Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Performance Analysis of Few-Shot Learning Approaches for Bangla Handwritten Character and Digit Recognition

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read SynergiProtoNet beats five few-shot baselines on Bangla handwriting recognition, from 1, 5, and 10 examples per class.

desk verdict Useful Bangla few-shot numbers undermined by an unsupported benchmark claim. read the letter →

arxiv 2506.00447 v1 pith:PP5IKQ3J submitted 2025-05-31 cs.CV

Mehedi Ahamed , Radib Bin Kabir , Tawsif Tashwar Dipto , Mueeze Al Mushabbir , Sabbir Ahmed , Md. Hasanul Kabir This is my paper · ORCID
classification cs.CV
keywords few-shotlearningBanglahandwrittencharacterrecognitionprototypicalnetworkhybridencoderResNet18cross-lingualtransferdigitepisodictraining
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 claims that a hybrid encoder, which concatenates low-level features from a small CNN with high-level features from a pre-trained ResNet18, outperforms five established few-shot learning methods for recognizing Bangla handwritten characters and digits. The claim matters because Bangla's intricate script and the scarcity of large labeled datasets make it a demanding test case for few-shot learning; if the hybrid approach works here, it could generalize to other low-resource scripts. The paper evaluates this across four settings: training and testing on different classes of the same dataset, training on one Bangla dataset and testing on another, training on Devanagari and testing on Bangla, and training on digits 0-5 while testing on digits 6-9. In most of these settings, SynergiProtoNet achieves the highest accuracy and F1 score, with the largest gains appearing in the 10-shot scenarios.

What carries the argument

The central object is the hybrid dual encoder in SynergiProtoNet: feature vectors $F_1(\theta)$ from a four-layer CNN and $F_2(\theta)$ from a pre-trained ResNet18 are concatenated along the feature dimension into a combined embedding $z = \text{concatenate}(f_{\text{CNN}}(x), f_{\text{ResNet}}(x), \text{dim}=1)$. Class prototypes are the means of support-set combined embeddings, $p_c = \frac{1}{K}\sum_{i=1}^K f_\phi(x_i^{(c)})$, and a query is classified by softmax over negative Euclidean distances $d(z_q, p_c) = \|z_q - p_c\|_2$. The machinery is standard episodic prototypical training with 500 tasks per epoch over 30 epochs, using SGD with a multi-step learning-rate schedule; the proposed contribution is specifically the feature concatenation step, which is what the paper credits for its accuracy gains.

What would settle it

Train and evaluate a prototypical network with only the ResNet18 encoder (no CNN branch) under the paper's exact 5-way, 3-way, 500-tasks-per-epoch protocol on the same four evaluation settings. If the ResNet18-only model matches or exceeds SynergiProtoNet's 10-shot accuracy in the monolingual intra-dataset and inter-dataset settings, the central claim that the concatenation mechanism drives the gains would be refuted.

Watch

Extended reading notes

Core claim

SynergiProtoNet, a prototypical network whose embedding is the concatenation of a CNN encoder and a pre-trained ResNet18 encoder, consistently achieves state-of-the-art few-shot accuracy for Bangla handwritten character and digit recognition. The claim is that the fusion of low-level features (edges, textures, corners) from the CNN with high-level features (complex textures, object configurations) from ResNet18 yields a richer embedding, so that class prototypes computed by averaging support-set embeddings become more discriminative. The paper reports top accuracy in monolingual intra-dataset, monolingual inter-dataset, and split-digit testing, and matches or exceeds the best baseline in cross-lingual transfer except in the 1-shot case where Relation Network is slightly higher. The authors present this as establishing a new benchmark for few-shot learning on handwritten characters and digits.

Load-bearing premise

The paper attributes SynergiProtoNet's gains to combining CNN features with pre-trained ResNet18 features, but it never runs an ablation that separates the contribution of the pre-trained backbone from the contribution of the concatenation itself; if a ResNet18-only prototype network already matches the hybrid's accuracy, the fusion is not the cause of the improvement.

Editorial extensions

If this is right

  • If the claim holds, few-shot recognition systems for Bangla documents could be built with only a handful of labeled examples per character class, reducing the need for large manual annotation efforts in banking, education, and government digitization workflows.
  • The reported cross-lingual transfer results (training on Devanagari, testing on Bangla) suggest that a model trained on one script may be adapted to a related script with minimal additional labeled data, which would support building multi-script recognition systems from shared representations.
  • The split-digit result, where the model trains on digits 0-5 and classifies 6-9, indicates that the hybrid encoder can separate classes never seen together during training, a property useful for incremental addition of new character classes.
  • The consistent advantage of the prototypical baseline over Matching, Relation, SimpleShot, and BD-CSPN in these experiments reinforces the choice of prototype-based classification with Euclidean distance as the right scaffolding for low-resource handwriting tasks.
  • Since the approach is not Bangla-specific, the same concatenation strategy could be applied to other low-resource scripts, though the paper only demonstrates this for the Devanagari-to-Bangla transfer direction.

Reading between the lines

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

  • The paper's strongest claim implicitly predicts that ResNet18 alone, under the same prototypical framework, would score below SynergiProtoNet; a direct comparison of ResNet18-only versus the hybrid would confirm whether the fusion mechanism, rather than the pre-trained backbone, is responsible for the gains.
  • The unusually low 1-shot accuracy in the split-digit experiment (37.4%) suggests a regime boundary: with one support example, the hybrid encoder may be too high-capacity to train stably, implying that a lighter or frozen-backbone variant could be preferable in true one-shot deployments.
  • A natural extension the authors do not test is a full cross-script matrix (Bangla to Devanagari, Tamil to Bangla, etc.); if the model transfers well in both directions, the 'complex-script generalization' hypothesis would be substantially stronger.
  • The availability of the code on GitHub makes the claimed benchmark directly checkable: rerunning the four evaluation settings with the same baselines and random seeds would settle whether the margins are stable or sensitive to the specific episodic tasks.
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

4 major / 5 minor

Summary. The paper proposes SynergiProtoNet, a few-shot classifier that concatenates features from a small CNN and a pre-trained ResNet18 into a prototypical-network framework, and evaluates it on Bangla handwritten character and digit recognition. The evaluation spans four settings—monolingual intra-dataset, monolingual inter-dataset, cross-lingual transfer, and split-digit testing—against five baselines (Matching Network, SimpleShot, Relation Network, BD-CSPN, Prototypical Network). The abstract and conclusion claim that SynergiProtoNet consistently achieves superior performance and establishes a new benchmark.

Significance. If the empirical claims were fully supported, the paper would provide a useful recipe for low-resource script recognition and a comparatively broad evaluation of few-shot methods on Bangla. Strengths include the public code release and the use of four distinct evaluation protocols. However, the headline claim of consistent superiority is contradicted by the paper's own tables, and the absence of ablations and error bars prevents the reader from attributing any observed advantage to the proposed architecture rather than to the pre-trained backbone or to chance.

major comments (4)
  1. [Section III, Tables III and IV] The abstract and conclusion assert that SynergiProtoNet 'consistently... superior' and establishes a new benchmark, but the reported results contradict this. In Table III (cross-lingual, 1-shot), Relation Network achieves 61.12% accuracy versus 58.59% for SynergiProtoNet, and in Table IV (split-digit, 1-shot), Matching Network achieves 73.67% versus 37.40% for SynergiProtoNet. The text acknowledges the cross-lingual gap in Section III-3 but does not disclose the split-digit 1-shot collapse in Section III-4. This internal inconsistency is load-bearing for the central claim and must be resolved, at minimum by qualifying the claims to multi-shot settings or by providing a technical explanation for the 1-shot failures.
  2. [Section III (all tables)] No error bars, confidence intervals, or significance tests are reported, so the reader cannot determine whether differences between models are meaningful. The instability within the tables (for example, Matching Network in Table I drops from 69.64% in 1-shot to 38.66% in 5-shot to 36.36% in 10-shot) indicates high variance that single-point accuracies cannot capture. The authors should report means and standard deviations over multiple seeds and task sets and perform paired significance tests before claiming superiority.
  3. [Section II-D] The contribution of the hybrid encoder is not isolated. The paper describes concatenating CNN and pre-trained ResNet18 features but provides no ablation that trains either the CNN-only or ResNet18-only encoder under the same prototypical framework. Consequently, the observed multi-shot gains cannot be attributed to the fusion itself rather than to the pre-trained ResNet18 backbone; this is a load-bearing gap for the claim that the proposed architecture is the cause of any improvements.
  4. [Section II-D and II-E] Reproducibility is currently insufficient. The paper states that a pre-trained ResNet18 is used, but does not specify the pretraining dataset, whether weights are frozen or fine-tuned, how the 84x84 input is normalized or resized, or how the CNN and ResNet18 output dimensions are aligned before concatenation. In addition, the relationship between the stated 5-way/3-way setups and the class splits (e.g., 39 training consonants versus 11 test vowels) is not described, leaving unclear how episodes are constructed and how the query-set accuracy is averaged over episodes.
minor comments (5)
  1. [Figure 1 caption] The caption spells the proposed model as 'SynergiProtonet'; it should be 'SynergiProtoNet' for consistency with the text.
  2. [Equation (4)] The denominator uses 'Nc' without defining it; it presumably denotes the number of classes and should be defined or renamed.
  3. [Table I] Matching Network accuracy decreases as the number of shots increases (69.64% to 38.66% to 36.36%); this is unusual and should be commented on, as it suggests either high variance or a training/protocol issue.
  4. [References] Reference [32] is cited for 'BD-CSPN' but the listed title is 'Prototype rectification for few-shot learning'; please clarify the relationship between the method name and the reference and ensure the citation is accurate.
  5. [Section II-B] The statement that 'the query set contains 10 random samples' is vague; the paper should specify how many episodes were used for the reported accuracies and how variance across episodes was handled.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the few-shot evaluation is self-contained and no prediction reduces to a fitted input or self-citation by construction.

full rationale

The paper's derivation chain is an empirical benchmark of few-shot classifiers. SynergiProtoNet concatenates features from a CNN encoder and a pre-trained ResNet18 encoder, forms class prototypes as means of support embeddings, and classifies queries by Euclidean distance (Eqs. 1-4). No target accuracy is used as an input, no parameter is fitted to the reported test outcomes, and the support/query split uses disjoint classes in every evaluation setting. The central claim of 'consistently superior performance' is an empirical assertion contradicted by the paper's own Table III (1-shot cross-lingual: 58.59 versus Relation Network 61.12) and Table IV (1-shot split digit: 37.40 versus Matching Network 73.67), but internal inconsistency is a correctness or reporting issue, not circularity. The cited prior works from the same authors (e.g., Refs. [3], [4], [10]-[12], [16], [19]) are contextual references and are not load-bearing for the proposed method. The only potentially under-supported component is the use of a pre-trained ResNet18 without an ablation separating backbone contribution from feature fusion; however, this is an architectural attribution problem rather than a circular reduction, because the model is still evaluated on held-out classes with no fitted constants renamed as predictions. Therefore the circularity burden is low and the appropriate score is 0.

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

No hand-fitted constants appear in the method; the only numbers are standard hyperparameters and network weights trained end-to-end. The main assumptions are the transferability of pre-trained ResNet18 features and the validity of the evaluation splits. SynergiProtoNet is a neural architecture, not a postulated physical entity.

assumptions (3)
  • domain assumption Pre-trained ResNet18 weights (presumably ImageNet) provide a useful feature space for Bangla handwriting.
    Section II-D describes the ResNet encoder as 'pre-trained' but does not specify the pretraining dataset or provide an ablation against a randomly initialized backbone.
  • domain assumption The evaluation splits correctly implement few-shot protocols with non-overlapping classes.
    Section II-E states that overlapping classes were avoided but does not detail how splits were constructed or whether test classes are truly disjoint.
  • domain assumption Reported accuracies are stable across random episodes despite no confidence intervals.
    Results are point estimates from an unreported number of test episodes, so the stability of the numbers is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Performance Analysis of Few-Shot Learning Approaches for Bangla Handwritten Character and Digit Recognition." pith.science (2026). https://pith.science/paper/PP5IKQ3J

@misc{pith2026250600447,
  author       = {Pith},
  title        = {Pith review of: Performance Analysis of Few-Shot Learning Approaches for Bangla Handwritten Character and Digit Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PP5IKQ3J}},
  note         = {Machine review of arXiv:2506.00447}
}
read the original abstract

This study investigates the performance of few-shot learning (FSL) approaches in recognizing Bangla handwritten characters and numerals using limited labeled data. It demonstrates the applicability of these methods to scripts with intricate and complex structures, where dataset scarcity is a common challenge. Given the complexity of Bangla script, we hypothesize that models performing well on these characters can generalize effectively to languages of similar or lower structural complexity. To this end, we introduce SynergiProtoNet, a hybrid network designed to improve the recognition accuracy of handwritten characters and digits. The model integrates advanced clustering techniques with a robust embedding framework to capture fine-grained details and contextual nuances. It leverages multi-level (both high- and low-level) feature extraction within a prototypical learning framework. We rigorously benchmark SynergiProtoNet against several state-of-the-art few-shot learning models: BD-CSPN, Prototypical Network, Relation Network, Matching Network, and SimpleShot, across diverse evaluation settings including Monolingual Intra-Dataset Evaluation, Monolingual Inter-Dataset Evaluation, Cross-Lingual Transfer, and Split Digit Testing. Experimental results show that SynergiProtoNet consistently outperforms existing methods, establishing a new benchmark in few-shot learning for handwritten character and digit recognition. The code is available on GitHub: https://github.com/MehediAhamed/SynergiProtoNet.

Figures

Figures reproduced from arXiv: 2506.00447 by the authors.

Figure 1
Figure 1. Overview of our proposed SynergiProtonet. The architecture consists of a dual-encoder system, integrating CNN and ResNet18 encoders, to process both Support Set and Query Set samples. Each image is passed through both encoders, yielding feature vectors F1(θ) from the CNN and F2(θ) from the ResNet18. These feature vectors are concatenated to form a combined representation Z for each Support image, and a Query feature… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages

  1. [1]

    Handwritten opti- cal character recognition (ocr): A comprehensive systematic literature review (slr),

    J. Memon, M. Sami, R. A. Khan, and M. Uddin, “Handwritten opti- cal character recognition (ocr): A comprehensive systematic literature review (slr),” IEEE Access, vol. 8, pp. 142 642–142 668, 2020

    work page 2020

  2. [2]

    A proposed approach for character recogni- tion using document analysis with ocr,

    H. Singh and A. Sachan, “A proposed approach for character recogni- tion using document analysis with ocr,” in 2018 Second International Conference on Intelligent Computing and Control Systems (ICICCS) . IEEE, 2018, pp. 190–195

    work page 2018

  3. [3]

    Improved speech emotion recognition in bengali language using deep learning,

    S. Aziz, N. H. Arif, S. Ahbab, S. Ahmed, T. Ahmed, and M. H. Kabir, “Improved speech emotion recognition in bengali language using deep learning,” in 2023 26th International Conference on Computer and Information Technology (ICCIT), 2023, pp. 1–6

    work page 2023

  4. [4]

    Csvc- net: Code-switched voice command classification using deep cnn-lstm network,

    A. Yasmeen, F. I. Rahman, S. Ahmed, and M. H. Kabir, “Csvc- net: Code-switched voice command classification using deep cnn-lstm network,” in 2021 Joint 10th International Conference on Informatics, Electronics & Vision (ICIEV) and 2021 5th International Conference on Imaging, Vision & Pattern Recognition (icIVPR) , 2021, pp. 1–8

    work page 2021

  5. [6]

    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. 92 597–92 632, 2022

    work page 2022

  6. [7]

    A complete bangla optical character recognition system: An effective approach,

    T. Ahmed, M. N. Raihan, R. Kushol, and M. S. Salekin, “A complete bangla optical character recognition system: An effective approach,” in 2019 22nd International Conference on Computer and Information Technology (ICCIT), 2019, pp. 1–7

    work page 2019

  7. [8]

    Bangla handwritten numeral character recognition using directional pattern,

    T. I. Aziz, A. S. Rubel, M. S. Salekin, and R. Kushol, “Bangla handwritten numeral character recognition using directional pattern,” in 2017 20th International Conference of Computer and Information Technology (ICCIT), 2017, pp. 1–5

    work page 2017

  8. [9]

    A systematic review on the chronological development of bangla sign language recognition systems,

    A. Khatun, M. S. Shahriar, M. H. Hasan, K. Das, S. Ahmed, and M. S. Islam, “A systematic review on the chronological development of bangla sign language recognition systems,” in 2021 Joint 10th International Conference on Informatics, Electronics & Vision (ICIEV) and 2021 5th International Conference on Imaging, Vision & Pattern Recognition (icIVPR), 2021...

    work page 2021

Show all 33 references
  1. [10]

    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. 68 868–68 884, 2022

    2022

  2. [11]

    Performance analy- sis of yolo-based architectures for vehicle detection from traffic images in bangladesh,

    R. M. Alamgir, A. A. Shuvro, M. Al Mushabbir, M. A. Raiyan, N. J. Rani, M. M. Rahman, M. H. Kabir, and S. Ahmed, “Performance analy- sis of yolo-based architectures for vehicle detection from traffic images in bangladesh,” in 2022 25th International Conference on Computer and ...

    2022

  3. [12]

    Recognizing bangladeshi traffic signs in the wild,

    A. N. Ashik, M. S. H. Shanto, R. H. Khan, M. H. Kabir, and S. Ahmed, “Recognizing bangladeshi traffic signs in the wild,” in 2022 25th International Conference on Computer and Information Technology (ICCIT), 2022, pp. 1004–1009

    2022

  4. [13]

    Handwritten bengali character recognition using deep convolution neural network,

    S. Maity, A. Dey, A. Chowdhury, and A. Banerjee, “Handwritten bengali character recognition using deep convolution neural network,” in Machine Learning, Image Processing, Network Security and Data Sciences. Singapore: Springer Singapore, 2020, pp. 84–92

    2020

  5. [14]

    Banglanet: Bangla handwritten character recognition using ensembling of convolutional neural network,

    C. Saha and M. M. Rahman, “Banglanet: Bangla handwritten character recognition using ensembling of convolutional neural network,” arXiv, Jan. 2024

    2024

  6. [15]

    Bornonet: Bangla handwritten characters recognition using convolu- tional neural network,

    A. S. A. Rabby, S. Haque, M. S. Islam, S. Abujar, and S. Hossain, “Bornonet: Bangla handwritten characters recognition using convolu- tional neural network,” Procedia Computer Science , vol. 143, pp. 528– 535, 01 2018

    2018

  7. [16]

    Huruf: An application for arabic handwritten character recognition using deep learning,

    M. Kamal, F. Shaiara, C. M. Abdullah, S. Ahmed, T. Ahmed, and M. H. Kabir, “Huruf: An application for arabic handwritten character recognition using deep learning,” in 25th International Conference on Computer and Information Technology (ICCIT) , 2022, pp. 1131–1136

    2022

  8. [17]

    Generalizing from a few examples: A survey on few-shot learning,

    Y . Wang, Q. Yao, J. T. Kwok, and L. M. Ni, “Generalizing from a few examples: A survey on few-shot learning,” ACM Comput. Surv., vol. 53, no. 3, jun 2020

    2020

  9. [18]

    Classification of plant disease from leaf images using few-shot learning,

    S. Ahmed, “Classification of plant disease from leaf images using few-shot learning,” MSc Thesis, Department of Computer Science and Engineering (CSE), Islamic University of Technology, 2022

    2022

  10. [19]

    Fused-net: Enhancing few-shot traffic sign detection with unfrozen parameters, pseudo-support sets, embedding normalization, and domain adaptation,

    M. A. Rahman, N. I. Asad, M. M. H. Omi, M. B. Hasan, S. Ahmed, and M. H. Kabir, “Fused-net: Enhancing few-shot traffic sign detection with unfrozen parameters, pseudo-support sets, embedding normalization, and domain adaptation,” arXiv, 2024

    2024

  11. [20]

    An enhanced prototypical network architec- ture for few-shot handwritten urdu character recognition,

    R. Sahay and M. Coustaty, “An enhanced prototypical network architec- ture for few-shot handwritten urdu character recognition,” IEEE Access, vol. PP, pp. 1–1, 01 2023

    2023

  12. [21]

    Offline handwritten amharic character recognition using few-shot learn- ing,

    M. Samuel, L. Schmidt-Thieme, D. P. Sharma, A. Sinamo, and A. Bruck, “Offline handwritten amharic character recognition using few-shot learn- ing,” in Pan-African Conference on Artificial Intelligence . Springer Nature Switzerland, 2023, pp. 233–244

    2023

  13. [22]

    Few-shot learning for tamil handwritten character recognition using deep siamese convolutional neural network,

    N. Shaffi and F. Hajamohideen, “Few-shot learning for tamil handwritten character recognition using deep siamese convolutional neural network,” in Applied Intelligence and Informatics . Cham: Springer International Publishing, 2021, pp. 204–215

    2021

  14. [23]

    Few- shot learning for character recognition in persian historical documents,

    A. Hajebrahimi, M. E. Santoso, M. Kovacs, and V . V . Kryssanov, “Few- shot learning for character recognition in persian historical documents,” in Machine Learning, Optimization, and Data Science. Cham: Springer Nature Switzerland, 2024, pp. 259–273

    2024

  15. [24]

    Siamese-transformer network for offline handwritten signature verification using few-shot,

    P. Majumder, A. Joaa, E. R. Rhythm, M. H. K. Mehedi, and A. A. Rasel, “Siamese-transformer network for offline handwritten signature verification using few-shot,” 12 2023, pp. 1–6

    2023

  16. [25]

    A novel siamese network for few/zero-shot handwritten character recognition tasks,

    N. Elaraby, S. Barakat, and A. Rezk, “A novel siamese network for few/zero-shot handwritten character recognition tasks,” Computers, Materials and Continua , vol. 74, pp. 1837–1854, 08 2022

    2022

  17. [26]

    Banglalekha-isolated: A multi-purpose comprehensive dataset of handwritten bangla isolated characters,

    M. Biswas, R. Islam, G. Shom, M. Shopon, N. Mohammed, S. Momen, and A. Abedin, “Banglalekha-isolated: A multi-purpose comprehensive dataset of handwritten bangla isolated characters,” Data in Brief, vol. 12, pp. 103–107, 2017

    2017

  18. [27]

    Bangla handwritten character db cmaterdb 3.1.2,

    M. M. Rahman, “Bangla handwritten character db cmaterdb 3.1.2,” 2023, [accessed 24-August-2024]. [Online]. Available: https://www. kaggle.com/datasets/mostafiz53/basicfinal

    2023

  19. [28]

    Hindi character recognition,

    S. Sinha, “Hindi character recognition,” Sep. 2021, [accessed 4-August- 2024]. [Online]. Available: https://www.kaggle.com/datasets/suvooo/ hindi-character-recognition

    2021

  20. [29]

    Numtadb - assembled bengali handwritten digits,

    S. Alam, T. Reasat, R. M. Doha, and A. I. Humayun, “Numtadb - assembled bengali handwritten digits,” arXiv, 2018

    2018

  21. [30]

    Matching networks for one shot learning,

    O. Vinyals, C. Blundell, T. Lillicrap, K. Kavukcuoglu, and D. Wier- stra, “Matching networks for one shot learning,” Advances in neural information processing systems , vol. 29, 2016

    2016

  22. [31]

    Learning to compare: Relation network for few-shot learning,

    F. Sung, Y . Yang, L. Zhang, T. Xiang, P. H. S. Torr, and T. M. Hospedales, “Learning to compare: Relation network for few-shot learning,” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

    2018

  23. [32]

    Prototype rectification for few-shot learning,

    J. Liu, L. Song, and Y . Qin, “Prototype rectification for few-shot learning,” in Computer Vision – ECCV 2020 . Springer International Publishing, 2020, pp. 741–756

    2020

  24. [33]

    Sim- pleshot: Revisiting nearest-neighbor classification for few-shot learning,

    Y . Wang, W.-L. Chao, K. Q. Weinberger, and L. van der Maaten, “Sim- pleshot: Revisiting nearest-neighbor classification for few-shot learning,” arXiv, 2019

    2019

  25. [34]

    Prototypical networks for few- shot learning,

    J. Snell, K. Swersky, and R. Zemel, “Prototypical networks for few- shot learning,” in Advances in Neural Information Processing Systems , vol. 30. Curran Associates, Inc., 2017

    2017

Pith tools

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