Pith. sign in

REVIEW 3 major objections 5 minor 57 references

Revisiting Pool-based Prompt Learning for Few-shot Class-incremental Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper argues that prompt-pool methods degrade in few-shot class-incremental learning because token capacity saturates, and that shifting prompts to the spatial dimension restores performance.

desk verdict Interesting spatial-pooling idea with large empirical gains, but the core fusion equation is undefined on the page and the results lack error bars; deserves review after major clarification. read the letter →

arxiv 2507.09183 v2 pith:WDMKRH2Y submitted 2025-07-12 cs.CV

Yongwei Jiang , Yixiong Zou , Yuhua Li , Ruixuan Li This is my paper · ORCID
classification cs.CV
keywords few-shotclass-incrementallearningpromptpoolvisualtuningtoken-dimensionsaturationspatialpromptingfrequencydomainLGSP-PromptVisionTransformer
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 studies why prompt-pool methods like L2P and DualPrompt, which work well in ordinary continual learning, hurt performance in few-shot class-incremental learning (FSCIL). It claims the root cause is token-dimension saturation: with scarce novel-session data, too many prompt tokens compete to encode limited task-relevant information, and losing prompts encode background noise, shifting attention away from the object. To avoid this, the paper proposes LGSP-Prompt, which replaces token-dimension prompts with spatial prompts: local convolutional filters select fine-grained regions while global frequency-domain rings capture holistic structure. The method reports state-of-the-art average accuracies on CUB-200 (81.39%), FGVCAircraft (25.82%), and iNF200 (57.62%). If the diagnosis is right, prompt-pool design for FSCIL should move away from token-count scaling and toward spatially structured prompting.

What carries the argument

Token-dimension saturation is the paper's name for the situation where appended prompt tokens carry more learnable capacity than the scarce training data can populate, so additional prompts end up encoding background or irrelevant patterns and interfering with the CLS token's attention. The machinery that carries the proposed fix is a pair of spatial prompt pools: a local pool whose entries are generated by two-layer convolutional filters (producing dp-dimensional vectors per image) selected by cosine similarity, and a global pool of K differentiable concentric-ring masks applied in the frequency domain after a discrete Fourier transform. The two are combined via Eq. (10), $X_{final} = X + \alpha_l P_{local} + \alpha_g X_{global}$, so the model modifies image features in space and frequency rather than adding tokens.

What would settle it

A reproduction that implements Eq. (10) with the shapes defined in Eqs. (1) and (9) will fail at the addition step unless a broadcast rule is specified, so checking the released code settles whether the method as described is the one that ran. Independently, if adding non-learnable zero tokens to the token dimension does not reproduce the novel-session accuracy drop, then token-dimension saturation as information competition is not the cause.

Watch

Extended reading notes

Core claim

Pool-based prompt learning, when applied on top of VPT in FSCIL, causes novel-session accuracy to drop as the number of prompts grows, even though base-session accuracy is maintained. The paper attributes this to token-dimension saturation: the prompt pool is a set of appended tokens, and with only a few training samples the pool's information capacity exceeds what the data can fill, so prompts compete and some attend to task-irrelevant background. The proposed remedy is LGSP-Prompt, which constructs a pool of spatial prompts in the image domain—local prompts generated by small convolutional networks and global prompts built from concentric frequency rings—and integrates them by addition with learnable weights, avoiding the token dimension entirely. This yields higher accuracy across all sessions and on all three benchmarks.

Load-bearing premise

The integration formula in Eq. (10) assumes the local prompt and global frequency feature have the same shape as the image tensor and can be added term-by-term; if no broadcast rule exists, the described mechanism cannot be run.

Editorial extensions

If this is right

  • If token-dimension saturation is real, then simply enlarging prompt pools in the token dimension cannot help FSCIL; the optimal pool size is set by training-data volume.
  • Spatial prompting can sidestep token-dimension conflicts and scale with pool capacity, as shown by the accuracy rising with local pool size in the parameter analysis.
  • Combining local spatial and global frequency prompts provides complementary fine-grained and holistic cues, which is why LGSP-Prompt outperforms either component alone in the ablations.
  • The method preserves base-session knowledge while learning novel sessions, as demonstrated by higher base and novel accuracies across all sessions.
  • Existing prompt-pool methods (L2P, DualPrompt, CODA-Prompt) should be re-evaluated in FSCIL; their reported continual-learning gains do not transfer without modification.

Reading between the lines

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

  • A testable extension: if saturation is the true cause, then injecting random or zero-initialized padding tokens to increase token capacity without adding learnable parameters should not reproduce the degradation; if it does, the competition story is incomplete.
  • The frequency-ring global prompts can be read as a soft band-pass filter bank; future work might predict the optimal ring count from dataset spectral statistics rather than by grid search.
  • The same spatial-prompting idea could transfer to other low-data continual settings, such as domain-incremental or few-shot transfer, where token-dimension competition should similarly arise.
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 / 5 minor

Summary. This paper studies pool-based prompt learning in Few-Shot Class-Incremental Learning (FSCIL). The authors observe that integrating a prompt pool with VPT degrades novel-session accuracy, attribute this to token-dimension saturation, and propose LGSP-Prompt, which shifts prompting to the spatial dimension via a local spatial prompt pool and a global frequency-domain ring-mask branch. They report state-of-the-art results on CUB-200, FGVCAircraft, and iNF200, with ablations and visualizations supporting the design. The main contribution is the diagnosis and the proposed spatial prompting mechanism, but the paper currently leaves a load-bearing integration step underspecified.

Significance. If the method is reproducible, the paper makes a useful contribution: it is among the first to analyze why prompt-pool methods fail in FSCIL, it offers a plausible mechanism (token-dimension saturation) with qualitative attention-map evidence, and it reports consistent gains over strong baselines on three fine-grained benchmarks. The code release and structured ablation are positive. However, the significance is currently conditional on resolving the dimensional mismatch in the core integration equation and on clarifying whether the global branch really implements dynamic prompt selection. These issues do not invalidate the empirical results by themselves, but they prevent the method from being implemented from the text.

major comments (3)
  1. [§4.3, Eq. (10)] Equation (10), X_final = X + αl·Plocal + αg·Xglobal, is not well-defined as written. X is an image tensor in R^{B×C×H×W} (defined in §4.1.1), while Plocal from Eq. (5) is a weighted sum of dp-dimensional vectors Pi = fi(X) ∈ R^{dp} (Eq. (1)). In addition, Xglobal is never defined; the global branch produces F_enhanced(X) in Eq. (9) in the frequency domain, and no inverse DFT or spatial reshape is specified. Since every reported accuracy depends on this integration, the authors must specify the exact shapes, the reshape/broadcast rule, and the definition of Xglobal, or replace Eq. (10) with a well-formed operation.
  2. [§4.2, Eq. (9)] The paper claims that LGSP-Prompt constructs two spatial prompt pools enabling dynamic prompt selection, but the global branch does not perform selection. In §4.2, all K rings are always used, with softmax weights over the K ring masks in Eq. (9); there is no selection step analogous to Eq. (3)–(5). Either provide an explicit selection mechanism for the global pool or revise the abstract, introduction, and methodology text so that the claims match the actual mechanism.
  3. [§5.5, Fig. 10(c)] The local-pool-size ablation in Figure 10(c) reports overall accuracy around 68.5–71.5%, whereas Table 2 reports an average accuracy of 81.39% for LGSP-Prompt on CUB-200. The figure does not state which dataset or evaluation setting is used, making the parameter analysis hard to interpret. Please specify the benchmark and protocol, or correct the figure/axis labels if the numbers refer to a different configuration.
minor comments (5)
  1. [§4.1.2, Eq. (4)] Equation (4) uses the symbol k in the softmax, but the query was denoted q in Eq. (2); please make the notation consistent.
  2. [§4.2.4, Eq. (9)] The expression softmax(wkτ) is ambiguous: it should be specified whether τ multiplies wk, divides wk, or is part of a temperature-scaled softmax.
  3. [§4.3, Eq. (10)] Even if Plocal is intended to be reshaped into a spatial feature map, no reshaping or interpolation rule is stated; please define the intended operation.
  4. [§5.1] Many implementation details (optimizer, learning rates, training epochs, αl/αg initialization, VPT depth, and kernel sizes for LSP) are deferred to the supplementary material. Since the code is referenced, include the key hyperparameters in the main text for reproducibility.
  5. [§2.1] The cited works for spatial prompting (e.g., Segment Anything [16] and InsVP [22]) are not obviously about prompt-pool mechanisms; the related-work discussion should distinguish spatial attention/prompting from the pool-based selection proposed here.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the main defect (Eq. 10's undefined tensor addition) is a reproducibility gap, not circular reasoning.

full rationale

The paper's central derivation—that token-dimension saturation explains prompt-pool degradation in FSCIL and that shifting prompting to the spatial domain (LGSP-Prompt) improves accuracy—is not circular. The saturation analysis is an empirical interpretation supported by accuracy-versus-pool-size curves (Fig. 2) and independent attention-map visualizations (Figs. 3–4); the proposed method is then evaluated against external FSCIL baselines (Tables 2–4) with ablations (Table 5) that separate VPT, LSP, and GSP contributions. No fitted parameter is later renamed as a prediction, and no load-bearing claim is justified solely by self-citation; the self-citations (e.g., [51], [53], [57]) appear as related-work baselines or comparison methods, not as premises of the method. The main weakness—Eq. (10)'s undefined tensor integration, where Plocal is a dp-dimensional vector per Eq. (1) while X is B×C×H×W, and Xglobal is never defined—is a reproducibility/correctness gap, not circularity, because it does not make any result true by definition or reduce the method's outputs to its inputs.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim relies on a pre-trained ViT backbone and the FSCIL protocol. The method introduces hyperparameters (pool size, ring count, kernel sizes, temperatures, beta) that are tuned on validation data and not reported in full. No new physical entities are introduced.

free parameters (6)
  • Local prompt pool size M = 60 (default, from Fig. 10c)
    Pool capacity chosen by validation; accuracy increases with pool size but with fluctuations.
  • Number of selected local prompts ns = unspecified
    Not stated in the paper; required for Eq. (3).
  • Frequency ring count K = 100
    Tuned via Fig. 10a; optimal around 100 rings.
  • Kernel sizes for local CNN = small (1 and 7)
    Selected from Fig. 10b; small kernels favored.
  • Ring boundary sharpness beta = unspecified
    Controls sigmoid sharpness in Eq. (8); not reported.
  • Softmax temperature tau = unspecified
    Used in Eqs. (4) and (9); not reported.
assumptions (3)
  • standard math Discrete Fourier Transform and sigmoid ring masks are standard operations.
    Used without proof in Section 4.2.
  • domain assumption ImageNet-21K pre-trained ViT features transfer to CUB-200, iNF200, and FGVCAircraft; the three datasets do not overlap with pre-training classes.
    The paper excludes CIFAR and miniImageNet due to overlap, but the fine-grained datasets are assumed free of pre-training leakage.
  • domain assumption The FSCIL evaluation protocol (10 novel sessions, N-way K-shot) is the correct benchmark.
    Adopted from prior FSCIL work without justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Revisiting Pool-based Prompt Learning for Few-shot Class-incremental Learning." pith.science (2026). https://pith.science/paper/WDMKRH2Y

@misc{pith2026250709183,
  author       = {Pith},
  title        = {Pith review of: Revisiting Pool-based Prompt Learning for Few-shot Class-incremental Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WDMKRH2Y}},
  note         = {Machine review of arXiv:2507.09183}
}
read the original abstract

Few-Shot Class-Incremental Learning (FSCIL) faces dual challenges of data scarcity and incremental learning in real-world scenarios. While pool-based prompting methods have demonstrated success in traditional incremental learning, their effectiveness in FSCIL settings remains unexplored. This paper presents the first study of current prompt pool methods in FSCIL tasks, revealing an unanticipated performance degradation in incremental sessions. Through comprehensive analysis, we identify that this phenomenon stems from token-dimension saturation: with limited data, excessive prompts compete for task-relevant information, leading to model overfitting. Based on this finding, we propose LGSP-Prompt (Local-Global Spatial Prompting), which innovatively shifts pool-based prompt learning from the token dimension to the spatial dimension. LGSP-Prompt generates spatial prompts by synergistically combining local spatial features and global frequency-domain representations to highlight key patterns in input images. We construct two spatial prompt pools enabling dynamic prompt selection to maintain acquired knowledge while effectively learning novel sessions. Extensive experiments demonstrate that our approach achieves state-of-the-art performance across multiple FSCIL benchmarks, showing significant advantages in both base knowledge preservation and incremental learning. Our implementation is available at https://github.com/Jywsuperman/LGSP.

Figures

Figures reproduced from arXiv: 2507.09183 by the authors.

Figure 1
Figure 1. Traditional pool-based methods (L2P) stack multiple [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Class token attention maps with increasing number of [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Attention heat maps of prompts selected from pool. The [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Overview of LGSP-Prompt framework. Our approach leverages both local and global spatial information through three key com [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Visualization of relative attention weight distributions. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Visualization of local spatial prompting (P1-P18): Di [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 10
Figure 10. Figure 10: Ablation studies: (a) frequency ring numbers showing [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 9
Figure 9. Figure 9: Comparison of attention patterns: orig shows origi [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 51 canonical work pages

  1. [1]

    Touqeer Ahmad, Akshay Raj Dhamija, Steve Cruz, Ryan Ra- binowitz, Chunchun Li, Mohsen Jafarzadeh, and Terrance E. Boult. Few-shot class incremental learning leveraging self- supervised features. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3900–3910, 2022. 2

    work page 2022

  2. [2]

    Subspace regularizers for few-shot class in- cremental learning, 2022

    Afra Feyza Aky ¨urek, Ekin Aky¨urek, Derry Tanti Wijaya, and Jacob Andreas. Subspace regularizers for few-shot class in- cremental learning, 2022. 2

    work page 2022

  3. [3]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakan- tan, Pranav Shyam, Girish Sastry, Amanda Askell, Sand- hini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz...

    work page 2020

  4. [4]

    E2vpt: An effec- tive and efficient approach for visual prompt tuning

    Han Cheng, Wang Qifan, Cui Yiming, Cao Zhiwen, Wang Wenguan, Qi Siyuan, and Liu Dongfang. E2vpt: An effec- tive and efficient approach for visual prompt tuning. In In- ternational Conference on Computer Vision (ICCV) , 2023. 2

    work page 2023

  5. [5]

    Uncertainty-aware distillation for semi-supervised few-shot class-incremental learning

    Yawen Cui, Wanxia Deng, Haoyu Chen, and Li Liu. Uncertainty-aware distillation for semi-supervised few-shot class-incremental learning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2

    work page 2023

  6. [6]

    Lpt: Long-tailed prompt tuning for image classifica- tion, 2023

    Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng Zuo. Lpt: Long-tailed prompt tuning for image classifica- tion, 2023. 2

    work page 2023

  7. [7]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 2

    2021

  8. [8]

    Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N. Metaxas. Visual prompt tuning for test-time domain adaptation, 2022. 2

    work page 2022

Show all 57 references
  1. [9]

    Decoupling represen- tation and knowledge for few-shot intent classification and slot filling

    Jie Han, Yixiong Zou, Haozhao Wang, Jun Wang, Wei Liu, Yao Wu, Tao Zhang, and Ruixuan Li. Decoupling represen- tation and knowledge for few-shot intent classification and slot filling. In Proceedings of the AAAI Conference on Arti- ficial Intelligence, pages 18171–18179, 2024. 2

    2024

  2. [10]

    The inaturalist species classification and detection dataset

    Grant Van Horn, Oisin Mac Aodha, Yang Song, Alex Shep- ard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8769–8778, 2018. 6

    2018

  3. [11]

    Diversity- aware meta visual prompting, 2023

    Qidong Huang, Xiaoyi Dong, Dongdong Chen, Weiming Zhang, Feifei Wang, Gang Hua, and Nenghai Yu. Diversity- aware meta visual prompting, 2023. 2, 4

    2023

  4. [12]

    Learning prompt with distribution-based feature replay for few-shot class-incremental learning, 2024

    Zitong Huang, Ze Chen, Zhixing Chen, Erjin Zhou, Xinx- ing Xu, Rick Siow Mong Goh, Yong Liu, Wangmeng Zuo, and Chunmei Feng. Learning prompt with distribution-based feature replay for few-shot class-incremental learning, 2024. 2

    2024

  5. [13]

    Vi- sual prompt tuning, 2022

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi- sual prompt tuning, 2022. 1, 2, 6

    2022

  6. [14]

    Haeyong Kang, Jaehong Yoon, Sultan Rizky Hikmawan Madjid, Sung Ju Hwang, and Chang D. Yoo. On the soft- subnetwork for few-shot class incremental learning, 2023. 2

    2023

  7. [15]

    Warping the space: Weight space rotation for class- incremental few-shot learning

    Do-Yeon Kim, Dong-Jun Han, Jun Seo, and Jaekyun Moon. Warping the space: Weight space rotation for class- incremental few-shot learning. InThe Eleventh International Conference on Learning Representations, 2023. 2, 7

    2023

  8. [16]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick. Segment anything, 2023. 2, 4

    2023

  9. [17]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 6

    2009

  10. [18]

    General- ized and incremental few-shot learning by explicit learning and calibration without forgetting

    Anna Kukleva, Hilde Kuehne, and Bernt Schiele. General- ized and incremental few-shot learning by explicit learning and calibration without forgetting. In ICCV, 2021. 2

    2021

  11. [19]

    The power of scale for parameter-efficient prompt tuning, 2021

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning, 2021. 2

    2021

  12. [20]

    Few-shot class incre- mental learning with attention-aware self-adaptive prompt,

    Chenxi Liu, Zhenyi Wang, Tianyi Xiong, Ruibo Chen, Yihan Wu, Junfeng Guo, and Heng Huang. Few-shot class incre- mental learning with attention-aware self-adaptive prompt,

  13. [21]

    Pre-train, prompt, and predict: A systematic survey of prompting methods in natu- ral language processing, 2021

    Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hi- roaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natu- ral language processing, 2021. 2

    2021

  14. [22]

    InsVP: Efficient instance visual prompting from image itself

    Zichen Liu, Yuxin Peng, and Jiahuan Zhou. InsVP: Efficient instance visual prompting from image itself. In ACM Multi- media 2024, 2024. 2, 4

    2024

  15. [23]

    Recon- struction target matters in masked image modeling for cross- domain few-shot learning

    Ran Ma, Yixiong Zou, Yuhua Li, and Ruixuan Li. Recon- struction target matters in masked image modeling for cross- domain few-shot learning. In Proceedings of the AAAI Con- ference on Artificial Intelligence, pages 19305–19313, 2025. 2

    2025

  16. [24]

    Fine-grained visual classi- fication of aircraft, 2013

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classi- fication of aircraft, 2013. 6

    2013

  17. [25]

    Pseudo-set frequency refinement architecture for fine-grained few-shot class- incremental learning

    Zicheng Pan, Weichuan Zhang, Xiaohan Yu, Miaohua Zhang, and Yongsheng Gao. Pseudo-set frequency refinement architecture for fine-grained few-shot class- incremental learning. Pattern Recognition , 155:110686,

  18. [26]

    Pre-trained vision and language transformers are few-shot incremental learners, 2024

    Keon-Hee Park, Kyungwoo Song, and Gyeong-Moon Park. Pre-trained vision and language transformers are few-shot incremental learners, 2024. 1, 2, 6, 7, 8

    2024

  19. [27]

    Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C. Lovell. Few-shot class-incremental learning from an open- set perspective, 2022. 2

    2022

  20. [28]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2

    2021

  21. [29]

    Self- supervised knowledge distillation for few-shot learning

    Jathushan Rajasegaran, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Mubarak Shah. Self- supervised knowledge distillation for few-shot learning. https://arxiv.org/abs/2006.09785, 2020. 2

    2006 arXiv

  22. [30]

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental clas- sifier and representation learning, 2017. 7

    2017

  23. [31]

    Few-shot class incremental learning with generative feature replay

    Abhilash Shankarampeta and Koichiro Yamauchi. Few-shot class incremental learning with generative feature replay. In International Conference on Pattern Recognition Applica- tions and Methods, 2021. 2

    2021

  24. [32]

    Coda-prompt: Contin- ual decomposed attention-based prompting for rehearsal-free continual learning, 2023

    James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Contin- ual decomposed attention-based prompting for rehearsal-free continual learning, 2023. 2, 7

    2023

  25. [33]

    Rethinking few-shot class-incremental learning: Learning from yourself

    Yu-Ming Tang, Yi-Xing Peng, Jingke Meng, and Wei-Shi Zheng. Rethinking few-shot class-incremental learning: Learning from yourself. In European Conference on Com- puter Vision, 2024. 2, 7, 8

    2024

  26. [34]

    Few-shot class- incremental learning

    Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class- incremental learning. In Proceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition, 2020. 1, 2

    2020

  27. [35]

    Matching networks for one shot learning, 2017

    Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning, 2017. 6

    2017

  28. [36]

    Belongie

    Catherine Wah, Steve Branson, Peter Welinder, Pietro Per- ona, and Serge J. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical report, California Institute of Technology,

    2011

  29. [37]

    Foster: Feature boosting and compression for class- incremental learning, 2022

    Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class- incremental learning, 2022. 2, 7

    2022

  30. [38]

    Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub- optimality, 2023

    Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub- optimality, 2023. 2

    2023

  31. [39]

    Few-shot class-incremental learning via training-free prototype calibration, 2023

    Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye. Few-shot class-incremental learning via training-free prototype calibration, 2023. 7

    2023

  32. [40]

    On the approximation risk of few-shot class- incremental learning

    Xuan Wang, Zhong Ji, Xiyao Liu, Yanwei Pang, and Jun- gong Han. On the approximation risk of few-shot class- incremental learning. In European Conference on Computer Vision, 2024. 2, 7, 8

    2024

  33. [41]

    S-prompts learning with pre-trained transformers: An occam’s razor for domain incremental learning, 2023

    Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam’s razor for domain incremental learning, 2023. 2

    2023

  34. [42]

    Dualprompt: Com- plementary prompting for rehearsal-free continual learning,

    Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Dualprompt: Com- plementary prompting for rehearsal-free continual learning,

  35. [43]

    Learning to prompt for continual learning, 2022

    Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jen- nifer Dy, and Tomas Pfister. Learning to prompt for continual learning, 2022. 1, 2, 7

    2022

  36. [44]

    Dynamic sup- port network for few-shot class incremental learning

    Boyu Yang, Mingbao Lin, Yunxiao Zhang, Binghao Liu, Xi- aodan Liang, Rongrong Ji, and Qixiang Ye. Dynamic sup- port network for few-shot class incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):2945–2951, 2023. 2

    2023

  37. [45]

    Neural collapse inspired feature- classifier alignment for few-shot class-incremental learning

    Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature- classifier alignment for few-shot class-incremental learning. In ICLR, 2023. 7

    2023

  38. [46]

    Few-shot incremental learning with contin- ually evolved classifiers

    Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with contin- ually evolved classifiers. In IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2021. 2, 7

    2021

  39. [47]

    Mi- haylova

    Tianhao Zhang, Zhixiang Chen, and Lyudmila S. Mi- haylova. Pvit: Prior-augmented vision transformer for out- of-distribution detection, 2025. 2

    2025

  40. [48]

    Few-shot class- incremental learning via class-aware bilateral distillation

    Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class- incremental learning via class-aware bilateral distillation. In 2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 11838–11847, 2023. 2

    2023

  41. [49]

    Forward compatible few-shot class-incremental learning

    Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shil- iang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In CVPR, 2022. 2, 7

    2022

  42. [50]

    Few-shot class-incremental learn- ing by sampling multi-phase tasks

    Da-Wei Zhou, Han-Jia Ye, Liang Ma, Di Xie, Shiliang Pu, and De-Chuan Zhan. Few-shot class-incremental learn- ing by sampling multi-phase tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence , 45(11):12816– 12831, 2023. 2

    2023

  43. [51]

    Delve into base-novel confusion: Redundancy exploration for few-shot class-incremental learning

    Haichen Zhou, Yixiong Zou, Ruixuan Li, Yuhua Li, and Kui Xiao. Delve into base-novel confusion: Redundancy exploration for few-shot class-incremental learning. arXiv preprint arXiv:2405.04918, 2024. 7

    2024 arXiv

  44. [52]

    Self-promoted prototype refinement for few-shot class- incremental learning, 2021

    Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class- incremental learning, 2021. 2

    2021

  45. [53]

    Margin-based few-shot class-incremental learning with class-level overfitting mitigation

    Yixiong Zou, Shanghang Zhang, Yuhua Li, and Ruixuan Li. Margin-based few-shot class-incremental learning with class-level overfitting mitigation. Advances in neural infor- mation processing systems, 35:27267–27279, 2022. 2, 7

    2022

  46. [54]

    Flatten long-range loss landscapes for cross-domain few- shot learning

    Yixiong Zou, Yicong Liu, Yiman Hu, Yuhua Li, and Ruixuan Li. Flatten long-range loss landscapes for cross-domain few- shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23575– 23584, 2024. 2

    2024

  47. [55]

    Atten- tion temperature matters in vit-based cross-domain few-shot learning

    Yixiong Zou, Ran Ma, Yuhua Li, and Ruixuan Li. Atten- tion temperature matters in vit-based cross-domain few-shot learning. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  48. [56]

    A closer look at the CLS token for cross-domain few-shot learning

    Yixiong Zou, Shuai Yi, Yuhua Li, and Ruixuan Li. A closer look at the CLS token for cross-domain few-shot learning. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2

    2024

  49. [57]

    Compositional few-shot class-incremental learning

    Yixiong Zou, Shanghang Zhang, Haichen Zhou, Yuhua Li, and Ruixuan Li. Compositional few-shot class-incremental learning. arXiv preprint arXiv:2405.17022, 2024. 2, 7

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.