Pith. sign in

REVIEW 4 major objections 4 minor 38 references

Few-Shot Learning with Global Class Representations

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read This paper claims that including novel-class training samples from the start, with a registration module and sample synthesis, improves both standard and generalized few-shot classification on Omniglot and miniImageNet.

desk verdict The method is coherent and the generalized FSL comparison is fair, but the standard FSL results are invalid because the test classes are used as novel classes during training, so the SOTA claim collapses. read the letter →

arxiv 1908.05257 v1 pith:5HJNBAXN submitted 2019-08-14 cs.CV

Tiange Luo , Aoxue Li , Tao Xiang , Weiran Huang , Liwei Wang This is my paper · ORCID
classification cs.CV
keywords few-shotlearningglobalclassrepresentationregistrationmodulesamplesynthesisepisodictraininggeneralizedOmniglotminiImageNet
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 argues that few-shot classification can be improved by learning a single global representation for every class, including the target novel classes, rather than only adapting to them at test time. The method trains with both base and novel class samples from the start, using a registration module to tie each episode's support-set mean to the corresponding global representation and a sample-synthesis module to inflate the tiny novel-class support. If the central claim holds, standard few-shot accuracy improves, and the model transfers naturally to generalized few-shot recognition, where test labels can come from both base and novel classes.

What carries the argument

The machinery has three parts. Global class representations are per-class vectors initialized as feature averages and updated during training. The registration module computes softmax similarity between an episodic class representation and every global representation, pulling the correct global vector toward the class and pushing others away. The sample synthesis module creates extra novel-class training examples by hallucinating variants and taking weighted convex combinations of their features, reducing the imbalance between abundant base classes and few-shot novel classes.

What would settle it

Take the same architecture and training schedule, remove all novel-class training samples from the training set so that only base classes appear during episodic training, and measure 5-way 5-shot accuracy on miniImageNet. If accuracy drops to or below the compared methods, the reported state-of-the-art result depends on the test-class training data rather than on the registration mechanism.

Watch

Extended reading notes

Core claim

The central discovery is that a class can be represented globally by jointly optimizing a feature extractor, a registration module, and per-class global vectors over an episode that mixes base and novel classes. In each episode, the support set of a class is condensed into an episodic representation (average for base classes, synthesized for novel classes); a softmax over distances to all global representations selects and combines them; and a query-set cross-entropy loss plus a registration loss update everything end-to-end. The result, the paper reports, is that novel-class accuracy rises on standard few-shot benchmarks and rises even more relative to baselines under generalized FSL.

Load-bearing premise

The standard few-shot results are computed with the 20 test classes' few-shot images included in training, so the comparison against methods that never see those classes is only valid if using test-class labels during training is an accepted few-shot protocol.

Editorial extensions

If this is right

  • On Omniglot, the method reports 99.72% for 5-way 1-shot and 99.90% for 5-way 5-shot, with 20-way accuracy also exceeding prior results.
  • On miniImageNet, the method reports 53.21% for 5-way 1-shot and 72.34% for 5-way 5-shot, with the 5-shot result showing a decisive jump over compared approaches.
  • Under generalized FSL on miniImageNet, joint-label accuracy rises from around 31-32% for compared methods to 39.14%, while novel-class accuracy rises from below 2% to 12.98%.
  • Because base and novel classes are trained jointly, the same model can classify test images from both label spaces and can add new unseen novel classes by learning only their global representations with existing parameters fixed.
  • The method's design makes it a unified model for both standard few-shot learning and generalized few-shot learning without changing the training pipeline.

Reading between the lines

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

  • If the field decides that using the 20 test classes' few-shot labels during training violates the standard few-shot protocol, the miniImageNet comparison in Section 4.1.4 should be re-run under a protocol where novel classes are never seen in training; the margin may shrink substantially.
  • The generality of the global-representation idea could be tested by plugging the registration module into other episodic few-shot learners, such as metric-based or optimization-based methods, and measuring whether the gain persists when novel labels are not used.
  • The generalized-FSL numbers suggest that the practical bottleneck for real deployment is not standard few-shot accuracy but the base-versus-novel imbalance; a natural extension is to use unlabeled base-class data or semi-supervised novel data to further push novel-class accuracy.
  • The convex-combination synthesis step is simple enough to be reused as a data-augmentation baseline in later few-shot work, independent of the registration module.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a few-shot learning (FSL) method that learns a global class representation for every class, base and novel, by jointly optimizing a registration module and a feature extractor in episodic training. Novel-class training samples are included from the beginning, with sample synthesis used to compensate for their scarcity. The method is evaluated on Omniglot and miniImageNet under the claimed standard FSL protocol and under a generalized FSL setting where test labels come from both base and novel classes. The main claim is state-of-the-art accuracy under both settings.

Significance. The conceptual direction is useful: unlike standard meta-learning methods that only see base classes during training, the proposed method explicitly uses few-shot labeled data of the target classes. The registration idea, which couples episodic class means with global representations, is clean, and the extension to generalized FSL is practically motivated. The ablation study is clearly organized and gives some evidence about the contribution of the synthesis and registration components. However, the central empirical claim is invalidated by the evaluation protocol: the manuscript states that the test classes themselves are used as novel classes during training, which is not the protocol followed by the baselines in Tables 1 and 2. As a result, the reported state-of-the-art performance under standard FSL is not supported.

major comments (4)
  1. [Sec. 4.1.1 and Algorithm 1] The standard FSL comparison is invalid because the test classes are included in training. Section 4.1.1 states, "The reported performance is obtained by our approach trained with 64 training classes as base classes and 20 test classes as novel classes." Algorithm 1 samples training episodes from Ctotal = Cbase ∪ Cnovel, and Equations (4) and (5) compute the registration loss and query classification loss on those episodes. Thus the 20 miniImageNet test classes and their few-shot labels are observed throughout meta-training. This violates the protocol used by every baseline in Table 2, which never sees the test classes during training. The reported 53.21% and 72.34% accuracies are therefore not comparable to the baseline numbers, and the state-of-the-art claim in Section 4.1.4 is unsupported.
  2. [Sec. 4.1.3 (Omniglot)] The same protocol problem is either present or undisclosed for Omniglot. The paper follows the standard split of 1,200 training characters and 423 test characters, but the method requires few-shot labeled samples of novel classes to be part of the training set. Section 4.1.3 says that in a training episode images are randomly selected from the whole training set; it does not say whether the 423 test characters are used as novel classes during training. If they are, Table 1 has the same test-class leakage as the miniImageNet experiment; if they are not, the paper never explains where Omniglot novel-class training samples come from. Either way, the Omniglot results in Table 1 are not a valid standard-FSL comparison as reported.
  3. [Sec. 4.1.4] The reported 100% registration accuracy on test data cannot be used as evidence of generalization under the standard protocol. Because the global representations of the test classes have been optimized using images and labels from those classes during training, near-perfect registration at test time is expected. The observation is an artifact of the training protocol and does not support the claim that the registration module generalizes to unseen classes.
  4. [Sec. 4.2 and Table 3] The generalized FSL comparison is more defensible because the baselines are also retrained with few-shot novel-class training samples under the new split, but the manuscript gives insufficient implementation detail for those baselines (episode composition, hyperparameters, training length). The extreme accun values in Table 3 (0.52% for PN vs. 12.98% for the proposed method) suggest that the evaluation setups may not be fully aligned. The generalized-FSL claim should not be accepted until the baseline adaptation procedure is documented in enough detail to be reproducible and verifiable.
minor comments (4)
  1. [Algorithm 1, step 4] Step 4 says features are computed for i = 1, ..., n_q × n_train, but the support set defined in step 2 contains n_s × n_train images; the index should be n_s.
  2. [Eq. (3)] The notation k_r_hat ~ U(0, k_t) with k_r = ceil(k_r_hat) can produce k_r = 0 when k_r_hat is in (0,1), in which case the sum over k_r is undefined. The intended distribution over {1, ..., k_t} should be stated explicitly.
  3. [Table 2] The Relation Network baseline is cited as "RN [31]", but the correct reference is [32]; the citation conflicts with Prototypical Networks [31] used in Table 1.
  4. [Sec. 3.3] The text says the loss will update "global representations, and the parameters of the registration module and feature extractor," but it is not stated whether the data hallucinator is updated or kept fixed during the joint training; this should be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the global-representation training is a self-contained optimization; the test-class-in-training protocol is a benchmark-validity concern, not a circularity.

full rationale

I find no step in which a claimed prediction is equivalent by construction to an input or to a fitted constant. The objective in Eq. (6) jointly optimizes global class representations, the registration module, and the feature extractor, and the reported accuracies are measured on query images rather than on the training objective itself. The self-citations, e.g., [18] as background literature and [32] as a baseline method, are not used to justify the core mechanism of the paper. The most serious issue is a benchmark-protocol concern: Section 4.1.1 states that the reported performance is obtained by training with the 20 test classes as novel classes, and Algorithm 1 samples training episodes from Ctotal, so the Table 2 numbers are not comparable to standard FSL baselines that never see those classes during meta-training. That is a validity and correctness problem, not circularity under the definition used here, because the query accuracy is not forced by a fitted parameter renamed as a prediction and no load-bearing claim reduces to a self-citation chain.

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

The method relies on standard deep learning machinery and two domain assumptions about feature-space geometry. No new physical entities or fitted constants are introduced; the learned global representations are model parameters, not free constants fit to external data. Hyperparameters for sample synthesis are chosen by hand and not reported.

free parameters (3)
  • kt (number of augmented samples per novel class) = unspecified
    Set by hand in the sample synthesis module, Section 3.2; not reported in the paper.
  • kr (number of samples selected for convex combination) = uniformly sampled from 1..kt
    Random per synthesis step, Section 3.2 Eq. 3.
  • Episode size n_train = 30 (1-shot) / 20 (5-shot)
    Chosen for miniImageNet episodes, Section 4.1.4.
assumptions (3)
  • domain assumption Pretrained feature extractor on base classes transfers to novel classes
    The global representations are initialized from features of a classifier trained on base classes only (Section 4.1.2).
  • domain assumption Convex combinations of feature vectors in a class lie near the true class distribution
    The synthesis strategy in Section 3.2, Eq. 3, assumes this geometry to generate valid novel-class samples.
  • standard math Softmax over embedding distances is a differentiable proxy for prototype selection
    Used in the registration module, Section 3.1 Eq. 1 and Section 3.3 Eq. 4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Few-Shot Learning with Global Class Representations." pith.science (2026). https://pith.science/paper/5HJNBAXN

@misc{pith2026190805257,
  author       = {Pith},
  title        = {Pith review of: Few-Shot Learning with Global Class Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5HJNBAXN}},
  note         = {Machine review of arXiv:1908.05257}
}
read the original abstract

In this paper, we propose to tackle the challenging few-shot learning (FSL) problem by learning global class representations using both base and novel class training samples. In each training episode, an episodic class mean computed from a support set is registered with the global representation via a registration module. This produces a registered global class representation for computing the classification loss using a query set. Though following a similar episodic training pipeline as existing meta learning based approaches, our method differs significantly in that novel class training samples are involved in the training from the beginning. To compensate for the lack of novel class training samples, an effective sample synthesis strategy is developed to avoid overfitting. Importantly, by joint base-novel class training, our approach can be easily extended to a more practical yet challenging FSL setting, i.e., generalized FSL, where the label space of test data is extended to both base and novel classes. Extensive experiments show that our approach is effective for both of the two FSL settings.

Figures

Figures reproduced from arXiv: 1908.05257 by the authors.

Figure 1
Figure 1. An illustration of our approach. The first block shows a base class and a novel class in an embedding space. The base class contains sufficient labeled data while the novel class has only a few labeled data. The two classes have intersections, and we aim to learn global representations for each class which are used for rec￾ognizing test data. The second block illustrates the two key com￾ponents of the proposed model… view at source ↗
Figure 2
Figure 2. Overview of the whole framework. First, we propose a sample synthesis method to synthesize [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Ablation study for FSL on the miniImageNet dataset un [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 33 canonical work pages

  1. [1]

    Meta-learning with differentiable closed-form solvers

    Luca Bertinetto, Joao F Henriques, Philip HS Torr, and An- drea Vedaldi. Meta-learning with differentiable closed-form solvers. In International Conference on Learning Represen- tations, 2019. 2

    work page 2019

  2. [2]

    Henriques, Jack Valmadre, Philip H

    Luca Bertinetto, Joo F. Henriques, Jack Valmadre, Philip H. S. Torr, and Andrea Vedaldi. Learning feed-forward one- shot learners. In Advances in Neural Information Processing Systems, pages 523–531, 2016. 2

    work page 2016

  3. [3]

    Memory matching networks for one-shot im- age recognition

    Qi Cai, Yingwei Pan, Ting Yao, and Tao Mei Cheng- gang Yan. Memory matching networks for one-shot im- age recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4080–4088,

    work page

  4. [4]

    Synthesized classifiers for zero-shot learning

    Soravit Changpinyo, Wei-Lun Chao, Boqing Gong, and Fei Sha. Synthesized classifiers for zero-shot learning. In Pro- ceedings of the IEEE conference on computer vision and pat- tern recognition, pages 5327–5336, 2016. 7

    work page 2016

  5. [5]

    Predicting visual exemplars of unseen classes for zero-shot learning

    Soravit Changpinyo, Wei-Lun Chao, and Fei Sha. Predicting visual exemplars of unseen classes for zero-shot learning. In IEEE International Conference on Computer Vision , pages 3476–3485, 2017. 7

    work page 2017

  6. [6]

    An empirical study and analysis of generalized zero- shot learning for object recognition in the wild

    Wei-Lun Chao, Soravit Changpinyo, Boqing Gong, and Fei Sha. An empirical study and analysis of generalized zero- shot learning for object recognition in the wild. In European Conference on Computer Vision, pages 52–68, 2016. 7

    work page 2016

  7. [7]

    Good semi-supervised learning that re- quires a bad gan

    Zihang Dai, Zhilin Yang, William Cohen Fan Yang, and Rus- lan Salakhutdinov. Good semi-supervised learning that re- quires a bad gan. In Advances in Neural Information Pro- cessing Systems, pages 6510–6520, 2017. 1

    work page 2017

  8. [8]

    Harrison Edwards and Amos J. Storkey. Towards a neural statistician. In International Conference on Learning Repre- sentations, 2017. 1

    work page 2017

Show all 38 references
  1. [9]

    Model- agnostic meta-learning for fast adaptation of deep networks

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model- agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning , pages 1126–1135, 2017. 2, 6, 7

    2017

  2. [10]

    Dynamic few-shot visual learning without forgetting

    Spyros Gidaris and Nikos Komodakis. Dynamic few-shot visual learning without forgetting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition, pages 4367–4375, 2018. 1

    2018

  3. [11]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 2

    2016

  4. [12]

    Batch normalization: Accelerating deep network training by reducing internal co- variate shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal co- variate shift. In International Conference on Machine Learn- ing, pages 448–456, 2015. 6

    2015

  5. [13]

    Learning to remember rare events

    Lukasz Kaiser, Ofir Nachum, Aurko Roy, and Samy Bengio. Learning to remember rare events. In International Confer- ence on Learning Representations, 2017. 1

    2017

  6. [14]

    Siamese neural networks for one-shot image recognition

    Gregory Koch, Richard Zemel, and Ruslan Salakhutdinov. Siamese neural networks for one-shot image recognition. In ICML Workshop, 2015. 2

    2015

  7. [15]

    Semantic autoencoder for zero-shot learning

    Elyor Kodirov, Tao Xiang, and Shaogang Gong. Semantic autoencoder for zero-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 3174–3183, 2017. 7

    2017

  8. [16]

    Lake, Ruslan Salakhutdinov, Jason Gross, and Joshua B

    Brenden M. Lake, Ruslan Salakhutdinov, Jason Gross, and Joshua B. Tenenbaum. One shot learning of simple visual concepts. In Proceedings of the Annual Meeting of the Cog- nitive Science Society, 2011. 6

    2011

  9. [17]

    Boser, John S

    Yann LeCun, Bernhard E. Boser, John S. Denker, Donnie Henderson, Richard E. Howard, and Lawrence D. Hubbard, Wayne E.and Jackel. Backpropagation applied to handwrit- ten zip code recognition. Neural Computation , 1(4):541– 551, 1989. 6

    1989

  10. [18]

    Large-scale few-shot learning: Knowledge transfer with class hierarchy

    Aoxue Li, Tiange Luo, Zhiwu Lu, Tao Xiang, and Liwei Wang. Large-scale few-shot learning: Knowledge transfer with class hierarchy. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 7212– 7220, 2019. 1

    2019

  11. [19]

    Meta networks

    Tsendsuren Munkhdalai and Hong Yu. Meta networks. In In- ternational Conference on Machine Learning , pages 2554– 2563, 2017. 2

    2017

  12. [20]

    Reptile: a scalable metalearning algorithm

    Alex Nichol and John Schulman. Reptile: a scalable metalearning algorithm. arXiv preprint arXiv:1803.02999 ,

    arXiv

  13. [21]

    A unified approach for conventional zero-shot, generalized zero-shot and few-shot learning

    Shafin Rahman, Salman H Khan, and Fatih Porikli. A unified approach for conventional zero-shot, generalized zero-shot and few-shot learning. arXiv preprint arXiv:1706.08653 ,

    arXiv

  14. [22]

    Adaptive posterior learn- ing: few-shot learning with a surprise-based memory mod- ule

    Tiago Ramalho and Marta Garnelo. Adaptive posterior learn- ing: few-shot learning with a surprise-based memory mod- ule. In International Conference on Learning Representa- tions, 2019. 6

    2019

  15. [23]

    Optimization as a model for few-shot learning

    Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International Conference on Learn- ing Representations, 2016. 1, 2, 7

    2016

  16. [24]

    Tenenbaum, Hugo Larochelle, and Richard S

    Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B. Tenenbaum, Hugo Larochelle, and Richard S. Zemel. Meta-learning for semi-supervised few- shot classification. In International Conference on Learning Representations, 2018. 1

    2018

  17. [25]

    One-shot generalization in deep generative models

    Danilo Jimenez Rezende, Shakir Mohamed, Ivo Danihelka, Karol Gregor, and Daan Wierstra. One-shot generalization in deep generative models. In International Conference on Machine Learning, pages 1521–1529, 2016. 1

    2016

  18. [26]

    Bernardino Romera-Paredes and Philip H. S. Torr. An em- barrassingly simple approach to zero-shot learning. In In- ternational Conference on Machine Learning , pages 2152– 2161, 2015. 7

    2015

  19. [27]

    Berg, and Li Fei-Fei

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, San- jeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpa- thy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recogni- tion challenge. International Journal of Computer Visio...

    2015

  20. [28]

    Meta-learning with latent embedding optimization

    Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Had- sell. Meta-learning with latent embedding optimization. In International Conference on Learning Representations ,

  21. [29]

    Lillicrap

    Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy P. Lillicrap. One-shot learning with memory-augmented neural networks. arXiv preprint arXiv:1605.06065, 2016. 2

    2016 arXiv

  22. [30]

    Few-shot learning with graph neural networks

    Victor Garcia Satorras and Joan Bruna Estrach. Few-shot learning with graph neural networks. In International Con- ference on Learning Representations, 2017. 1

    2017

  23. [31]

    Swersky Kevin Snell, Jake and Richard S. Zemel. Prototyp- ical networks for few-shot learning. In Advances in Neural Information Processing Systems, pages 4080–4090, 2017. 2, 4, 6, 7, 8, 9

    2017

  24. [32]

    Flood Sung, Yongxin Yang, Li Zhang, Philip H. S. Torr Tao Xiang, and Timothy M. Hospedales. Learning to com- pare: Relation network for few-shot learning. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 1199–1208, 2018. 2, 6, 7, 8

    2018

  25. [33]

    Going deeper with convolutions

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 1–9, 2015. 1

    2015

  26. [34]

    Christian Szegedy, Vincent Vanhoucke Sergey Ioffe, and Alexander A. Alemi:. Inception-v4, inception-resnet and the impact of residual connections on learning. In Proceed- ings of the AAAI Conference on Artificial Intelligence, pages 4278–4284, 2017. 1

    2017

  27. [35]

    Zemel, and Raquel Urtasun

    Eleni Triantafillou, Richard S. Zemel, and Raquel Urtasun. Few-shot learning through an information retrieval lens. In Advances in Neural Information Processing Systems , pages 2255–2265., 2017. 6, 7

    2017

  28. [36]

    Matching networks for one shot learning

    Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Advances in Neural Information Pro- cessing Systems, pages 3630–3638, 2016. 2, 6, 7, 8

    2016

  29. [37]

    Low-shot learning from imaginary data

    Yu-Xiong Wang, Ross Girshick1, Martial Hebert, and Bharath Hariharan. Low-shot learning from imaginary data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7229–7238, 2018. 3, 6

    2018

  30. [38]

    Metagan: An adversarial ap- proach to few-shot learning

    Ruixiang Zhang, Tong Che, Zoubin Ghahramani, Yoshua Bengio, and Yangqiu Song. Metagan: An adversarial ap- proach to few-shot learning. In Advances in Neural Informa- tion Processing Systems, pages 2371–2380, 2018. 6, 7

    2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.