. Sevrani, S. Iqbal, Deep learning approaches for classification tasks in medical x-ray, mri, and ultrasound images: a scoping review, BMC Medical Imaging 25 (1) (2025) 156

    2025

  19. [19]

    Kumar, P

    R. Kumar, P. Kumbharkar, S. Vanam, S. Sharma, Medical images classifica- tion using deep learning: a survey, Multimedia Tools and Applications 83 (7) (2024) 19683–19728

    2024

  20. [20]

    Dosovitskiy, An image is worth 16x16 words: Transformers for image recognition at scale, arXiv preprint arXiv:2010.11929 (2020)

    A. Dosovitskiy, An image is worth 16x16 words: Transformers for image recognition at scale, arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2010

  21. [21]

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, B. Guo, Swin trans- former: Hierarchical vision transformer using shifted windows, in: Proceed- ings of the IEEE/CVF international conference on computer vision, 2021, pp. 10012–10022

    2021

  22. [22]

    Touvron, M

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, H. Jégou, Training data-efficient image transformers & distillation through attention, in: International conference on machine learning, PMLR, 2021, pp. 10347– 10357

    2021

  23. [23]

    Shobayo, R

    O. Shobayo, R. Saatchi, Developments in deep learning artificial neural net- work techniques for medical image analysis and interpretation, Diagnostics 15 (9) (2025) 1072

    2025

  24. [24]

    T. C. Nauen, S. Palacio, F. Raue, A. Dengel, Which transformer to fa- vor: A comparative analysis of efficiency in vision transformers, in: 35 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV), IEEE, 2025, pp. 6955–6966

    2025

  25. [25]

    Menghani, Efficient deep learning: A survey on making deep learning models smaller, faster, and better, ACM Computing Surveys 55 (12) (2023) 1–37

    G. Menghani, Efficient deep learning: A survey on making deep learning models smaller, faster, and better, ACM Computing Surveys 55 (12) (2023) 1–37

    2023

  26. [26]

    Raghu, C

    M. Raghu, C. Zhang, J. Kleinberg, S. Bengio, Transfusion: Understanding transfer learning for medical imaging, Advances in neural information pro- cessing systems 32 (2019)

    2019

  27. [27]

    Gholizade, H

    M. Gholizade, H. Soltanizadeh, M. Rahmanimanesh, S. S. Sana, A review of recent advances and strategies in transfer learning, International Journal of System Assurance Engineering and Management (2025) 1–40

    2025

  28. [28]

    Z. Han, C. Gao, J. Liu, J. Zhang, S. Q. Zhang, Parameter-efficient fine-tuning for large models: A comprehensive survey, arXiv preprint arXiv:2403.14608 (2024)

    Pith/arXiv arXiv 2024

  29. [29]

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen, et al., Lora: Low-rank adaptation of large language models., ICLR 1 (2) (2022) 3

    2022

  30. [30]

    E. B. Zaken, Y . Goldberg, S. Ravfogel, Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models, in: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 2: Short Papers), 2022, pp. 1–9

    2022

  31. [31]

    M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, S.-N. 36 Lim, Visual prompt tuning, in: European conference on computer vision, Springer, 2022, pp. 709–727

    2022

  32. [32]

    Wilson, D

    G. Wilson, D. J. Cook, A survey of unsupervised deep domain adaptation, ACM Transactions on Intelligent Systems and Technology (TIST) 11 (5) (2020) 1–46

    2020

  33. [33]

    W.-N. Hsu, A. Sriram, A. Baevski, T. Likhomanenko, Q. Xu, V . Pratap, J. Kahn, A. Lee, R. Collobert, G. Synnaeve, et al., Robust wav2vec 2.0: Analyzing domain shift in self-supervised pre-training, arXiv preprint arXiv:2104.01027 (2021)

    Pith/arXiv arXiv 2021

  34. [34]

    Liang, R

    J. Liang, R. He, T. Tan, A comprehensive survey on test-time adaptation under distribution shifts, International Journal of Computer Vision 133 (1) (2025) 31–64

    2025

  35. [35]

    J. O. Neill, An overview of neural network compression, arXiv preprint arXiv:2006.03669 (2020)

    Pith/arXiv arXiv 2006

  36. [36]

    Y . He, L. Xiao, Structured pruning for deep convolutional neural networks: A survey, IEEE transactions on pattern analysis and machine intelligence 46 (5) (2023) 2900–2919

    2023

  37. [37]

    K. He, J. Sun, Convolutional neural networks at constrained time cost, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015

    2015

  38. [38]

    H. Li, Z. Wang, X. Yue, W. Wang, H. Tomiyama, L. Meng, An architecture- level analysis on deep learning models for low-impact computations, Artifi- cial Intelligence Review 56 (3) (2023) 1971–2010. 37

    2023

  39. [39]

    Y . Li, N. Wang, J. Shi, X. Hou, J. Liu, Adaptive batch normalization for practical domain adaptation, Pattern Recognition 80 (2018) 109–117

    2018

  40. [40]

    Patel, P

    D. Patel, P. Timsina, L. Gorenstein, B. S. Glicksberg, G. Raut, S. N. Cheeti- rala, F. Santana, J. Tamegue, A. Kia, E. Zimlichman, et al., Traditional ma- chine learning, deep learning, and bert (large language model) approaches for predicting hospitalizations from nurse triage notes: Comparative evalua- tion of resource management, JMIR AI 3 (1) (2024) e52190

    2024

  41. [41]

    Miotto, F

    R. Miotto, F. Wang, S. Wang, X. Jiang, J. T. Dudley, Deep learning for healthcare: review, opportunities and challenges, Briefings in bioinformatics 19 (6) (2018) 1236–1246

    2018

  42. [42]

    M. M. Rahman, Brain Cancer - MRI dataset (2024). doi:10.17632/mk56jw9rns.1. URLhttps://data.mendeley.com/datasets/mk56jw9rns/1

    doi:10.17632/mk56jw9rns.1 2024

  43. [43]

    F. A. Spanhol, L. S. Oliveira, C. Petitjean, L. Heutte, A dataset for breast can- cer histopathological image classification, IEEE Transactions on Biomedical Engineering 63 (7) (2016) 1455–1462. doi:10.1109/TBME.2015.2496264

    arXiv 2016

  44. [44]

    B. S. Veeling, J. Linmans, J. Winkens, T. Cohen, M. Welling, Rotation equiv- ariant cnns for digital pathology (2018). arXiv:1806.03962. URLhttps://arxiv.org/abs/1806.03962

    Pith/arXiv arXiv 2018

  45. [45]

    M. Saha, M. Chakraborty, S. Maiti, D. Das, Breast-NET: A lightweight DCNN model for breast cancer detection and grading using histological sam- ples, Neural Computing and Applications 36 (32) (2024) 20067–20087. 38

    2024