Pith. sign in

REVIEW 4 major objections 5 minor 47 references

Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning

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

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

Pith's one-line read A tripartite weight-space ensemble that interpolates the base, previous, and current classification heads, regularized by distillation on amplified few-shot data, achieves state-of-the-art accuracy in few-shot class-incremental learning…

desk verdict A solid empirical FSCIL paper with a genuinely new weight-space ensembling idea and strong ablations, but the ADKD mechanism is oversold and the SOTA margins rest on single runs without code. read the letter →

arxiv 2506.15720 v1 pith:26ZBENNO submitted 2025-06-04 cs.LG cs.CV

Juntae Lee , Munawar Hayat , Sungrack Yun This is my paper
classification cs.LGcs.CV
keywords few-shotclass-incrementallearningweight-spaceensembleknowledgedistillationdataamplificationcatastrophicforgettingmodelinterpolationcosineclassifiercontinual
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 argues that in few-shot class-incremental learning, fixing the feature extractor after base training limits adaptability to new classes. It proposes updating the entire model each session while preventing forgetting by constructing the classification head as a convex combination of three heads: the original base head, the previous session's head, and the current session's head. To keep the previous model's knowledge alive despite scarce data, the paper amplifies the few-shot training examples by mixing them and then applies knowledge distillation on this amplified data. The combined method, Tri-WE plus ADKD, reports new state-of-the-art average accuracies on miniImageNet, CUB200, and CIFAR100.

What carries the argument

The central object is the tripartite weight-space ensemble (Tri-WE), defined by the interpolation rule in Eq. (1): for each class, the final head weight is a convex combination of the corresponding class weights from the base head, the previous head, and the current all-class head, with learnable scalars $\alpha_1$ and $\alpha_2$ normalized by Eq. (2). Only the classification head is ensembled; the feature extractor is updated with a reduced learning rate. The companion mechanism is amplified data knowledge distillation (ADKD), which uses CutMix to amplify the $N$-way $K$-shot training set to $16NK$ samples per class and then applies feature-level and logit-level distillation losses from the frozen previous model. A third loss, $L_{\text{Cls-Old}}$, keeps the previous head consistent with the lightly updated feature extractor.

What would settle it

Measure linear mode connectivity between the base head, the previous head, and the current head after incremental training: if the interpolation path between any two heads passes through a region of substantially higher loss than either endpoint, or if the interpolated head's accuracy is no better than the better endpoint on a held-out set, the core Tri-WE claim is refuted. A second concrete test is to run the method on an architecture whose weights are known not to be linearly connected after fine-tuning and observe whether the accuracy advantage over a single-head baseline disappears.

Watch

Extended reading notes

Core claim

The central claim is that a model can be continually updated on only a few examples per class without catastrophic forgetting if its classification head is periodically rebuilt by weight-space ensembling of the base, immediately previous, and current session heads, and if the previous model is distilled on amplified few-shot data. Concretely, the head weight for a class is a normalized convex combination with two learnable scalars, and the feature extractor is lightly updated rather than frozen. The paper reports 70.62% average accuracy on miniImageNet (60.13% in the last session), surpassing prior methods while deploying a smaller classifier ensemble than BiDistill or SAVC and without knowing the total number of incremental classes in advance, unlike NC-FSCIL and OrCo.

Load-bearing premise

The three classification heads of the base, previous, and current sessions are close enough in weight space that a convex interpolation of their weights yields a classifier that is at least as good as any one of them, and this closeness is assumed rather than measured.

Editorial extensions

If this is right

  • If the head interpolation works as claimed, FSCIL methods no longer need to freeze the feature extractor after base training, opening the door to whole-model adaptation in low-data regimes.
  • The deployed model is a single classification head, so the approach avoids the test-time model-size growth of BiDistill and SAVC while still outperforming them.
  • The method does not require knowing the number of future incremental classes at base time, a constraint imposed by NC-FSCIL and OrCo, making it applicable to open-ended learning scenarios.
  • Data amplification by mixing, specifically CutMix, is shown to be a simple and effective way to make knowledge distillation useful when only a handful of examples per new class are available.
  • The learnable interpolation coefficients adapt per session, suggesting that the balance between base knowledge and new-class adaptability can be set automatically rather than tuned by hand.

Reading between the lines

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

  • The paper does not analyze linear mode connectivity between the three classification heads; if those heads do not lie on a near-convex connected loss surface for other architectures or datasets, the interpolation could fail, so the method's generality hinges on an empirically untested geometric property.
  • A natural extension beyond the paper is to make the interpolation weights class-conditioned or predicted from a small meta-model, rather than using only two global scalars, which could better handle heterogeneous old and new class distributions.
  • One testable implication is that ADKD with CutMix works because mixing forces the distilled representation to preserve locally invariant features; a different mixing scheme that destroys local structure should degrade performance, and this could be verified directly.
  • The Tri-WE idea could transfer to large-scale continual learning settings with more abundant data, where weight-space interpolation of successive checkpoints is already known to help, though the paper only demonstrates it in the few-shot regime.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes a few-shot class-incremental learning (FSCIL) method combining a tripartite weight-space ensemble (Tri-WE) and an amplified-data knowledge distillation loss (ADKD). Tri-WE interpolates the per-class classification-head weight vectors of the base, immediately previous, and current session models, with two learnable scalar coefficients, while keeping the feature extractor only lightly updated. ADKD augments the current session's few-shot data with CutMix and distills the previous model's features and logits on the amplified data. The base model is trained with ALICE-style techniques plus a geometric auxiliary classification head. Experiments on miniImageNet, CUB200, and CIFAR100 report state-of-the-art average accuracies, e.g., 70.62% on miniImageNet with a last-session accuracy of 60.13%, and ablation studies attribute gains to both Tri-WE and ADKD.

Significance. If the results are reproducible, the paper makes a useful empirical contribution: Tri-WE is a simple, low-cost mechanism that improves average accuracy by 2.70% over No-WE on miniImageNet, ADKD adds 1.57% over no data amplification, and the deployed model is smaller than BiDistill and SAVC. The ablation suite is more thorough than is typical for this area, and the comparison table is internally consistent. However, the central claims are currently supported only by single-run joint accuracies: no uncertainty estimates are reported, and the ADKD loss is not shown to preserve old-class knowledge specifically, as opposed to acting as a generic feature-smoothness regularizer. These issues need to be addressed before the stated conclusions can be accepted.

major comments (4)
  1. [Sec. 3.2, Eqs. (5)-(7); Table 5] ADKD is computed on D_amp^t, which is generated from the current session's new-class data D^t. The frozen teacher f^{t-1} was trained only on classes C^{0:t-1}, so its features and logits on these inputs are not a record of old-class decision boundaries. The manuscript nevertheless motivates ADKD as transferring generalized knowledge for previous classes and credits it with preventing catastrophic forgetting, but Table 5 reports only joint accuracy over all classes. The observed gain could come entirely from a generic feature/logit-smoothness regularizer on new-class data. Please either recharacterize ADKD as a feature-smoothness regularizer, or provide an ablation that isolates old-class accuracy (e.g., accuracy on C^{0:t-1} only, or the base/novel breakdown already used in Fig. 6) with and without the L_ADKD terms.
  2. [Tables 1-6 and Fig. 4] All reported accuracies come from a single run, with no multiple seeds or error bars. For 5-shot tasks, stochasticity in optimization and data sampling can plausibly shift results by more than the reported SOTA margins, such as the 1.82% average gain on miniImageNet. Please report means and standard deviations over at least three seeds, or explicitly justify a fully deterministic pipeline; without this, the state-of-the-art claim is not statistically grounded.
  3. [Sec. 3, Eqs. (3)-(4)] The prototype buffer M is described as storing "the averaged feature of the K examples per class," but Eq. (3) computes CE(h^{(t)}_phi(g^{(t)}_theta(p)), y) for p drawn from M, which is only meaningful if p is an input image. In contrast, Eq. (4) passes p directly to the classification head, consistent with p being a feature vector. Please clarify the representation stored in M and correct Eqs. (3)-(4) accordingly; the current inconsistency makes the optimization procedure ambiguous and harms reproducibility.
  4. [Sec. 4, Implementation details; Table 4; Fig. 5] Several key hyperparameters (gamma1=1.2, gamma2=10.0, the 16x amplification factor, the CutMix scheme, and the initial values of alpha1 and alpha2) are selected using the same three benchmark datasets on which state-of-the-art performance is claimed. The main text reports no validation protocol or sensitivity analysis for the loss weights, and the amplification-factor analysis is deferred to the supplementary material. Please provide an explicit model-selection protocol or report sensitivity of the results to these hyperparameters; otherwise the risk of selection bias on the test sets is not addressed.
minor comments (5)
  1. [Sec. 4, Implementation details] The implementation details refer to "Loss weights lambda1 and lambda2," while Eq. (8) uses gamma1 and gamma2; the notation should be unified.
  2. [Fig. 2] Figure 2 is labeled "10-way 5-shot miniImageNet," but the main FSCIL protocol for miniImageNet is described as 5-way 5-shot; please clarify whether Fig. 2 is a separate transfer-learning experiment and how it relates to the reported protocol.
  3. [Eq. (7)] The notation f^{(t)}_{theta,phi_{0:N_{t-1}}}(x) should explicitly define which slice of the output vector is used in the cross-entropy with the teacher's logits.
  4. [Abstract / Contributions] The contributions bullet "attain SOTA results across on miniImageNet" contains a grammatical error ("across on") and should be corrected.
  5. [Eq. (2)] Since alpha1 and alpha2 are learnable scalars initialized to 1.0 in every session, please state explicitly whether the normalized coefficients in Eq. (2) are recomputed at every optimizer step and how the special case alpha2=0 at t=1 is handled during backpropagation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Tri-WE and ADKD are model-construction choices, not predictions reduced to their inputs.

full rationale

The paper's central claims are empirical: Tri-WE interpolates classification-head weights via Eq. (1), ADKD adds feature and logit distillation losses in Eqs. (5)-(7), and the combined method is evaluated against external benchmarks. None of these components is defined in terms of the reported accuracy numbers, and no fitted constant is renamed as a prediction. The learnable scalars alpha1 and alpha2 are optimized per session; the fixed loss weights gamma1=1.2 and gamma2=10.0 are hyperparameters tuned on the same benchmarks, which is a potential overfitting risk but not a circular reduction. The ADKD mechanism is motivated as transferring previous-class knowledge, but Eqs. (6)-(7) are computed on D_amp^(t), amplified current-session new-class data, and Table 5 reports only joint accuracy; this is a limitation in mechanism validation, not circularity. No load-bearing self-citations are present, and the base session training borrows ALICE as an external baseline technique. The derivation chain is self-contained against external benchmarks, so the appropriate finding is no significant circularity.

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

The method's central gain rests on two domain assumptions: that weight-space interpolation of three session heads is valid (Eq. 1), and that distillation on amplified current-class data preserves old-class knowledge (Eq. 5). Both are supported only empirically, without error bars or a linear-mode-connectivity analysis. All scalar hyperparameters were tuned on the same benchmarks used for reporting SOTA.

free parameters (5)
  • Alpha_1 (alpha1) = learned per session, init 1.0
    Eq. (2); weights the base head contribution to the interpolated classifier; optimized on the few-shot training data.
  • Alpha_2 (alpha2) = learned per session, init 1.0 (set to 0 at t=1)
    Eq. (2); weights the previous-session head contribution; optimized on few-shot data.
  • Gamma_1 (gamma1) = 1.2
    Loss weight for L_Cls-Old in Eq. (8), empirically determined.
  • Gamma_2 (gamma2) = 10.0
    Loss weight for L_ADKD in Eq. (8), empirically determined.
  • Amplification factor = 16 x N x K
    D^(t) amplified to 16N*K examples per class for ADKD; chosen empirically.
assumptions (5)
  • domain assumption The FSCIL setting grants access only to current few-shot data plus one prototype per old class.
    Section 'Problem setting'; inherited from TOPIC/CEC, standard for the benchmark.
  • domain assumption Convex interpolation of the base, previous, and current classification heads in weight space yields a valid classifier that balances old and new knowledge.
    Section 3.1, Eq. (1); motivated by Wise-FT/Model Soup, validated only empirically in Table 2, no linear-mode-connectivity analysis.
  • domain assumption Updating the feature extractor with a strongly reduced learning rate preserves the feature space sufficiently to keep prototypes meaningful.
    Section 3.1 and implementation details (LR 0.001 vs 0.1); used for L_Cls-Old and the interpolation to work.
  • domain assumption Distilling the frozen previous model's outputs on amplified images of only the current new classes transfers knowledge relevant to old classes.
    Section 3.2, Eq. (5)-(7); D_amp is built only from D^(t), so old-class preservation via ADKD is an assumption, supported only indirectly by Table 4.
  • domain assumption CutMix is a valid amplifier for KD in FSCIL.
    Table 4 shows CutMix beats MixUp/CutOut/RandAug; the paper explains it qualitatively, with no formal justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning." pith.science (2026). https://pith.science/paper/26ZBENNO

@misc{pith2026250615720,
  author       = {Pith},
  title        = {Pith review of: Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/26ZBENNO}},
  note         = {Machine review of arXiv:2506.15720}
}
read the original abstract

Few-shot class incremental learning (FSCIL) enables the continual learning of new concepts with only a few training examples. In FSCIL, the model undergoes substantial updates, making it prone to forgetting previous concepts and overfitting to the limited new examples. Most recent trend is typically to disentangle the learning of the representation from the classification head of the model. A well-generalized feature extractor on the base classes (many examples and many classes) is learned, and then fixed during incremental learning. Arguing that the fixed feature extractor restricts the model's adaptability to new classes, we introduce a novel FSCIL method to effectively address catastrophic forgetting and overfitting issues. Our method enables to seamlessly update the entire model with a few examples. We mainly propose a tripartite weight-space ensemble (Tri-WE). Tri-WE interpolates the base, immediately previous, and current models in weight-space, especially for the classification heads of the models. Then, it collaboratively maintains knowledge from the base and previous models. In addition, we recognize the challenges of distilling generalized representations from the previous model from scarce data. Hence, we suggest a regularization loss term using amplified data knowledge distillation. Simply intermixing the few-shot data, we can produce richer data enabling the distillation of critical knowledge from the previous model. Consequently, we attain state-of-the-art results on the miniImageNet, CUB200, and CIFAR100 datasets.

Figures

Figures reproduced from arXiv: 2506.15720 by the authors.

Figure 1
Figure 1. Overview of the proposed method (FE: feature extrac￾tor, Cls. head: classification head). Through Tri-WE, at tth incre￾mental session, Cls. head is composed by interpolating the weights of the Cls. heads covering the classes until base, (t−1)th, tth ses￾sions, respectively. For ADKD regularization loss, the input few￾shot data are amplified to reliably mine generalized knowledge of the model of the previous (t − 1)t… view at source ↗
Figure 2
Figure 2. Few-shot (10-way 5-shot miniImageNet) transfer learning (TL) starting from the base session model on each session. Classification accuracy scores on incremented 10 classes (top) & 40 base classes (bottom) are shown for each session. 2. Related Works Class-incremental learning (CIL) aims to seamlessly in￾tegrate new classes without access to previously seen data, preventing the catastrophic forgetting problem. To ens… view at source ↗
Figure 3
Figure 3. The pipeline of the proposed method. At the tth incremental session S (t) , the model is initialized using the resulting model of session S (t−1). Then, a set of N-way K-shot few-shot examples D (t) and the prototype buffer M are given. The weights ϕ0, ϕold, ϕall of classification heads hϕ0 , h (t) ϕold , h (t) ϕall are interpolated with learnable scalars α1 and α2. The main loss LCls is computed on the resulting cl… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparative results on CUB200 and CIFAR100 datasets. For each dataset, the curve of the session-wise Acc (%) is first provided. And also, the base, last, average Acc are reported on our method comparing with three recent SOTA methods, NC-FSCIL [40], SAVC [29], YourSelf…
Figure 5
Figure 5. Figure 5: Analysis on learnable scalars α1 and α2 on mini￾ImageNet. (a) Comparison with multiple combinations of fixed (not-learnable) α1 and α2 in terms of averaged Acc over all the sessions, (b) session-wise learned α1 and α2 in ours. in [PITH_FULL_IMAGE:figures/full_fig_p007…
Figure 6
Figure 6. Figure 6: Analysis on the new class adaptability on CUB200. The average of accuracy scores are provided for base and novel classes, denoted by Base and New respectively. And, harmonic means are also compared. base and novel classes, and the harmonic mean (HM). Since most previou…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 41 canonical work pages

  1. [1]

    Orco: To- wards better generalization via orthogonality and contrast for few-shot class-incremental learning

    Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: To- wards better generalization via orthogonality and contrast for few-shot class-incremental learning. InIEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6

    work page 2024

  2. [2]

    Il2m: Class incremen- tal learning with dual memory

    Eden Belouadah and Adrian Popescu. Il2m: Class incremen- tal learning with dual memory. InIEEE Conf. Comput. Vis. Pattern Recog., 2019. 2

    work page 2019

  3. [3]

    Efficient lifelong learning with a- gem.arXiv preprint arXiv:1812.00420, 2018

    Arslan Chaudhry, Marc’Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a- gem.arXiv preprint arXiv:1812.00420, 2018. 2

    arXiv 2018

  4. [4]

    Incremental few-shot learn- ing via vector quantization in deep embedded space

    Kuilin Chen and Chi-Guhn Lee. Incremental few-shot learn- ing via vector quantization in deep embedded space. InInt. Conf. Learn. Represent., 2020. 6

    work page 2020

  5. [5]

    Synthesized feature based few-shot class- incremental learning on a mixture of subspaces

    Ali Cheraghian, Shafin Rahman, Sameera Ramasinghe, Pengfei Fang, Christian Simon, Lars Petersson, and Mehrtash Harandi. Synthesized feature based few-shot class- incremental learning on a mixture of subspaces. InInt. Conf. Comput. Vis., 2021. 6

    work page 2021

  6. [6]

    Metafscil: A meta-learning approach for few- shot class incremental learning

    Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. Metafscil: A meta-learning approach for few- shot class incremental learning. InIEEE Conf. Comput. Vis. Pattern Recog., 2022. 6

    work page 2022

  7. [7]

    Randaugment: Practical automated data augmentation with a reduced search space

    Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. InIEEE Conf. Comput. Vis. Pattern Recog. Worksh., 2020. 7, 8

    work page 2020

  8. [8]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. InIEEE Conf. Comput. Vis. Pattern Recog., 2009. 5

    work page 2009

Show all 47 references
  1. [9]

    Improved regular- ization of convolutional neural networks with cutout.arXiv preprint arXiv:1708.04552, 2017

    Terrance DeVries and Graham W Taylor. Improved regular- ization of convolutional neural networks with cutout.arXiv preprint arXiv:1708.04552, 2017. 7, 8

    2017 arXiv

  2. [10]

    Ensemble methods in machine learn- ing

    Thomas G Dietterich. Ensemble methods in machine learn- ing. InInternational Workshop on Multiple Classifier Sys- tems, 2000. 3

    2000

  3. [11]

    Few-shot class- incremental learning via relation knowledge distillation

    Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class- incremental learning via relation knowledge distillation. In AAAI, 2021. 6

    2021

  4. [12]

    Podnet: Pooled outputs distilla- tion for small-tasks incremental learning

    Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distilla- tion for small-tasks incremental learning. InEur. Conf. Com- put. Vis., 2020. 2

    2020

  5. [13]

    Neural network en- sembles.IEEE Trans

    Lars Kai Hansen and Peter Salamon. Neural network en- sembles.IEEE Trans. Pattern Anal. Mach. Intell., 12(10): 993–1001, 1990. 3

    1990

  6. [14]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InIEEE Conf. Comput. Vis. Pattern Recog., 2016. 5

    2016

  7. [15]

    Learning a unified classifier incrementally via re- balancing

    Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via re- balancing. InIEEE Conf. Comput. Vis. Pattern Recog., 2019. 2

    2019

  8. [16]

    Snapshot ensembles: Train 1, get m for free

    Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E Hopcroft, and Kilian Q Weinberger. Snapshot ensembles: Train 1, get m for free. InInt. Conf. Learn. Represent.2017. 3

    2017

  9. [17]

    S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning

    Jayateja Kalla and Soma Biswas. S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning. InEur. Conf. Comput. Vis., 2022. 1, 2, 3, 6, 8

    2022

  10. [18]

    On the soft- subnetwork for few-shot class incremental learning

    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. InInt. Conf. Learn. Represent., 2023. 2, 5, 6

    2023

  11. [19]

    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. InInt. Conf. Learn. Repre- sent., 2023. 2, 6

    2023

  12. [20]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5

    2009

  13. [21]

    Imagenet classification with deep convolutional neural net- works

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net- works. InAdv. Neural Inform. Process. Syst.2012. 3

    2012

  14. [22]

    Learning without forgetting

    Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Trans. Pattern Anal. Mach. Intell., 40(12):2935–2947,

  15. [23]

    Few-shot class-incremental learn- ing via entropy-regularized data-free replay

    Huan Liu, Li Gu, Zhixiang Chi, Yang Wang, Yuanhao Yu, Jun Chen, and Jin Tang. Few-shot class-incremental learn- ing via entropy-regularized data-free replay. InEur. Conf. Comput. Vis., 2022. 6

    2022

  16. [24]

    Closer: Towards better representation learning for few-shot class- incremental learning

    Junghun Oh, Sungyong Baik, and Kyoung Mu Lee. Closer: Towards better representation learning for few-shot class- incremental learning. InEur. Conf. Comput. Vis., 2024. 6

    2024

  17. [25]

    Few-shot class-incremental learning from an open- set perspective

    Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open- set perspective. InEur. Conf. Comput. Vis., 2022. 1, 2, 5, 6

    2022

  18. [26]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. InInt. Conf. Machine Learning, 2021. 3

    2021

  19. [27]

    icarl: Incremental clas- sifier and representation learning

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental clas- sifier and representation learning. InIEEE Conf. Comput. Vis. Pattern Recog., 2017. 2, 4, 6

    2017

  20. [28]

    Experience replay for continual learning

    David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lil- licrap, and Gregory Wayne. Experience replay for continual learning. 2019. 2

    2019

  21. [29]

    Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning

    Zeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng, Li Yuan, and Yonghong Tian. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. InIEEE Conf. Comput. Vis. Pattern Recog., 2023. 1, 2, 3, 5, 6

    2023

  22. [30]

    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. InIEEE Conf. Comput. Vis. Pattern Recog

  23. [31]

    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. InEur. Conf. Comput. Vis., 2024. 6

    2024

  24. [32]

    Few-shot class- incremental learning

    Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class- incremental learning. InIEEE Conf. Comput. Vis. Pattern Recog., 2020. 2, 3, 5, 6

    2020

  25. [33]

    Matching networks for one shot learning

    Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. InNeurIPS, 2016. 5, 8

    2016

  26. [34]

    The caltech-ucsd birds-200-2011 dataset

    Catherine Wah, Steve Branson, Peter Welinder, Pietro Per- ona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5

    2011

  27. [35]

    Learning to prompt for continual learning

    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. InIEEE Conf. Comput. Vis. Pattern Recog., 2022. 2

    2022

  28. [36]

    Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing in- ference time

    Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Re- becca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Ko- rnblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing ...

    2022

  29. [37]

    Robust fine-tuning of zero-shot models

    Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gon- tijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. InIEEE Conf. Comput. Vis. Pattern Recog., 2022. 2, 3

    2022

  30. [38]

    Large scale incre- mental learning

    Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incre- mental learning. InIEEE Conf. Comput. Vis. Pattern Recog.,

  31. [39]

    Der: Dynam- ically expandable representation for class incremental learn- ing

    Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynam- ically expandable representation for class incremental learn- ing. InIEEE Conf. Comput. Vis. Pattern Recog., 2021. 2

    2021

  32. [40]

    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. InInt. Conf. Learn. Represent., 2023. 2, 3, 5, 6

    2023

  33. [41]

    Cutmix: Regu- larization strategy to train strong classifiers with localizable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regu- larization strategy to train strong classifiers with localizable features. InInt. Conf. Comput. Vis., 2019. 4, 7, 8

    2019

  34. [42]

    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. InIEEE Conf. Comput. Vis. Pattern Recog., 2021. 1, 2, 3, 5, 6, 8

    2021

  35. [43]

    mixup: Beyond empirical risk minimiza- tion.arXiv preprint arXiv:1710.09412, 2017

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimiza- tion.arXiv preprint arXiv:1710.09412, 2017. 4, 7, 8

    2017 arXiv

  36. [44]

    Maintaining discrimination and fairness in class incremental learning

    Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu- Tao Xia. Maintaining discrimination and fairness in class incremental learning. InIEEE Conf. Comput. Vis. Pattern Recog., 2020. 2

    2020

  37. [45]

    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 IEEE Conf. Comput. Vis. Pattern Recog., 2023. 1, 2, 3, 5, 6, 8

    2023

  38. [46]

    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. InIEEE Conf. Comput. Vis. Pat- tern Recog., 2022. 6, 8

    2022

  39. [47]

    Gkeal: Gaussian kernel embedded analytic learning for few-shot class incremental task

    Huiping Zhuang, Zhenyu Weng, Run He, Zhiping Lin, and Ziqian Zeng. Gkeal: Gaussian kernel embedded analytic learning for few-shot class incremental task. InIEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 6

    2023

Pith tools

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