Pith. sign in

REVIEW 3 major objections 6 minor 73 references

DSS-Prompt: Dynamic-Static Synergistic Prompting for Few-Shot Class-Incremental Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read DSS-Prompt claims that a frozen ImageNet-pretrained Vision Transformer, given static prompts plus BLIP-generated instance-aware prompts in every block, tops all prior methods on four few-shot class-incremental benchmarks with zero training

desk verdict Solid FSCIL paper with a clean prompting design, but the SOTA claim is confounded—BLIP may be carrying the gain, and the missing BLIP-only baseline is the make-or-break experiment. read the letter →

arxiv 2508.09785 v1 pith:HWEEUWP2 submitted 2025-08-13 cs.CV

Linpu He , Yanan Li , Bingze Li , Elvis Han Cui , Donghui Wang This is my paper
classification cs.CV
keywords few-shotclass-incrementallearningprompttuningdynamicpromptsstaticinstance-awaremulti-modalknowledgevision-languagepre-trainingprototypeclassifiercatastrophicforgetting
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

Few-shot class-incremental learning (FSCIL) asks a model to absorb new classes from a handful of examples without losing the classes it already knows. This paper tries to establish that the problem can be largely dissolved by prompt design: keep an ImageNet-pretrained Vision Transformer completely frozen and give each of its blocks two kinds of prompts — static prompts that adapt the network to the downstream domain, and dynamic prompts generated per image from BLIP's visual and textual semantics, which carry the instance-level information that transfers from base to novel classes. With only these prompts trained in the initial base session, a training-free prototype classifier (each class is the mean embedding of its few samples) is claimed to beat every prior method on all four benchmarks tested, with the largest margin on the fine-grained CUB200 dataset. If correct, the result would mean catastrophic forgetting in FSCIL can be handled without rehearsal, regularization, or network growth, and incremental sessions reduce to storing a few mean vectors. The paper also reports that on its forgetting measure (performance drop) it places second or third among compared methods (Sec. 4.2, observation 4), a gap it flags for future work, which qualifies but does not retract the headline accuracy claims.

What carries the argument

The carrying object is the two-prompt bundle inserted into each ViT block. Static prompts $P^n_s$ follow deep visual prompt tuning: learnable tokens prepended at every layer, fixed after the base session, so they absorb the domain shift from ImageNet to the downstream benchmark. Dynamic prompts are produced on the fly by the Instance Semantic Perceptor (ISP) — BLIP's frozen image encoder, plus BLIP's image-grounded captioning followed by its text encoder — feeding the Dynamic Prompt Generator (DPG), a network with one encoder $g_e$ and a two-headed decoder $(g^p_d, g^\alpha_d)$ that emits prompt tokens $\tilde{P}^v_i$ and sigmoid-activated layer-wise scaling coefficients $\alpha_i$. The per-

What would settle it

Give DSS-Prompt a fine-grained test set on which BLIP's captions are known to be generic or wrong, such as a domain BLIP was not trained on, and compare the text-guided variant against the vision-guided-only ablation: the paper's mechanism predicts the text channel's contribution on CUB200 (84.49% vs 83.79% AVG with both channels) collapses or reverses exactly where caption quality fails. A cheaper computational check: feed randomized or mismatched captions through the same generator at test time and confirm that CUB200 accuracy falls toward the vision-only baseline.

Watch

Extended reading notes

Core claim

DSS-Prompt's central claim is that one frozen ViT-B/16 backbone becomes a state-of-the-art FSCIL classifier when every Transformer block receives two complementary prompt families. Static prompts (learnable, shared across images) close the gap between ImageNet pre-training and the downstream dataset. Dynamic prompts are generated per image from BLIP: its image encoder supplies visual knowledge, and its image-grounded captioning plus text encoder supplies textual knowledge; a small two-headed generator projects these into prompt tokens and per-layer scaling coefficients, so each layer receives a differently weighted instance-aware prompt. After base-session training of only the prompts, gener

Load-bearing premise

The load-bearing premise is that BLIP's image encoder and its captioning-plus-text-encoder supply reliable, instance-discriminative semantics for the benchmark images, and that those semantics survive the small two-headed projection into prompt tokens; if BLIP produces noisy or off-domain captions, the dynamic prompts should add noise rather than signal, eroding the reported gains over static-prompt baselines.

Editorial extensions

If this is right

  • Incremental sessions become training-free: after the base session, each new class is added simply by averaging the embeddings of its few shots, and the same frozen network classifies all past and new classes by cosine similarity.
  • Because the backbone is never updated, base-class knowledge cannot be overwritten during incremental sessions; the reported performance drop comes from base-session training choices, not from later adaptation.
  • Fine-grained datasets with very few samples per class (e.g., 60 per class in CUB200) benefit most from instance-aware prompts, suggesting the dynamic channel supplies discriminative detail that static adaptation misses.
  • The parameter budget shrinks to 1.61M trainable parameters versus 2.00M for the closest prompt-based rival, while reporting higher accuracy on the same backbone and benchmarks.
  • The design predicts that richer or better-aligned multi-modal knowledge extractors would further improve accuracy, since the dynamic prompt channel is only as informative as the semantics fed into it.

Reading between the lines

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

  • If the mechanism is exactly as described, the generator acts as a semantic bottleneck: shrinking its 256-dimensional latent code should degrade accuracy smoothly and predictably, a monotonicity the paper does not test.
  • A swap test follows from the architecture: replacing BLIP with another vision-language encoder or a purely visual extractor should change CUB200 accuracy more than CIFAR100 accuracy, because the paper attributes its fine-grained gains to instance-aware multi-modal semantics.
  • Because the dynamic prompts are computed from the input alone at inference, the same machinery could in principle serve open-set recognition or test-time adaptation with a frozen backbone, applications the paper does not mention.
  • The text-guided channel's value likely depends on caption quality; on domains where BLIP produces generic or incorrect captions (medical, satellite, or highly specialized imagery), the dynamic text prompts may add noise rather than signal, a boundary condition the paper leaves unexamined.
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 / 6 minor

Summary. The paper proposes DSS-Prompt, a few-shot class-incremental learning (FSCIL) method that freezes a ViT-B/16 pretrained on ImageNet and adds two types of prompt tokens in every transformer block: learnable static prompts for domain adaptation and input-dependent dynamic prompts produced by a small generator from BLIP-derived visual and textual embeddings. Only the prompts and generator are trained in the base session; in incremental sessions the model is frozen and classification is done with a cosine prototype classifier. The paper reports state-of-the-art average accuracies on CIFAR100, CUB200, miniImageNet, and ImageNet-R, with component ablations, scaling-coefficient analyses, and t-SNE/attention visualizations. The central claim is that this dynamic-static synergy allows training-free transfer to novel classes while using few trainable parameters.

Significance. If the empirical claims hold, the paper makes a useful contribution: it shows that a frozen ViT can be adapted with very few trainable parameters and that multi-modal BLIP knowledge can be injected through prompts, without any incremental training. The paper is clearly written, the setup follows standard FSCIL protocols, and the ablations confirm that each proposed component moves the average accuracy in the intended direction. However, the central SOTA claim is currently not fully supported because the experimental design does not isolate the prompting mechanism from the external BLIP prior, and the comparison to visual-only baselines is not fairness- or cost-controlled. These issues are addressable with additional experiments rather than being fundamental to the method.

major comments (3)
  1. [§3.2 (Step 1), Table 3] The ablation in Table 3 does not isolate the contribution of BLIP from that of the dynamic-prompt mechanism. The baseline row is a ViT-only prototype classifier, and the rows that add DP bundle (i) BLIP visual/text embeddings, (ii) the DPG, and (iii) layer-wise scaling. A reader cannot tell whether the gains come from the prompting machinery or from BLIP's external knowledge. Please add (a) a BLIP-feature baseline that classifies directly on v_i and/or t_i, with or without concatenation to the ViT [CLS] embedding, and (b) a control where the DPG is driven by a frozen encoder of comparable capacity trained only on ImageNet (or by randomly projected patch features). Without such controls, the statement in §4.3 that 'instance-aware prompts' are responsible for the improvement is not established.
  2. [§4.2, Tables 1 and 5; §3.2 Step 1] The state-of-the-art comparison is not apples-to-apples. DSS-Prompt is the only evaluated method that invokes a second frozen VLM (BLIP) at inference time, including an autoregressive captioning step for every image; the 1.61M trainable parameters in Table 5 exclude BLIP, understating both the computational cost and the amount of external pretrained knowledge. Since most competitors are visual-only, the reported margins (e.g., 1.42% AVG on CUB200) could reflect BLIP's web-scale class knowledge rather than the prompting design. I request a control: augment a strong visual-only baseline (e.g., ASP) with BLIP features, or replace the ISP with a non-multimodal pretrained encoder, and report inference FLOPs/latency including BLIP. Also, reporting whether generated captions contain downstream class names would clarify the zero-shot-prior risk.
  3. [§4.1, Table 1, Fig. 5] All accuracy numbers are single-run point estimates with no variance or error bars, and many baselines are quoted from prior papers. Since the claimed advantages over the strongest competitor are small (0.64% AVG on CIFAR100, 1.42% on CUB200), the ranking may not be statistically robust. Please report mean±std over at least three seeds for DSS-Prompt and, where feasible, rerun the closest competitor(s) under the identical protocol.
minor comments (6)
  1. [§5] Typo: 'promtps' should be 'prompts'. Also, Fig. 3 caption has 'a encoder' instead of 'an encoder'.
  2. [§3.2, Training Procedure] The text-branch parameters W_h', W_t, W_alpha' are mentioned in the Training Procedure but never formally defined; Eq. (4) only covers the vision branch. Please add the analogous equations or state the notation explicitly.
  3. [§4.1, Implementation Details] State which BLIP checkpoint/version is used and confirm that all BLIP modules are frozen during both training and inference.
  4. [§4.3, Fig. 6] The caption says 'two random classes'; please specify how the classes were chosen and whether the displayed distribution is representative.
  5. [References] References [26] and [27] have the same title and appear to be the same paper; please merge or disambiguate them.
  6. [Table 3] The full model has a higher PD (6.73) than the static-prompt-only model (4.58); since PD is a headline metric, a sentence explaining this trade-off (higher average accuracy but larger base-session drop) would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DSS-Prompt's claims are supported by held-out benchmark evaluation; the only self-citations are not load-bearing.

full rationale

The paper's derivation chain is: train static and dynamic prompts on the base session using cross-entropy loss; freeze the whole model; compute prototypes for each class from the frozen prompted embeddings; classify by cosine similarity. No parameter is fitted to the incremental-session test accuracy, and the dynamic-prompt scaling coefficients are learned only on base classes. The BLIP-derived multi-modal knowledge is an input to the prompt generator, but this is an architectural component, not a term that makes the reported accuracy equal to the training objective by construction. The only self-citations are [26] (related work) and [27] (a baseline in Table 1); neither is used to justify the core claim that DSS-Prompt outperforms state-of-the-art methods. The comparison against external benchmark results from other methods is a standard held-out evaluation. A possible concern that BLIP may provide privileged class knowledge is an experimental-fairness/leakage issue, not a circularity of the derivation, because the paper does not define its prediction in terms of the benchmark labels or use those labels to construct the prompts. Therefore, no circular step meets the evidence bar requiring an explicit reduction or a fitted parameter renamed as a prediction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method introduces trainable prompt and generator parameters plus several hand-chosen hyperparameters; it relies on external pre-trained models (ViT-B/16, BLIP) and the standard FSCIL evaluation protocol. No new physical or conceptual entities are postulated.

free parameters (4)
  • trainable prompt and generator parameters = ~1.61M parameters (Tab. 5)
    Includes static prompts, DPG weights (W_h, W_h', W_v, W_t, W_alpha, W_alpha') and classifier head W_z, all optimized on base session with CE loss.
  • prompt token length L_s = L_v = 10 (ImageNet-R), 3 (CIFAR100, CUB200, miniImageNet)
    Selected via ablation on CUB200 (Fig. 9 left); affects the trade-off between adaptation and overfitting.
  • latent dimension d' = 256
    FC layer reduces 768-dimensional BLIP features by 3x; architectural choice not justified by ablations.
  • training epochs = 60 (CUB200), 40 (others)
    Hand-chosen per dataset; longer training on CUB200 likely reflects smaller per-class data.
assumptions (4)
  • domain assumption ViT-B/16 pre-trained on ImageNet provides a transferable visual feature space.
    Used as frozen backbone throughout; Section 4.1 and Fig. 1. Its generalization ability is the basis for training-free transfer.
  • domain assumption BLIP's image encoder and captioning/text encoder yield complementary, reliable semantics for images in the target FSCIL benchmarks.
    The Instance Semantic Perceptor (Sec 3.2, Step 1) assumes BLIP outputs are useful input for generating dynamic prompts; poor captions would degrade the method.
  • domain assumption The standard FSCIL benchmark splits and evaluation protocols are correct and comparable across methods.
    Comparisons in Tab. 1 and Fig. 5 rely on prior papers reporting on the same fixed splits with ViT-B/16.
  • domain assumption Training only in the base session with a linear classifier and then freezing suffices to learn prompts that generalize to novel classes.
    Core design decision stated in Sec 3.2 and Algorithm 1; the paper follows prior work (TEEN, ASP) but does not justify it from first principles.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DSS-Prompt: Dynamic-Static Synergistic Prompting for Few-Shot Class-Incremental Learning." pith.science (2026). https://pith.science/paper/HWEEUWP2

@misc{pith2026250809785,
  author       = {Pith},
  title        = {Pith review of: DSS-Prompt: Dynamic-Static Synergistic Prompting for Few-Shot Class-Incremental Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HWEEUWP2}},
  note         = {Machine review of arXiv:2508.09785}
}
read the original abstract

Learning from large-scale pre-trained models with strong generalization ability has shown remarkable success in a wide range of downstream tasks recently, but it is still underexplored in the challenging few-shot class-incremental learning (FSCIL) task. It aims to continually learn new concepts from limited training samples without forgetting the old ones at the same time. In this paper, we introduce DSS-Prompt, a simple yet effective approach that transforms the pre-trained Vision Transformer with minimal modifications in the way of prompts into a strong FSCIL classifier. Concretely, we synergistically utilize two complementary types of prompts in each Transformer block: static prompts to bridge the domain gap between the pre-training and downstream datasets, thus enabling better adaption; and dynamic prompts to capture instance-aware semantics, thus enabling easy transfer from base to novel classes. Specially, to generate dynamic prompts, we leverage a pre-trained multi-modal model to extract input-related diverse semantics, thereby generating complementary input-aware prompts, and then adaptively adjust their importance across different layers. In this way, on top of the prompted visual embeddings, a simple prototype classifier can beat state-of-the-arts without further training on the incremental tasks. We conduct extensive experiments on four benchmarks to validate the effectiveness of our DSS-Prompt and show that it consistently achieves better performance than existing approaches on all datasets and can alleviate the catastrophic forgetting issue as well.

Figures

Figures reproduced from arXiv: 2508.09785 by the authors.

Figure 1
Figure 1. Comparison between (a) the naive FSCIL baseline [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The schematic diagram of our DSS-Prompt model. For each input image, the Instance Semantic Perceptor (ISP) first [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Details of our dynamic prompt generator. Given the multi-modal prior knowledge [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Illustration of the training procedure. On top of [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: FSCIL performance comparison with ViT-B/16 as the backbone on CIFAR100 (left), miniImageNet (middle) and [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Left: the scaling coefficient distribution across all [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: T-SNE Visualizations of feature embeddings on CUB200, obtained by the pre-trained ViT, the ViT-Prompt baseline [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Attention Rollout visualizations of the pre-trained [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

73 extracted references · 47 canonical work pages

  1. [26]

    Alex Krizhevsky, Geoffrey Hinton, et al. 2009. Learning multiple layers of features from tiny images.Master’s thesis, University of Tront(2009)

    work page 2009

  2. [27]

    Yanan Li, Linpu He, Feng Lin, and Donghui Wang. 2025. Few-Shot Class- Incremental Learning via Cross-Modal Alignment with Feature Replay. InPattern Recognition and Computer Vision. Springer Nature Singapore

    work page 2025

  3. [1]

    Samira Abnar and Willem Zuidema. 2020. Quantifying attention flow in trans- formers.arXiv preprint arXiv:2005.00928(2020)

    arXiv 2020

  4. [2]

    Aishwarya Agarwal, Biplab Banerjee, Fabio Cuzzolin, and Subhasis Chaudhuri

    work page

  5. [3]

    Noor Ahmed, Anna Kukleva, and Bernt Schiele. 2024. OrCo: Towards Better Generalization via Orthogonality and Contrast for Few-Shot Class-Incremental Learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 28762–28771

    work page 2024

  6. [4]

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. 2020. Unsupervised learning of visual features by contrasting cluster assignments.Advances in neural information processing systems33 (2020), 9912–9924

    2020

  7. [5]

    Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. 2022. Metafscil: A meta-learning approach for few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 14166–14175

    work page 2022

  8. [6]

    Marco D’Alessandro, Alberto Alonso, Enrique Calabrés, and Mikel Galar. 2023. Multimodal parameter-efficient few-shot class incremental learning. InProceed- ings of the IEEE/CVF International Conference on Computer Vision. 3393–3403

    work page 2023

Show all 73 references
  1. [7]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition. Ieee, 248–255

    2009

  2. [8]

    Mohammad Mahdi Derakhshani, Ivona Najdenkoska, Tom van Sonsbeek, Xi- antong Zhen, Dwarikanath Mahapatra, Marcel Worring, and Cees GM Snoek

  3. [9]

    Thang Doan, Sima Behpour, Xin Li, Wenbin He, Liang Gou, and Liu Ren. 2024. A streamlined Approach to Multimodal Few-Shot Class Incremental Learning for Fine-Grained Datasets.arXiv preprint arXiv:2403.06295(2024)

    2024 arXiv

  4. [10]

    InInterna- tional conference on medical image computing and computer-assisted intervention

    Lifelonger: A benchmark for continual disease classification. InInterna- tional conference on medical image computing and computer-assisted intervention. Springer, 314–324

  5. [11]

    Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. 2022. Scaling open- vocabulary image segmentation with image-level labels. InEuropean conference on computer vision. Springer, 540–557

    2022

  6. [12]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprin...

    2020 arXiv

  7. [13]

    Ziqi Gu, Chunyan Xu, Jian Yang, and Zhen Cui. 2023. Few-shot continual infomax learning. InProceedings of the IEEE/CVF International Conference on Computer Vision. 19224–19233

    2023

  8. [14]

    Dipam Goswami, Bartłomiej Twardowski, and Joost Van De Weijer. 2024. Cali- brating Higher-Order Statistics for Few-Shot Class-Incremental Learning with Pre-trained Vision Transformers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 4075–4084

    2024

  9. [15]

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

    2016

  10. [16]

    Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. 2022. A survey on vision transformer.IEEE transactions on pattern analysis and machine intelligence45, 1 (2022), 87–110

    2022

  11. [17]

    Michael Hersche, Geethan Karunaratne, Giovanni Cherubini, Luca Benini, Abu Sebastian, and Abbas Rahimi. 2022. Constrained few-shot class-incremental learning. InCVPR. 9057–9067

    2022

  12. [18]

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. 2021. The many faces of robustness: A critical analysis of out-of-distribution generalization. InProceedings of the IEEE/CVF internatio...

    2021

  13. [19]

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. 2022. Visual prompt tuning. InEuro- pean conference on computer vision. Springer, 709–727

    2022

  14. [20]

    Zitong Huang, Ze Chen, Zhixing Chen, Erjin Zhou, Xinxing Xu, Rick Siow Mong Goh, Yong Liu, Wangmeng Zuo, and Chunmei Feng. 2024. Learning prompt with distribution-based feature replay for few-shot class-incremental learning.arXiv preprint arXiv:2401.01598(2024)

    2024 arXiv

  15. [21]

    Muhammad Gul Zain Ali Khan, Muhammad Ferjad Naeem, Luc Van Gool, Didier Stricker, Federico Tombari, and Muhammad Zeshan Afzal. 2023. Introducing language guidance in prompt-based continual learning. InProceedings of the IEEE/CVF International Conference on Computer Vision. 11463–11473

    2023

  16. [22]

    Shilong Jia, Tingting Wu, Yingying Fang, Tieyong Zeng, Guixu Zhang, and Zhi Li. 2024. Purified Distillation: Bridging Domain Shift and Category Gap in Incremental Object Detection. InProceedings of the 32nd ACM International Conference on Multimedia. 1197–1205

    2024

  17. [23]

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 2013. 3d object repre- sentations for fine-grained categorization. InProceedings of the IEEE international conference on computer vision workshops. 554–561

    2013

  18. [24]

    Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. 2023. Maple: Multi-modal prompt learning. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. 19113–19122

    2023

  19. [25]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InInternational conference on machine learning. PMLR, 12888–12900

    2022

  20. [28]

    Yanan Li, Linpu He, Feng Lin, and Donghui Wang. 2024. Few-Shot Class- Incremental Learning via Cross-Modal Alignment with Feature Replay. InChinese Conference on Pattern Recognition and Computer Vision (PRCV). Springer, 19–33

    2024

  21. [29]

    Huan Liu, Li Gu, Zhixiang Chi, Yang Wang, Yuanhao Yu, Jun Chen, and Jin Tang

  22. [30]

    Chenxi Liu, Zhenyi Wang, Tianyi Xiong, Ruibo Chen, Yihan Wu, Junfeng Guo, and Heng Huang. 2025. Few-shot class incremental learning with attention-aware self-adaptive prompt. InEuropean Conference on Computer Vision. Springer, 1–18

    2025

  23. [31]

    Pratik Mazumder, Pravendra Singh, and Piyush Rai. 2021. Few-shot lifelong learning. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 2337–2345

    2021

  24. [32]

    InEuropean Conference on Computer Vision

    Few-shot class-incremental learning via entropy-regularized data-free replay. InEuropean Conference on Computer Vision. Springer, 146–162

  25. [33]

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi

  26. [34]

    Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. 2022. Few-shot class-incremental learning from an open-set perspective. InEuropean Conference on Computer Vision. Springer, 382–397

    2022

  27. [35]

    Wenhao Qiu, Sichao Fu, Jingyi Zhang, Chengxiang Lei, and Qinmu Peng. 2023. Semantic-visual guided transformer for few-shot class-incremental learning. In 2023 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2885– 2890

    2023

  28. [36]

    Keon-Hee Park, Kyungwoo Song, and Gyeong-Moon Park. 2024. Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 23881–23890

    2024

  29. [37]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library.Advances in neural information processing systems...

    2019

  30. [38]

    Sachin Ravi and Hugo Larochelle. 2017. Optimization as a model for few-shot learning. InInternational conference on learning representations

    2017

  31. [39]

    Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki

  32. [40]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on machine learnin...

    2021

  33. [41]

    Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. 2021. Do vision transformers see like convolutional neural networks?Advances in neural information processing systems34 (2021), 12116– 12128

    2021

  34. [42]

    Songsong Tian, Lusi Li, Weijun Li, Hang Ran, Li Li, and Xin Ning. 2024. Pl-fscil: Harnessing the power of prompts for few-shot class-incremental learning.arXiv preprint arXiv:2401.14807(2024)

    2024 arXiv

  35. [43]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research9, 11 (2008). MM ’25, October 27–31, 2025, Dublin, Ireland Linpu He, Yanan Li, Bingze Li, Elvis Han Cui, & Donghui Wang

    2008

  36. [44]

    Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie

  37. [45]

    James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira

  38. [46]

    Xuan Wang, Zhong Ji, Xiyao Liu, Yanwei Pang, and Jungong Han. 2025. On the Approximation Risk of Few-Shot Class-Incremental Learning. InEuropean Conference on Computer Vision. Springer, 162–178

    2025

  39. [47]

    Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. 2020. Few-shot class-incremental learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 12183–12192

    2020

  40. [48]

    Chen Xing, Negar Rostamzadeh, Boris Oreshkin, and Pedro O O Pinheiro. 2019. Adaptive cross-modal few-shot learning.Advances in neural information process- ing systems32 (2019)

    2019

  41. [49]

    Yuqiao Xu, Shucheng Huang, and Haoliang Zhou. 2024. CA-CLIP: category-aware adaptation of CLIP model for few-shot class-incremental learning.Multimedia Systems30, 3 (2024), 1–14

    2024

  42. [50]

    Boyu Yang, Mingbao Lin, Binghao Liu, Mengying Fu, Chang Liu, Rongrong Ji, and Qixiang Ye. 2021. Learnable expansion-and-compression network for few-shot class-incremental learning.arXiv preprint arXiv:2104.02281(2021)

    2021 arXiv

  43. [51]

    Boyu Yang, Mingbao Lin, Yunxiao Zhang, Binghao Liu, Xiaodan Liang, Rongrong Ji, and Qixiang Ye. 2022. Dynamic support network for few-shot class incremental learning.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 3 (2022), 2945–2951

    2022

  44. [52]

    Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye

  45. [53]

    Hantao Yao, Rui Zhang, and Changsheng Xu. 2023. Visual-language prompt tun- ing with knowledge-guided context optimization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 6757–6767

    2023

  46. [54]

    Zhidong Yu, Zhenbo Shi, Xiaoman Liu, and Wei Yang. 2024. PFFAA: Prototype- based Feature and Frequency Alteration Attack for Semantic Segmentation. In Proceedings of the 32nd ACM International Conference on Multimedia. 4562–4571

    2024

  47. [55]

    Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. 2022. Learning to prompt for continual learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 139–149

    2022

  48. [56]

    Jinghua Zhang, Li Liu, Olli Silvén, Matti Pietikäinen, and Dewen Hu. 2025. Few- Shot Class-Incremental Learning for Classification and Object Detection: A Sur- vey.IEEE Transactions on Pattern Analysis and Machine Intelligence(2025)

    2025

  49. [57]

    Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. 2023. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. InProceedings of the IEEE/CVF conference on computer vision and pattern recogni...

    2023

  50. [58]

    Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. 2023. Few-shot class-incremental learning via class-aware bilateral distillation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11838–11847

    2023

  51. [59]

    Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. 2024. Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need.International Journal of Computer Vision(2024), 1–21

    2024

  52. [60]

    Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. 2023. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning.arXiv preprint arXiv:2302.03004(2023)

    2023 arXiv

  53. [61]

    Da-Wei Zhou, Han-Jia Ye, Liang Ma, Di Xie, Shiliang Pu, and De-Chuan Zhan

  54. [62]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Conditional prompt learning for vision-language models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 16816–16825

    2022

  55. [63]

    Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu

  56. [64]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 12455–12464

  57. [69]

    Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. 2022. Forward compatible few-shot class-incremental learning. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9046–9056

    2022

  58. [71]

    Few-shot class-incremental learning by sampling multi-phase tasks.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 11 (2022), 12816– 12831

    2022

  59. [73]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Learning to prompt for vision-language models.International Journal of Computer Vision 130, 9 (2022), 2337–2348

    2022

  60. [2011]

    The caltech-ucsd birds-200-2011 dataset.CNS-TR-2011-001(2011)

    2011

  61. [2013]

    Fine-grained visual classification of aircraft.arXiv preprint arXiv:1306.5151 (2013)

    2013 arXiv

  62. [2021]

    Laion-400m: Open dataset of clip-filtered 400 million image-text pairs.arXiv preprint arXiv:2111.02114(2021)

    2021 arXiv

  63. [2022]

    InProceedings of the 30th ACM International Conference on Multimedia

    Semantics-Driven Generative Replay for Few-Shot Class Incremental Learning. InProceedings of the 30th ACM International Conference on Multimedia. 5246–5254

  64. [2023]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 11909–11919

  65. [2024]

    Advances in Neural Information Processing Systems36 (2024)

    Few-shot class-incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems36 (2024)

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.