Pith. sign in

REVIEW 3 major objections 4 minor 32 references

MangoLeafViT: Leveraging Lightweight Vision Transformer with Runtime Augmentation for Efficient Mango Leaf Disease Classification

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 MobileViT s, a 4.94-million-parameter vision transformer, reaches 99.43% accuracy on eight-class mango leaf disease classification with only 1.44 GFLOPs.

desk verdict A useful lightweight-ViT benchmark on mango leaf disease whose headline accuracy claim is undercut by a self-contradictory evaluation protocol and an overclaimed comparison. read the letter →

arxiv 2505.23961 v1 pith:PE73GNWO submitted 2025-05-29 cs.CV

Rafi Hassan Chowdhury , Sabbir Ahmed This is my paper
classification cs.CV
keywords mangoleafdiseaseclassificationMobileViTlightweightvisiontransformerruntimeaugmentationtransferlearningBDdatasetself-attentionedgedeployment
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 paper claims that a lightweight Vision Transformer, MobileViT s, can classify eight classes of mango leaf disease from the MangoLeafBD dataset at 99.43% accuracy, while using far fewer parameters and less compute than the larger models that previously topped this benchmark. The authors compare three lightweight transformers, EfficientViT b0, TinyViT 5m, and MobileViT s, against standard CNNs and prior specialized models, and report that MobileViT s offers the best accuracy-efficiency trade-off. If the claim holds, accurate disease screening could run on the kind of low-end devices available in rural farming settings instead of requiring server-grade hardware. The contribution is an extension, combining existing components, a known lightweight architecture, transfer learning, and runtime augmentation, on an existing public dataset.

What carries the argument

The central object is MobileViT s, a lightweight hybrid architecture that starts with convolutional layers and MobileNetV2 blocks for local feature extraction, then applies transformer blocks with self-attention to capture global context across image patches, with positional encodings and a classification token. Around this, the pipeline wraps three supporting mechanisms: transfer learning from ImageNet-1k weights, which the ablation shows is decisive for ViTs because they lack CNN inductive biases; runtime augmentation, including CLAHE contrast enhancement, rotation up to 45 degrees, shifting, flips, and grayscale conversion, applied dynamically during training; and a 60/20/20 five-fold cross-validation protocol. The self-attention mechanism is what the paper credits for letting the model separate visually similar classes such as Powdery Mildew and Sooty Mould.

What would settle it

Re-run the fine-tuning and evaluation with test-time augmentation disabled (or with all random seeds fixed), and report per-fold mean accuracy and standard deviation: if the accuracy drops materially below 99.43% or varies widely across seeds, the central claim is not supported in its current form.

Watch

Extended reading notes

Core claim

On MangoLeafBD, MobileViT s pretrained on ImageNet-1k and fine-tuned with runtime augmentation achieves 99.43% accuracy with 4.94 million trainable parameters, a 4.71 MB model, and 1.44 GFLOPs. The paper reports that this is 0.44 percentage points below the highest previously reported accuracy (99.87% from a 20.58-million-parameter model), but with roughly a quarter of the parameters and about half the FLOPs; it also beats a custom CNN baseline (LeafNet, 99.55%, 2.26 GFLOPs) on efficiency. EfficientViT b0 reaches 98.80% with 2.14 million parameters and only 0.1 GFLOPs, positioning the model family as suitable for resource-constrained deployment.

Load-bearing premise

The load-bearing premise is that the reported 99.43% accuracy is a clean measurement of generalization; if the randomized augmentations were applied to validation and test images, as the paper's own figure caption states, then the test predictions, and the headline number, depend on random draws and are not a deterministic, reproducible estimate.

Editorial extensions

If this is right

  • At 4.94 million parameters and 1.44 GFLOPs, MobileViT s is small enough for on-device inference on low-end hardware, while matching the accuracy of much larger models.
  • EfficientViT b0 offers an even lighter option at 2.14 million parameters and 0.1 GFLOPs, with 98.80% accuracy, useful when the compute budget is the binding constraint.
  • Transfer learning from ImageNet-1k is decisive: without it, all three ViTs fall to 67–76% accuracy, while with it they exceed 98%.
  • Runtime augmentation lifts MobileViT s from 99.00% to 99.43%, and adds smaller gains to the other two models.
  • Class-wise analysis reports 100% recall on six of eight classes, with errors concentrated in visually similar pairs such as Bacterial Canker versus Anthracnose and Powdery Mildew versus Sooty Mould.

Reading between the lines

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

  • Editorial inference: if test-time augmentations are what lift the reported accuracy, then a head-to-head comparison with LeafNet under identical, augmentation-free test conditions would likely show a smaller gap, because the prior baselines were not evaluated with those transformations.
  • Editorial inference: the same pipeline is a natural candidate for other crop disease datasets, but MangoLeafBD images are close-up shots on white backgrounds; field conditions with dirt, shadows, and overlapping leaves would probably degrade accuracy and would stress-test whether the augmentations confer real invariance.
  • Editorial inference: since EfficientViT b0 already reaches 98.80% at 0.1 GFLOPs, the practical optimum may be a model between EfficientViT b0 and MobileViT s, possibly quantized, that keeps the accuracy drop under one point at a fraction of the compute.
  • Editorial inference: reporting per-fold standard deviation and a fixed random seed would turn the 99.43% figure from a point estimate into a reproducible benchmark, which is what deployment decisions actually need.
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 proposes MangoLeafViT, a lightweight Vision Transformer pipeline for mango leaf disease classification on the MangoLeafBD dataset. The authors compare three lightweight ViT variants (EfficientViT b0, TinyViT 5m, MobileViT s) against CNN baselines and prior state-of-the-art methods, reporting that MobileViT s achieves 99.43% accuracy with 4.94M parameters and 1.44 GFLOPs, and claiming superiority over existing methods in accuracy, model size, parameter count, and FLOPs. The paper also includes an ablation study on pretraining and augmentation, class-wise metrics, confusion matrix, and Grad-CAM visualizations.

Significance. If the central claims hold, the work is practically relevant: it demonstrates that a relatively small Vision Transformer can match or exceed larger CNN classifiers on an eight-class mango leaf disease benchmark while being deployable on low-end devices. The structural efficiency numbers (parameters, model size, FLOPs) are credible and useful for practitioners. The ablation study and comparison with multiple baselines are also valuable. However, the primary accuracy claim and the efficiency superiority claim are not fully substantiated due to ambiguities in the evaluation protocol and an internal contradiction in the comparison table, so the significance is currently conditional on fixing these issues.

major comments (3)
  1. [§II-B / Figure 2 caption] The evaluation protocol is internally inconsistent: the caption of Figure 2 states that augmentations (rotation, shifts, flips, grayscale) are performed on images 'during training, validation, and testing', while the following paragraph in §II-B says these transformations were 'applied dynamically during training' and that 'no identical image appears across the training, testing, or validation splits'. If random augmentations are applied to test images, the reported 99.43% accuracy is a random variable whose value depends on the random seed and the number of augmented evaluations, making it non-reproducible and not directly comparable with the fixed-protocol baselines in Table IV. If test-time augmentation was not used, the text and figure caption are wrong. Please clarify which protocol was actually used, and if test-time augmentation was used, report results with a fixed seed and/or multiple runs with variance.
  2. [§II-A / Results] The description of the train/validation/test split is ambiguous: the paper states a '5-fold cross-validation strategy, splitting the dataset into 60% for training, 20% for testing, and 20% for validation,' which does not correspond to a standard 5-fold protocol (typically 80/20 or 60/20/20 would be a fixed split, not 5-fold). No per-fold accuracies, standard deviations, or a definition of how the 60/20/20 split relates to the 5 folds are provided. As a result, the reader cannot determine whether the reported 99.43% is the best fold, a mean over folds, or a single run, which is load-bearing for the central accuracy claim. Please report the exact split procedure and per-fold results, including mean and standard deviation.
  3. [Abstract / Table IV] The abstract's claim that the proposed method 'outperforming existing methods in terms of model size, parameter count, and FLOPs count' is contradicted by the paper's own comparison table: Table IV lists Rizvee et al. [17] with 3.26M trainable parameters, which is fewer than MobileViT s's 4.94M parameters. Even though MobileViT s has a smaller model size (4.71 MB vs 13.1 MB) and lower FLOPs (1.44 vs 2.26 GFLOPS), the parameter count claim is false for this comparison. Please correct the claim to accurately state which metrics are improved relative to which baselines, or provide additional context explaining the discrepancy.
minor comments (4)
  1. [References] Reference [1] is titled 'Global production of tomato' but the text in the Introduction discusses mango production; this appears to be a citation error that should be corrected.
  2. [Figure 2 caption] The caption 'Sample augmentations performed on the images during training, validation, and testing phase' directly conflicts with the text in §II-B that says augmentations were applied dynamically during training; this must be reconciled.
  3. [§II-E] The early stopping description is ambiguous: 'halted training if the average validation loss of the current epoch exceeded the best validation loss recorded from previous epochs' with a patience of three epochs is clearer as 'halted if validation loss did not improve for three consecutive epochs'; please rephrase.
  4. [§III-D] In Table IV, the FLOPs values for Rizvee et al. (2.26 GFLOPS) and Varma et al. (2.85 GFLOPS) are reported without specifying the input resolution or measurement method; please state the conditions under which FLOPs were computed so that the comparison is reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical benchmark on an external dataset; self-citations are background, not load-bearing.

full rationale

There is no derivation chain to be circular. The paper fine-tunes three publicly pretrained lightweight Vision Transformers on the MangoLeafBD dataset and reports test-set accuracy, parameters, model size, and FLOPs. The 99.43% accuracy is an empirical result from an external benchmark, not derived from a fitted parameter that is subsequently renamed as a prediction, and no equation in the paper maps an input back onto the claimed output by construction. Several references are co-authored by the present authors (the MangoLeafBD dataset citation [15], the augmentation-practice citation [2], and other background works), but these are supporting citations rather than load-bearing proofs: the dataset is a published external artifact and the augmentation technique is a standard implementation choice, not a uniqueness claim or a fitted ansatz. The internal inconsistencies noted in the evaluation protocol (the Figure 2 caption stating augmentations are performed during validation and testing versus Section II-B stating they are applied dynamically during training, and the abstract's parameter-count claim contradicting Rizvee et al.'s 3.26M parameters in Table IV) are reproducibility and factual-consistency concerns, not circularity. No self-citation chain forces the result, no known result is merely renamed, and no fitted input is called a prediction. Overall circularity score is therefore 0.

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

This work introduces no free parameters beyond standard training hyperparameters, no unproved mathematical axioms, and no invented entities. The central claim rests on the correctness of the dataset labels, the transferability of ImageNet features, and the fairness of the evaluation protocol, all of which are domain assumptions from prior literature or experimental practice.

free parameters (5)
  • Learning rate = 0.00001
    Set to 1e-5 in Section II-E; standard choice for fine-tuning, not fitted to MangoLeafBD.
  • Batch size = 32
    Set in Section II-E; standard value, not fitted.
  • Early stopping patience = 3 epochs
    Chosen in Section II-E; affects when training stops and can influence the final accuracy.
  • Augmentation parameters = rotation up to 45 degrees, shift up to 0.1, flip 50%, grayscale 10%, CLAHE clip 2.0
    Selected in Section II-B; these discrete choices affect generalization but are not optimized against the test set.
  • Input resolution = 224x224
    Fixed by the ViT input requirement in Section II-B.
assumptions (3)
  • domain assumption MangoLeafBD dataset images are correctly labeled and representative of eight mango leaf conditions.
    The paper treats the labels from [15] as ground truth throughout Section II-A and III-C.
  • domain assumption ImageNet-1k pretrained weights provide a beneficial starting point for fine-grained leaf disease classification.
    Invoked in Section II-D as the basis for transfer learning; no specific evidence is given beyond the ablation showing pretraining helps.
  • domain assumption The 60/20/20 per-fold split and the early stopping rule yield a test evaluation that is unbiased and leak-free.
    Section II-A and II-E describe the protocol, but no per-fold results or standard deviations are reported, and the test-augmentation statement in Section II-B raises uncertainty.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MangoLeafViT: Leveraging Lightweight Vision Transformer with Runtime Augmentation for Efficient Mango Leaf Disease Classification." pith.science (2026). https://pith.science/paper/PE73GNWO

@misc{pith2026250523961,
  author       = {Pith},
  title        = {Pith review of: MangoLeafViT: Leveraging Lightweight Vision Transformer with Runtime Augmentation for Efficient Mango Leaf Disease Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PE73GNWO}},
  note         = {Machine review of arXiv:2505.23961}
}
read the original abstract

Ensuring food safety is critical due to its profound impact on public health, economic stability, and global supply chains. Cultivation of Mango, a major agricultural product in several South Asian countries, faces high financial losses due to different diseases, affecting various aspects of the entire supply chain. While deep learning-based methods have been explored for mango leaf disease classification, there remains a gap in designing solutions that are computationally efficient and compatible with low-end devices. In this work, we propose a lightweight Vision Transformer-based pipeline with a self-attention mechanism to classify mango leaf diseases, achieving state-of-the-art performance with minimal computational overhead. Our approach leverages global attention to capture intricate patterns among disease types and incorporates runtime augmentation for enhanced performance. Evaluation on the MangoLeafBD dataset demonstrates a 99.43% accuracy, outperforming existing methods in terms of model size, parameter count, and FLOPs count.

Figures

Figures reproduced from arXiv: 2505.23961 by the authors.

Figure 1
Figure 1. Sample mango leaf images of the different classes from the MangoLeafBD dataset [ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Sample augmentations performed on the images during [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Confusion Matrix of MobileViT s [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Attention map output of correctly predicted samples. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Misclassified sample with visually similar samples of [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 28 canonical work pages

  1. [17]

    Leafnet: A proficient convolutional neural network for detecting seven prominent mango leaf diseases,

    R. A. Rizvee, T. H. Orpa, A. Ahnaf, M. A. Kabir, M. R. Ahmmad Rashid, M. M. Islam, M. Islam, T. Jabid, and M. S. Ali, “Leafnet: A proficient convolutional neural network for detecting seven prominent mango leaf diseases,” Journal of Agriculture and F ood Research , vol. 14, p. 100787, 2023

    work page 2023

  2. [1]

    Global production of tomato

    L. Tridge Co., “Global production of tomato.” Website: https://www. tridge.com/intelligences/tomato/production, 2023. Accessed: Aug. 17, 2024

    work page 2023

  3. [2]

    Less is more: Lighter and faster deep neural architecture for tomato leaf disease classification,

    S. Ahmed, M. B. Hasan, T. Ahmed, M. R. K. Sony, and M. H. Kabir, “Less is more: Lighter and faster deep neural architecture for tomato leaf disease classification,” IEEE Access , vol. 10, pp. 68868–68884, 2022

    work page 2022

  4. [3]

    D. P. Hughes and M. Salath ´e, “An open access repository of images on plant health to enable the development of mobile disease diagnostics (a) Label: Sooty Mould Predicted: Powdery Mildew (b) Label: Powdery Mildew Predicted: Sooty Mould (c) Label: Bacterial Canker Predicted: Anthracnose (d) Label: Anthracnose Fig. 5: Misclassified sample with visually si...

    work page 2015

  5. [4]

    Plant disease detection and classification by deep learning—a review,

    L. Li, S. Zhang, and B. Wang, “Plant disease detection and classification by deep learning—a review,” IEEE Access , vol. 9, pp. 56683–56698, 2021

    work page 2021

  6. [5]

    A critical analysis of deep learning applications in crop pest classification: Promising pathways and limitations,

    M. H. Rafi, M. Ratul Mahjabin, M. S. Rahman, M. Hasanul Kabir, and S. Ahmed, “A critical analysis of deep learning applications in crop pest classification: Promising pathways and limitations,” in 2023 26th International Conference on Computer and Information Technology (ICCIT), pp. 1–6, 2023

    work page 2023

  7. [6]

    Vision transformer meets convolutional neural network for plant disease classification,

    P. S. Thakur, S. Chaturvedi, P. Khanna, T. Sheorey, and A. Ojha, “Vision transformer meets convolutional neural network for plant disease classification,” Ecological Informatics , vol. 77, p. 102245, 2023

    work page 2023

  8. [7]

    Rethinking cooking state recognition with vision transformers,

    A. M. Khan, A. Ashrafee, R. Sayera, S. Ivan, and S. Ahmed, “Rethinking cooking state recognition with vision transformers,” in 25th International Conference on Computer and Information Technology (ICCIT) , pp. 170– 175, 2022

    work page 2022

Show all 32 references
  1. [8]

    Vision transformer for plant disease detection: Plantvit,

    P. S. Thakur, P. Khanna, T. Sheorey, and A. Ojha, “Vision transformer for plant disease detection: Plantvit,” in Computer Vision and Image Processing, pp. 501–511, Springer International Publishing, 2022

    2022

  2. [9]

    Exe-net: Explainable ensemble network for potato leaf disease classification,

    T. Ahmed, M. B. Hasan, S. Ahmed, and M. H. Kabir, “Exe-net: Explainable ensemble network for potato leaf disease classification,” in 2024 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE) , pp. 335–339, 2024

    2024

  3. [10]

    An efficient transfer learning-based approach for apple leaf disease classification,

    M. H. Ashmafee, T. Ahmed, S. Ahmed, M. B. Hasan, M. N. Jahan, and A. Ashikur Rahman, “An efficient transfer learning-based approach for apple leaf disease classification,” in 2023 International Conference on Electrical, Computer and Communication Engineering (ECCE) , pp. 1–6, 2023

    2023

  4. [11]

    Deep learning in agriculture: A survey,

    A. Kamilaris and F. X. Prenafeta-Bold ´u, “Deep learning in agriculture: A survey,” Computers and Electronics in Agriculture , vol. 147, pp. 70–90, 2018

    2018

  5. [12]

    Cotton leaf disease identification using transfer learning,

    A. Herok and S. Ahmed, “Cotton leaf disease identification using transfer learning,” in 2023 International Conference on Information and Communication Technology for Sustainable Development (ICICT4SD) , pp. 158–162, 2023

    2023

  6. [13]

    Automatic mango leaf disease detection using different transfer learning models,

    T. Varma, P. Mate, N. A. Azeem, S. Sharma, and B. Singh, “Automatic mango leaf disease detection using different transfer learning models,” Multimedia Tools and Applications , May 7 2024

    2024

  7. [14]

    A novel machine learning approach for fast and efficient detection of mango leaf diseases,

    S. M. M. R. Swapno, S. N. Nobel, M. B. Islam, R. Haque, V . P. Meena, and F. Benedetto, “A novel machine learning approach for fast and efficient detection of mango leaf diseases,” in IEEE 3rd International Conference on Computing and Machine Intelligence (ICMI) , pp. 1–7, 2024

    2024

  8. [15]

    Mangoleafbd: A comprehensive image dataset to classify diseased and healthy mango leaves,

    S. I. Ahmed, M. Ibrahim, M. Nadim, M. M. Rahman, M. M. Shejunti, T. Jabid, and M. S. Ali, “Mangoleafbd: A comprehensive image dataset to classify diseased and healthy mango leaves,” Data in Brief , vol. 47, p. 108941, 2023

    2023

  9. [16]

    Detect bangladeshi mango leaf diseases using lightweight convolutional neural network,

    N. I. Mahbub, F. Naznin, M. I. Hasan, S. M. R. Shifat, M. A. Hossain, and M. Z. Islam, “Detect bangladeshi mango leaf diseases using lightweight convolutional neural network,” in International Conference on Electrical, Computer and Communication Engineering (ECCE) , pp. 1–6, 2023

    2023

  10. [18]

    Mobilevit: Light-weight, general-purpose, and mobile-friendly vision transformer,

    S. Mehta and M. Rastegari, “Mobilevit: Light-weight, general-purpose, and mobile-friendly vision transformer,” 2022

    2022

  11. [19]

    Efficientvit: Multi-scale linear attention for high-resolution dense prediction,

    H. Cai, J. Li, M. Hu, C. Gan, and S. Han, “Efficientvit: Multi-scale linear attention for high-resolution dense prediction,” 2024

    2024

  12. [20]

    Tinyvit: Fast pretraining distillation for small vision transformers,

    K. Wu, J. Zhang, H. Peng, M. Liu, B. Xiao, J. Fu, and L. Yuan, “Tinyvit: Fast pretraining distillation for small vision transformers,” in Computer Vision – ECCV 2022 , (Cham), pp. 68–85, Springer Nature Switzerland, 2022

    2022

  13. [21]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems , vol. 30, Curran Associates, Inc., 2017

    2017

  14. [22]

    Gaitgcn++: Improving gcn-based gait recognition with part-wise attention and dropgraph,

    M. B. Hasan, T. Ahmed, S. Ahmed, and M. H. Kabir, “Gaitgcn++: Improving gcn-based gait recognition with part-wise attention and dropgraph,” Journal of King Saud University - Computer and Information Sciences, vol. 35, no. 7, p. 101641, 2023

    2023

  15. [23]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” 2021

    2021

  16. [24]

    Mobilenetv2: Inverted residuals and linear bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mobilenetv2: Inverted residuals and linear bottlenecks,” in Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 4510–4520, 2018

    2018

  17. [25]

    Two decades of bengali handwritten digit recognition: A survey,

    A. B. M. Ashikur Rahman, M. B. Hasan, S. Ahmed, T. Ahmed, M. H. Ashmafee, M. R. Kabir, and M. H. Kabir, “Two decades of bengali handwritten digit recognition: A survey,”IEEE Access, vol. 10, pp. 92597– 92632, 2022

    2022

  18. [26]

    Fruit quality assessment with densely connected convolutional neural network,

    M. S. Morshed, S. Ahmed, T. Ahmed, M. U. Islam, and A. Ashikur Rah- man, “Fruit quality assessment with densely connected convolutional neural network,” in 2022 12th International Conference on Electrical and Computer Engineering (ICECE) , pp. 1–4, 2022

    2022

  19. [27]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE Conference on Computer Vision and Pattern Recognition , pp. 248–255, 2009

    2009

  20. [28]

    Do vision transformers see like convolutional neural networks?,

    M. Raghu, T. Unterthiner, S. Kornblith, C. Zhang, and A. Dosovitskiy, “Do vision transformers see like convolutional neural networks?,” 2022

    2022

  21. [29]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” 2017

    2017

  22. [30]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in Neural Information Processing Systems , vol. 25, Curran Associates, Inc., 2012

    2012

  23. [31]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” in Proceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR) , July 2017

    2017

  24. [32]

    Grad-cam: Visual explanations from deep networks via gradient-based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” International Journal of Computer Vision , vol. 128, p. 336–359, Oct. 2019

    2019

Pith tools

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