Pith. sign in

REVIEW 4 major objections 5 minor 33 references

The paper demonstrates that a cooperative meta-learning framework can identify individual cattle from a handful of muzzle images, with a reported best accuracy of 98.46% and F1 score of 97.91% in a 5-way 5-shot setting.

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

2026-08-04 16:52 UTC pith:SDD2TSS4

load-bearing objection Useful benchmark, but the headline numbers are tuned on the test set and the abstract cherry-picks the better split; the method itself is a minor variant of CML. the 4 major comments →

arxiv 2509.11219 v1 pith:SDD2TSS4 submitted 2025-09-14 cs.CV

CCoMAML: Efficient Cattle Identification Using Cooperative Model-Agnostic Meta-Learning

Rabin Dulal , Lihong Zheng , Ashad Kabir This is my paper
classification cs.CV
keywords cattle identificationmuzzle biometricsfew-shot learningmodel-agnostic meta-learninggradient noisemulti-head attention feature fusioncooperative meta-learninglivestock management
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 aims to show that few-shot meta-learning can make cattle identification practical with only a handful of muzzle images per animal, avoiding full retraining when herd composition changes. It proposes CCoMAML, which augments cooperative MAML with a small CNN co-learner that injects learnable noise into outer-loop gradient updates, paired with a multi-head attention feature fusion backbone. Across two cattle muzzle datasets and two benchmark datasets, the method is reported to outperform 14 few-shot learning baselines, reaching 98.46% accuracy and 97.91% F1 in the 5-way 5-shot setting. The practical payoff, if correct, is that cameras can replace RFID ear tags for dynamic cattle traceability with minimal data collection.

Core claim

The paper's central claim is that CCoMAML, a cooperative meta-learning method with a custom CNN co-learner, achieves top cattle identification performance from few muzzle images. The co-learner modifies gradient updates during meta-training by adding learnable noise, which the paper argues regularizes learning and leads to a meta-initialization that adapts quickly to new cattle with one to five images per animal. Evaluated against 14 few-shot learning methods, CCoMAML reports the highest accuracy and F1 under both 5-way 1-shot and 5-way 5-shot settings on cattle muzzle data, with a top accuracy of 98.46% and F1 of 97.91%. The paper also shows that CCoMAML outperforms conventional deep learni

What carries the argument

The central mechanism is the co-learner: a small CNN with two convolutional layers, ReLU activations, adaptive average pooling, and two fully connected layers. It operates only during meta-training, taking features from the base learner and producing a 64-dimensional noise vector that augments the outer-loop gradient. A scalar gamma controls the strength of this noise, and the paper selects gamma=0.2 after ablation. The base feature extractor is MHAFF, which fuses CNN and Vision Transformer features through multi-head cross-attention. Because the co-learner is discarded at test time, the method adds no inference cost, a property the paper emphasizes.

Load-bearing premise

The load-bearing assumption is that the co-learner intensity (gamma=0.2) and co-learner architecture were chosen without using the final test data, so the reported accuracies are unbiased estimates of real-world generalization.

What would settle it

Repeat the 5-way 5-shot cattle evaluation with gamma and co-learner depth selected on a held-out meta-validation split (e.g., classes disjoint from both training and test sets) rather than on the test data; if accuracy drops materially below the reported 96–98% range, or if the gap over CML vanishes, the claimed co-learner benefit is an artifact of test-set tuning.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the claim holds, new cattle can be enrolled with as few as one to five muzzle images, and existing cows can be re-identified without retraining the model from scratch.
  • The cross-dataset experiments imply that a model meta-trained on one farm's muzzle images can adapt to animals from another farm or dataset with high accuracy, which would make farm-to-farm deployment feasible.
  • The method's advantage is largest in low-shot regimes: with three training images per class, CCoMAML reportedly stays above 90% accuracy while conventional transfer-learned deep models fall below 75%.
  • The co-learner's gradient noise acts as a task-level regularizer, and because it is removed at inference, the framework could be dropped into existing MAML-based identification systems without extra per-prediction compute.
  • Validation on CIFAR10 and Flower102 suggests the same meta-learning recipe transfers beyond cattle to other fine-grained classification problems with limited labeled data.

Where Pith is reading between the lines

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

  • A reader could test whether the reported advantage over plain CML is robust by choosing gamma and the co-learner architecture on a separate validation split of training classes, rather than on the test set; if the gap shrinks, the headline numbers are partly an artifact of test-set tuning.
  • Because the paper's own N-way results show accuracy declining as N increases, a practical herd-scale system might need hierarchical classification or class-subset strategies to maintain the reported accuracy when the number of animals is large.
  • The same co-learner design could be applied to other fine-grained biometric identification tasks, such as ear or coat-pattern recognition in other livestock, where data is scarce and herd membership is dynamic.

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

4 major / 5 minor

Summary. The paper proposes CCoMAML, a few-shot cattle identification method that combines Cooperative MAML (CML) with a Multi-Head Attention Feature Fusion (MHAFF) backbone and a custom CNN co-learner that injects learnable gradient noise in the outer loop. The method is evaluated on two cattle muzzle datasets (UNE, UNL) and two benchmark datasets (CIFAR10, Flower102) against 14 existing few-shot learning methods. The paper claims top performance of 98.46% accuracy and 97.91% F1-score in the 5-way 5-shot setting on muzzle data, and argues that the method avoids full retraining when herd composition changes.

Significance. If the reported results hold under a clean evaluation protocol, the work would be a useful application-oriented contribution: it would demonstrate that meta-learning with an attention-fused CNN/transformer backbone can identify individual cattle from very few muzzle images, and it would provide a systematic FSL benchmark for muzzle-based cattle identification. The use of publicly available datasets, the large number of compared FSL methods, and the focus on a practically motivated problem are strengths. However, the central performance claims are currently undermined by evaluation-protocol issues: hyperparameters and co-learner intensity appear to be selected on the test set, and the abstract/conclusion report the better of two cross-dataset directions rather than the main comparison table. The contribution relative to CML is also incremental, but the empirical benchmark could still be valuable once the protocol is fixed.

major comments (4)
  1. [§6.2.1, Tables 6–8, and §6.2.2, Fig. 5] The co-learner architecture and the co-learner intensity γ are selected using the same target data that is later reported as the final test performance. For example, Table 8 shows that 2 convolutional layers gives 87.32% (1-shot) and 96.07% (5-shot), and Fig. 5 shows that γ=0.2 gives exactly the same values, including F1=97.03% in Table 4. No separate validation split is described; the 'validation loss' mentioned in §6 is not tied to a defined validation set. This means the headline numbers are test-set-selected and are optimistic estimates of generalization. A proper held-out validation set, or nested cross-validation, is needed before reporting final test accuracy/F1.
  2. [Abstract, §8, Table 4, Fig. 7] The abstract and conclusion advertise 98.46% accuracy and 97.91% F1-score, but the main comparison on the primary split (UNE train → UNL test, §5.1.1 and Table 4) reports 96.07% accuracy and 97.03% F1 for 5-shot. The headline numbers come from the reversed split (UNL train → UNE test) in Fig. 7, i.e., the better of the two cross-dataset directions. Reporting only the best direction without averaging or disclosure inflates the claimed performance. Both directions should be reported and the primary result should be stated consistently.
  3. [§7 and Fig. 8] The comparison between CCoMAML and the deep learning baselines is not protocol-equivalent. MHAFF, ViT, and MobileNetV3 are trained on K examples per class over the full class set (65 UNE classes / 227 UNL classes), whereas CCoMAML is evaluated in 5-way episodic tasks. The number of classes and the task structure differ, so the claim of a 'fair and consistent comparison' is unsupported. The authors should either evaluate the DL baselines under the same N-way K-shot protocol, evaluate CCoMAML in a closed-set setting over all classes, or clearly frame Fig. 8 as an approximate comparison and state the limitation.
  4. [§5.2 and Table 4] The selection of CML as the base FSL method is itself based on performance on the same split that is later used to evaluate CCoMAML. This is an additional source of test-set dependence: the 'best' baseline was chosen by looking at the test results, and CCoMAML is then constructed to outperform it on the same data. A validation split should be used for method selection, or the selection step should be described as exploratory with the final comparison repeated on a truly held-out split.
minor comments (5)
  1. [Figures 5–8] All four figures have the same caption 'Figure 1' in the rendered text; the captions need to be relabeled.
  2. [§6.1] The text says results are 'over three runs' but the table captions say 'Results are averaged over 5 runs.' This inconsistency should be fixed.
  3. [§6.2 heading] Typo: 'Abalation' should be 'Ablation'.
  4. [Table 6] The 5-shot value for 2 FC layers is printed as '91.932', which appears to be a typo for 91.93. Also, the 4-layer row repeats 93.98; clarify whether this is exact or rounded.
  5. [§5.3.2 and §6] The text states that the co-learner is removed at meta-testing, but it is not explicit whether the hyperparameters chosen in the ablation (e.g., γ=0.2) are also removed during inference or only during meta-training. Clarify this in the methodology.

Circularity Check

2 steps flagged

Headline numbers are test-selected: co-learner architecture and γ are tuned on the same evaluation data, then reported as CCoMAML's performance; the 98.46%/97.91% headline also cherry-picks the better cross-dataset direction.

specific steps
  1. fitted input called prediction [Section 6.2.1 (Tables 6-8)]
    "The accuracies of 5-way 1-shot and 5-shot experiments were evaluated to determine the optimal number of FC layers. ... As shown in Table 8, introducing two convolutional layers yielded the highest accuracy in both 1-shot (87.32%) and 5-shot (96.07%) settings."

    The co-learner architecture (FC count, pooling/conv strategy, number of conv layers) is selected by maximizing 5-way 1-shot and 5-shot accuracy on the cattle muzzle evaluation. The winning configuration's accuracies (87.32%, 96.07%) are exactly the numbers later reported as CCoMAML's result in Table 4 (87.32±0.31%, 96.07±0.28%). Thus the reported 'performance' is the objective function used for architecture search, not an independent evaluation. No separate validation set is described for these choices.

  2. fitted input called prediction [Section 6.2.2 (Fig. 5)]
    "Different values for γ were used to study the noise intensity. The used values were 0.2, 0.4, 0.6, 0.8, and 1.0. The performance of the effect of γ is shown in Fig. 5. ... When γ is set to 0.2, the model achieves the highest accuracy and F1 scores in 5-way 1-shot and 5-way 5-shot settings."

    The co-learner intensity γ is chosen from the same accuracy/F1 curves that later serve as the final evidence. Figure 5 at γ=0.2 shows 87.32/96.07 accuracy and 87.46/97.03 F1, identical to Table 4's CCoMAML row. Selecting γ by the best score on the evaluation data and then presenting that score as proof of superiority is a fitted-input-called-prediction loop; with no separate validation set, the headline numbers incorporate selection bias.

full rationale

The central circularity is evaluation leakage presented as prediction: both the co-learner architecture (§6.2.1, Tables 6-8) and the noise intensity γ (§6.2.2, Fig. 5) are tuned on the same 5-way 1-shot/5-shot cattle muzzle test tasks whose results are then reported as CCoMAML's performance in Table 4 and the abstract. This makes the headline 98.46%/97.91% (taken from the better reversed split in Fig. 7) a product of fitting to the evaluation data, so the claimed superiority over 14 FSL methods is not yet supported by an unbiased held-out test. The reversed-split selection is a reporting/statistical concern rather than a derivation loop. The core update equations (Eqs. 3-6) are standard MAML/CML and are not definitionally circular; self-citations to MHAFF are not load-bearing because the same backbone is used for all compared methods, and CIFAR10/Flower102 results provide some external check. However, the main cattle-identification claim rests on test-selected hyperparameters, warranting a partial-circularity score of 6.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim rests on several hand-selected hyperparameters and the inherited validity of prior domain assumptions. The co-learner architecture and noise intensity were tuned on the test data, which is the most significant ledger item.

free parameters (6)
  • gamma (co-learner intensity) = 0.2
    Tuned via ablation on the target muzzle test set (Fig. 5); controls noise in the outer-loop update.
  • Number of FC layers in co-learner = 2
    Selected from Table 6 based on 5-way 1-shot and 5-shot accuracy on the target test data.
  • Number of conv layers in co-learner = 2
    Selected from Table 8 based on accuracy on the target test data.
  • Inner-loop learning rate alpha = 1e-1
    Stated as set in Section 6; not subjected to sensitivity analysis.
  • Meta-learning rate beta = 1e-3
    Stated as set in Section 6; hand-chosen.
  • Weight decay = 1e-5
    Stated as set in Section 6; hand-chosen.
axioms (5)
  • domain assumption Muzzle patterns are unique per individual and stable over time
    Inherited from earlier literature (Petersen 1922, Awad 2016); not revalidated here.
  • domain assumption Episodic few-shot training with disjoint classes simulates real-world dynamic herds
    Stated in Section 5.1.1; necessary for the evaluation protocol.
  • domain assumption Training on UNE and testing on UNL (or reverse) is representative of deployment
    Assumed in Section 6.3 for cross-dataset generalization.
  • ad hoc to paper The CNN co-learner's learnable gradient noise improves meta-generalization
    Adopted from CML (Shin et al. 2024) with a redesigned architecture; no independent theoretical justification beyond prior work.
  • ad hoc to paper MHAFF is an appropriate base model for all FSL methods
    Borrowed from the authors' previous paper (Dulal et al. 2025); not compared with alternative backbones in this study.

pith-pipeline@v1.3.0-alltime-deepseek · 27533 in / 12628 out tokens · 132311 ms · 2026-08-04T16:52:12.473048+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of CCoMAML: Efficient Cattle Identification Using Cooperative Model-Agnostic Meta-Learning." pith.science (2026). https://pith.science/paper/SDD2TSS4

@misc{pith2026250911219,
  author       = {Pith},
  title        = {Pith review of: CCoMAML: Efficient Cattle Identification Using Cooperative Model-Agnostic Meta-Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SDD2TSS4}},
  note         = {Machine review of arXiv:2509.11219}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Cattle identification is critical for efficient livestock farming management, currently reliant on radio-frequency identification (RFID) ear tags. However, RFID-based systems are prone to failure due to loss, damage, tampering, and vulnerability to external attacks. As a robust alternative, biometric identification using cattle muzzle patterns similar to human fingerprints has emerged as a promising solution. Deep learning techniques have demonstrated success in leveraging these unique patterns for accurate identification. But deep learning models face significant challenges, including limited data availability, disruptions during data collection, and dynamic herd compositions that require frequent model retraining. To address these limitations, this paper proposes a novel few-shot learning framework for real-time cattle identification using Cooperative Model-Agnostic Meta-Learning (CCoMAML) with Multi-Head Attention Feature Fusion (MHAFF) as a feature extractor model. This model offers great model adaptability to new data through efficient learning from few data samples without retraining. The proposed approach has been rigorously evaluated against current state-of-the-art few-shot learning techniques applied in cattle identification. Comprehensive experimental results demonstrate that our proposed CCoMAML with MHAFF has superior cattle identification performance with 98.46% and 97.91% F1 scores.

Figures

Figures reproduced from arXiv: 2509.11219 by Ashad Kabir, Lihong Zheng, Rabin Dulal.

Figure 1
Figure 1. Figure 1: An overview of the methodology of this research. [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Representative samples from each dataset, showcasing sample images from di [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: An overview of the CCoMAML method. The upper light green color represents the outer loop, and the lower light blue color represents [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Architecture of the proposed custom CNN-based co-learner module. The co-learner takes feature representations as input and leverages [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 1
Figure 1. Figure 1: Effect of co-learner intensity (γ) on CCoMAML’s performance in 1-shot and 5-shot settings. References Preprint submitted to Elsevier September 14, 2025 [PITH_FULL_IMAGE:figures/full_fig_p024_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: N-way K-shot performance of CCoMAML using UNE muzzle as training and UNL muzzle as testing dataset. Figure 6: N-way K-shot performance of CCoMAML using UNE muzzle as training and UNL muzzle as testing dataset. [PITH_FULL_IMAGE:figures/full_fig_p025_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: Performance of CCoMAML on N-way K-shot classification using UNL muzzle as training data and UNE muzzle as testing data. Figure 7: Performance of CCoMAML on N-way K-shot classification using UNL muzzle as training data and UNE muzzle as testing data. [PITH_FULL_IMAGE:figures/full_fig_p026_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: Comparison of accuracy vs. number of training images for UNL and UNE muzzle data across diff Figure 8: Comparison of accuracy vs. number of training images for UNL and UNE muzzle data across di [PITH_FULL_IMAGE:figures/full_fig_p027_1.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

33 extracted references · 21 linked inside Pith

  1. [4]

    Research Track: European Conference, ECML PKDD 2021, Bilbao, Spain, September 13–17, 2021, Proceedings, Part I 21, Springer

    Bridging few-shot learning and adaptation: New challenges of support- query shift, in: Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2021, Bilbao, Spain, September 13–17, 2021, Proceedings, Part I 21, Springer. pp. 554–569. Bergman, N., Yitzhaky, Y ., Halachmi, I.,

    2021

  2. [7]

    Lightweight cow face recognition algorithm based on few-shot learning for edge computing application, in: 2021 ASABE Annual International Virtual Meeting, American Society of Agricultural and Biological Engineers. p

    2021

  3. [9]

    arXiv preprint arXiv:2010.11929

    An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 . Du, Y ., Kou, Y ., Li, B., Qin, L., Gao, D.,

    Pith/arXiv arXiv 2010

  4. [12]

    arXiv preprint arXiv:2104.02638

    Comparing transfer and meta learning approaches on a unified few-shot classification benchmark. arXiv preprint arXiv:2104.02638 . Finn, C., Abbeel, P., Levine, S.,

    Pith/arXiv arXiv

  5. [13]

    A broader study of cross-domain few- shot learning, in: Computer vision–ECCV 2020: 16th European conference, glasgow, UK, August 23–28, 2020, proceedings, part XXVII 16, Springer. pp. 124–141. Guo, Y ., Hong, W., Wu, J., Huang, X., Qiao, Y ., Kong, H.,

    2020

  6. [14]

    arXiv preprint arXiv:1703.07737

    In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 . Hinton, G.E., Roweis, S.,

    Pith/arXiv arXiv

  7. [17]

    arXiv preprint arXiv:1707.09835

    Meta-sgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835 . Liu, H., Reibman, A.R., Boerman, J.P.,

    Pith/arXiv arXiv

  8. [19]

    arXiv preprint arXiv:1707.03141

    A simple neural attentive meta-learner. arXiv preprint arXiv:1707.03141 . Mogan, J.N., Lee, C.P., Lim, K.M.,

    Pith/arXiv arXiv

  9. [20]

    arXiv preprint arXiv:1511.06807

    Adding gradient noise improves learning for very deep networks. arXiv preprint arXiv:1511.06807 . Nichol, A., Schulman, J.,

    Pith/arXiv arXiv

  10. [21]

    arXiv preprint arXiv:1803.02999 2,

    Reptile: a scalable metalearning algorithm. arXiv preprint arXiv:1803.02999 2,

    Pith/arXiv arXiv

  11. [23]

    (accessed: 15.01.2023)

    How does the nlis work? https://www.integritysystems.com.au/identification–traceability/national-livestock-identification-system/. (accessed: 15.01.2023). Oh, J., Yoo, H., Kim, C., Yun, S.Y .,

    2023

  12. [24]

    arXiv preprint arXiv:2008.08882

    Boil: Towards representation change for few-shot learning. arXiv preprint arXiv:2008.08882 . Ojala, T., Pietikainen, M., Harwood, D.,

    Pith/arXiv arXiv 2008

  13. [25]

    arXiv preprint arXiv:1909.09157

    Rapid learning or feature reuse? towards understanding the effectiveness of maml. arXiv preprint arXiv:1909.09157 . Ren, K., Bernes, G., Hetta, M., Karlsson, J.,

    Pith/arXiv arXiv 1909

  14. [26]

    arXiv preprint arXiv:1807.05960

    Meta-learning with latent embedding optimization. arXiv preprint arXiv:1807.05960 . Saar, M., Edan, Y ., Godo, A., Lepar, J., Parmet, Y ., Halachmi, I.,

    Pith/arXiv arXiv

  15. [27]

    arXiv preprint arXiv:2406.04639

    Cooperative meta-learning with gradient augmentation. arXiv preprint arXiv:2406.04639 . Shojaeipour, A., Falzon, G., Kwan, P., Hadavi, N., Cowley, F.C., Paul, D.,

    Pith/arXiv arXiv

  16. [28]

    The journal of machine learning research 15, 1929–1958

    Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15, 1929–1958. Starshak, T.,

    1929

  17. [29]

    arXiv preprint arXiv:2203.10185

    Negative inner-loop learning rates learn universal features. arXiv preprint arXiv:2203.10185 . Sung, F., Yang, Y ., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.,

    Pith/arXiv arXiv

  18. [30]

    arXiv preprint arXiv:1903.03096

    Meta-dataset: A dataset of datasets for learning to learn from few examples. arXiv preprint arXiv:1903.03096 . Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I.,

    Pith/arXiv arXiv 1903

  19. [33]

    arXiv preprint arXiv:1710.09412

    mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 . Zhang, X., Xuan, C., Ma, Y ., Su, H., 2023a. A high-precision facial recognition method for small-tailed han sheep based on an optimised vision transformer. animal 17, 100886. Zhang, Y ., Liu, X., Sun, C., 2023b. A survey on meta-learning: Concepts, methods, applications, and per...

    Pith/arXiv arXiv

  20. [156]

    Using a cnn-lstm for basic behaviors detection of a single dairy cow in a complex environment

    Wu, D., Wang, Y ., Han, M., Song, L., Shang, Y ., Zhang, X., Song, H., 2021a. Using a cnn-lstm for basic behaviors detection of a single dairy cow in a complex environment. Computers and Electronics in Agriculture 182, 106016. Wu, Y ., Guo, H., Li, Z., Ma, Q., Zhao, Y ., Pezzuolo, A., 2021b. Body condition score for dairy cows method based on vision trans...

    2021

  21. [2008]

    Automated flower classification over a large number of classes, in: 2008 Sixth Indian conference on computer vision, graphics & image processing, IEEE. pp. 722–729. NLIS,

    2008

  22. [2010]

    A rfid-based traceability system for cattle breeding in china, in: 2010 International Conference on Computer Application and System Modeling (ICCASM 2010), IEEE. pp. V2–567. Weng, Z., Liu, S., Zheng, Z., Zhang, Y ., Gong, C.,

    2010

  23. [2014]

    arXiv:1412.6980

    A method for stochastic optimization. arXiv:1412.6980 . Andrew, W.,

    Pith/arXiv arXiv

  24. [2015]

    Convolutional neural network-based image representation for visual loop closure detection, in: 2015 IEEE international conference on information and automation, IEEE. pp. 2238–2245. Hu, H., Dai, B., Shen, W., Wei, X., Sun, J., Li, R., Zhang, Y .,

    2015

  25. [2017]

    arXiv preprint arXiv:1708.04552

    Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 . Ding, C., Pereira, T., Xiao, R., Lee, R.J., Hu, X.,

    Pith/arXiv arXiv

  26. [2018]

    arXiv preprint arXiv:1810.09502

    How to train your maml. arXiv preprint arXiv:1810.09502 . Awad, A.I.,

    Pith/arXiv arXiv

  27. [2019]

    arXiv preprint arXiv:1909.11722

    A theoretical analysis of the number of shots in few-shot learning. arXiv preprint arXiv:1909.11722 . Chen, Y .S., Kuan, C.Y ., Hsu, J.T., Lin, T.T.,

    Pith/arXiv arXiv 1909

  28. [2020]

    arXiv preprint arXiv:2004.10934

    Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934 . Bottou, L.,

    Pith/arXiv arXiv 2004

  29. [2021]

    arXiv preprint arXiv:2110.02095

    Exploring the limits of large scale pre-training. arXiv preprint arXiv:2110.02095 . Adam, D.P.K.J.B., et al.,

    Pith/arXiv arXiv

  30. [2022]

    Automatic cattle identification using yolov5 and mosaic augmentation: A comparative analysis, in: 2022 International Conference on Digital Image Computing: Techniques and Applications (DICTA), IEEE. pp. 1–8. Dumoulin, V ., Houlsby, N., Evci, U., Zhai, X., Goroshin, R., Gelly, S., Larochelle, H.,

    2022

  31. [2023]

    arXiv preprint arXiv:2311.08148

    Cattle identification using muzzle images and deep learning techniques. arXiv preprint arXiv:2311.08148 . Krizhevsky, A., Hinton, G., et al.,

    Pith/arXiv arXiv

  32. [2024]

    Few-shot classification with shared and private branches for cow face recognition, in: 2024 IEEE International Workshop on Radio Frequency and Antenna Technologies (iWRF&AT), IEEE. pp. 1–5. Mishra, N., Rohaninejad, M., Chen, X., Abbeel, P.,

    2024

  33. [2025]

    IEEE Transactions on AgriFood Electronics , 1–12doi:10.1109/TAFE.2025.3574708

    Mhaff: Multihead attention feature fusion of cnn and transformer for cattle identification. IEEE Transactions on AgriFood Electronics , 1–12doi:10.1109/TAFE.2025.3574708. Dulal, R., Zheng, L., Kabir, M.A., McGrath, S., Medway, J., Swain, D., Swain, W.,

    arXiv 2025