Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

EKPC: Elastic Knowledge Preservation and Compensation for Class-Incremental Learning

T0 review · 3 major / 4 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 EKPC reduces forgetting in class-incremental learning by selectively regularizing shared adapter parameters according to their importance and by compensating old class prototypes using a trainable semantic drift estimate.

desk verdict A solid 10/20-split CIL method with a strong ablation story, but the long-horizon Table 4 is not credible as reported and should not carry the 'advantage grows with task count' claim. read the letter →

arxiv 2506.12351 v1 pith:2OKLTMAD submitted 2025-06-14 cs.CV

Huaijie Wang , De Cheng , Lingfeng He , Yan Li , Jie Li , Nannan Wang , Xinbo Gao This is my paper
classification cs.CV
keywords knowledgecompensationmethodmethodssemanticacquiredadapterclass-incremental
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

Class-incremental learning asks a model to keep learning new classes over time without seeing old data again. Most modern approaches freeze a pre-trained vision model and only train a small adapter plus a classifier. Shared adapters are cheap but they forget old classes easily.

EKPC protects old knowledge in two ways. First, Importance-aware Parameter Regularization (IPR) measures which adapter parameters matter most for old classes. It looks at channel averages and variances of features, and at how strongly each adapter responds to an input, then heavily penalizes changes to important parameters while leaving others free. Second, Trainable Semantic Drift Compensation (TSDC) handles the fact that old class prototypes drift when new classes arrive. It estimates that drift from current training samples, shrinks the estimate during training so it is not too noisy, and uses the corrected prototypes to train a fresh unified classifier.

The paper reports that EKPC outperforms existing methods on ImageNetR, ImageNetA, CUB-200, CIFAR-100, and DomainNet, in both 10-task and 20-task splits, and shows large advantages on 50-task and 100-task settings. Ablations confirm that both components contribute, with TSDC giving the larger boost. No code is released, and some implementation details are left unspecified, so independent re-implementation would require some effort.

Extended reading notes

Core claim

The paper claims EKPC 'achieves superior performances to existing state-of-the-art methods' on five CIL benchmarks while adding no parameters at inference. The load-bearing empirical assertion is that combining importance-aware regularization (IPR) and trainable semantic drift compensation (TSDC) outperforms the shared-adapter baseline SSIAT on every reported setting, e.g., 80.60 vs 79.38 ALast on 10-split ImageNetR and 87.22 vs 85.11 on 10-split DomainNet, and that the advantage grows with task count.

Load-bearing premise

Section 3.3, Eq. 18 assumes the semantic drift of old class prototypes can be estimated as a weighted average of the feature drift of current-task samples: the displacement suffered by new-class features when the model is updated is assumed to transfer to old classes. If old and new classes shift in different directions, the compensated prototypes are wrong and the unified classifier is trained on incorrect samples. TSDC's ablation shows it contributes the larger part of the gain, so this assumption is load-bearing for the reported results.

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 / 4 minor

Summary. This manuscript proposes EKPC, a rehearsal-free class-incremental learning method built on a frozen ViT backbone with shared adapters. The first component, IPR, estimates per-parameter importance from global channel statistics and local perturbation sensitivity, then applies weighted L2 regularization to constrain adapter updates. The second component, TSDC, estimates the semantic drift of old-class prototypes from current-task feature changes, regularizes that drift estimate, and uses compensated prototypes to sample features for retraining a unified classifier. Experiments are reported on five CIL benchmarks at 10/20 splits, plus a long-horizon 50/100-split comparison, with ablations for each component, comparisons against alternative regularization and drift-compensation strategies, hyperparameter sensitivity, and a forgetting metric.

Significance. If the results hold, EKPC is an attractive practical method: it adds no inference-time parameters and shows consistent improvements over the shared-adapter baseline SSIAT at 10/20 splits across five benchmarks. The empirical core is broad and generally well executed: three seeds, per-component ablations, comparisons with uniform/amplitude/Fisher regularization, comparisons with static drift compensation, hyperparameter sweeps, and an average-forgetting analysis. However, the significance is moderated by two load-bearing weaknesses: the long-horizon results in Table 4 are presented without the protocol details and error bars used elsewhere, and the semantic-drift transfer assumption in Eq. (18) is unvalidated even though Table 5 shows that TSDC contributes the larger share of the reported gains.

major comments (3)
  1. [§4.2, Table 4] The long-horizon (50/100-split) results are reported without error bars or a statement of the number of seeds, unlike Tables 1-3, and the class/domain ordering, classifier-retraining schedule, and the direct baseline SSIAT are omitted. The 100-split DomainNet margin over the next-best method is 23.6 points (81.24 vs 57.60), whereas the 10-split margin over the next-best method is about 2 points; no mechanism in IPR/TSDC is given for this discontinuity. These omissions are load-bearing for the claim that EKPC's advantage grows with task count, and Table 4 should be rerun and reported under the same protocol as the rest of the paper, with error bars and SSIAT included.
  2. [§3.3, Eq. (18)] The semantic drift for an old class c is estimated as a similarity-weighted average of the displacement δ^{t-1→t}_i of current-task samples, with weights α_i computed from the distance of those samples to the old prototype f^{t-1}_c. This assumes that the displacement of new-class features is a good proxy for the displacement of old-class features; if old and new classes shift in different directions, the compensated prototypes used to train the unified classifier will be wrong. Because Table 5 shows that TSDC alone provides the larger share of the improvement (e.g., DomainNet ALast 86.24 in Idx3 vs 82.75 in Idx1), this assumption is load-bearing. Please validate the transfer assumption (e.g., by comparing compensated old prototypes against actual old-class features from a withheld subset, or by measuring per-class drift directions) or restrict the claim accordingly.
  3. [§3.2, Eqs. (3)-(7)] The global-importance derivation optimizes a linear classifier score s = w·x + b for a single channel, but the framework actually uses a cosine classifier (Eq. 21) on deep feature vectors, and Eq. (7) uses |f_c|/σ²_c without the 2λ factor or any channel-correlation term. As written, the derivation does not establish that this statistic is the optimal weight for the deployed classifier. Since the novelty claim rests on the importance algorithm, either align the derivation with the actual classifier or explicitly present the statistic as a heuristic whose validity is established by the ablations.
minor comments (4)
  1. [Throughout] There are several typos and formatting issues: "dowm-sampling" in Section 3.1, "T rainable" in Section 3.3, "DmainNet" in the Table 3 caption, "sampling sapce" in Algorithm 2, and "of of" in the contributions list in Section 1.
  2. [§3.2, Eq. (4)] The trade-off parameter λ appears in the derivation but is never given a value or used in the implementation; the authors should state whether it is a conceptual device or an actual hyperparameter.
  3. [§3.2, Eq. (2)] The squared differences of matrices in Eq. (2) should be defined as elementwise operations to avoid ambiguity.
  4. [§4.2, Fig. 3] The per-task accuracy curves in Figure 3 do not show error bars despite the paper reporting three seeds; adding them would help assess whether the apparent growing advantage over SSIAT is significant.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the IPR regularizer and TSDC drift compensation are computed from feature statistics and external measurements, not from the final accuracies they predict.

full rationale

Walked the derivation chain. IPR: Eq. 2 is a weighted L2 regularizer; the importance weights in Eqs. 7-17 are computed from (i) channel-wise mean/variance of current-task features extracted by the frozen previous adapter and (ii) perturbation sensitivity of individual adapter outputs, then fused via the hyperparameters eta1 and eta2. These are data-dependent statistics, not functions of the final accuracy, so the regularizer is a constraint rather than a disguised predictor. TSDC: Eq. 18 estimates old-prototype drift as a similarity-weighted average of current-sample feature displacement between the (t-1)-th and t-th feature extractors; Eq. 19 penalizes the drift norm toward zero; Eq. 20 adds the regularized drift to stored prototypes before Gaussian sampling and unified-classifier training in Eq. 23. This is an assumption-laden heuristic (the displacement of new-class features is assumed to transfer to old classes), but no term is fitted to the accuracy that is later reported, so it is not circular. Hyperparameters (eta1, eta2, w1, w2) are ablated with sensitivity analyses and are not renamed predictions. The only same-group citation with any load-bearing potential is Lu et al. (2024), used for the DomainNet 200-class selection protocol and as the VPT-NSP baseline; this is a dataset/convention citation, not a theorem or ingredient of EKPC, and other same-group citations are related-work background. Table 4's missing error bars and omitted SSIAT baseline, and the very large 100-split DomainNet margin, are reproducibility/protocol concerns rather than circularity. Verdict: no circular step found.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The empirical claim rests on common deep-learning assumptions (pre-trained ViT transfer, Gaussian class-conditional features for sampling) plus two method-specific assumptions: Gaussianity of channel activations (Section 3.2) and transferability of drift from new to old classes (Eq. 18). The IPR derivation also substitutes a linear classifier for the actual cosine classifier. These are heuristic assumptions with empirical support in the ablations, but they are not proven. No invented entities (new forces, dimensions, or objects) are introduced.

free parameters (6)
  • lambda
    Introduced in Eq. 4 as the variance trade-off in the objective; the final importance formula uses only the proportionality to mu/sigma^2, so lambda cancels and no value is needed. Stated as a derivation constant.
  • eta1 and eta2 = 100.0 and 1.0
    Scale factors fusing global and local importance in Eq. 17. Chosen by sweeping on Split-ImageNetR and Split-ImageNetA (Fig. 6).
  • w1 and w2 = 1.0 and 1.0
    Loss weights in Eq. 22. Tuned on Split-ImageNetR (Table 11).
  • s and m = 20.0 and 0.01
    Scale and margin of the cosine classifier (Eq. 21), following prior work.
  • Ns = not stated
    Number of sampled features per class for unified classifier training in Algorithm 2; the text never gives its value.
  • sigma_c in Eq. 18 = not stated
    Standard deviation of class c used in the drift weighting; the estimation procedure is not specified.
assumptions (5)
  • domain assumption Channel activations of a class are approximately Gaussian (invoked via the Central Limit Theorem).
    Used in Section 3.2, Global importance, to justify the mean/variance ratio in Eq. 6-7. This is asserted, not verified on the actual features.
  • domain assumption The classifier score is linear (w·x+b) for the importance derivation.
    Eq. 3 assumes a linear classifier, but the method uses a cosine classifier (Eq. 21). The derivation is a proxy.
  • domain assumption First-order Taylor approximations capture perturbation effects in adapters.
    Eq. 8-11 suppose small perturbations and ReLU-activated units; higher-order terms are dropped.
  • domain assumption Semantic drift of old prototypes can be transferred from current-task samples.
    Eq. 18 estimates old-prototype drift by weighting the feature drift of new-task samples. This is the key premise of TSDC and follows Yu et al. 2020, but is not proven for these settings.
  • domain assumption Class features are Gaussian for prototype sampling in the unified classifier.
    Section 3.4 samples features from N(fc, Σc); the covariance estimate is not specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EKPC: Elastic Knowledge Preservation and Compensation for Class-Incremental Learning." pith.science (2026). https://pith.science/paper/2OKLTMAD

@misc{pith2026250612351,
  author       = {Pith},
  title        = {Pith review of: EKPC: Elastic Knowledge Preservation and Compensation for Class-Incremental Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2OKLTMAD}},
  note         = {Machine review of arXiv:2506.12351}
}
read the original abstract

Class-Incremental Learning (CIL) aims to enable AI models to continuously learn from sequentially arriving data of different classes over time while retaining previously acquired knowledge. Recently, Parameter-Efficient Fine-Tuning (PEFT) methods, like prompt pool-based approaches and adapter tuning, have shown great attraction in CIL. However, these methods either introduce additional parameters that increase memory usage, or rely on rigid regularization techniques which reduce forgetting but compromise model flexibility. To overcome these limitations, we propose the Elastic Knowledge Preservation and Compensation (EKPC) method, integrating Importance-aware Parameter Regularization (IPR) and Trainable Semantic Drift Compensation (TSDC) for CIL. Specifically, the IPR method assesses the sensitivity of network parameters to prior tasks using a novel parameter-importance algorithm. It then selectively constrains updates within the shared adapter according to these importance values, thereby preserving previously acquired knowledge while maintaining the model's flexibility. However, it still exhibits slight semantic differences in previous knowledge to accommodate new incremental tasks, leading to decision boundaries confusion in classifier. To eliminate this confusion, TSDC trains a unified classifier by compensating prototypes with trainable semantic drift. Extensive experiments on five CIL benchmarks demonstrate the effectiveness of the proposed method, showing superior performances to existing state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2506.12351 by the authors.

Figure 1
Figure 1. (a) illustrates the expandable parameters ap [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of our proposed framework. (I) a backbone network with shared adapter tuning, (II) Importance [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The performance of each session on ImageNetR and CIFAR100 [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The task-by-task accuracy changing curves of our baseline(Adapter), Adapter+IPR, Adapter+IPR+TSDC(EKPC). The effectiveness of Global and Local Im￾portance. The IPR module comprises two key components: Global Importance (GI) and Local Importance(LI). To evaluate their e…
Figure 5
Figure 5. Figure 5: old and new task accuracy for incremental ses [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: hyper-parameter analysis of η1 and η2 on 10- tasks of S-ImageNetR and S-ImageNetA. bustness: The experimental results demonstrate that the proposed method exhibits high robustness to vari￾ations in w1 and w2, with performance fluctuations re￾maining within 0.33% when a…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Symbiosis-Inspired Knowledge Distillation for Incremental Object Detection

    cs.CV 2026-07 conditional novelty 6.0 of 10

    SIKD improves incremental object detection by distilling old-model responses in co-occurring, overlapping regions plus old-class rank structure, instead of discarding high-overlap predictions.

  2. CKAA: Cross-subspace Knowledge Alignment and Aggregation for Robust Continual Learning

    cs.CV 2025-07 conditional novelty 5.0 of 10

    CKAA aligns features and decision boundaries across task-specific subspaces and mixes adapters by task confidence, improving class-incremental learning accuracy under misidentified task labels.

Reference graph

Works this paper leans on

74 extracted references · 70 canonical work pages · cited by 2 Pith papers

  1. [1]

    In: Proceedings of the European conference on computer vision (ECCV), pp 139--154

    Aljundi R, Babiloni F, Elhoseiny M, Rohrbach M, Tuytelaars T (2018) Memory aware synapses: Learning what (not) to forget. In: Proceedings of the European conference on computer vision (ECCV), pp 139--154

    work page 2018

  2. [2]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 8218--8227

    Bang J, Kim H, Yoo Y, Ha JW, Choi J (2021) Rainbow memory: Continual learning with a memory of diverse samples. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 8218--8227

    work page 2021

  3. [3]

    Neural Networks 135:38--54

    Belouadah E, Popescu A, Kanellos I (2021) A comprehensive study of class incremental learning algorithms for visual tasks. Neural Networks 135:38--54

    work page 2021

  4. [4]

    In: European Conference on Computer Vision, Springer, pp 692--709

    Boschini M, Bonicelli L, Porrello A, Bellitto G, Pennisi M, Palazzo S, Spampinato C, Calderara S (2022) Transfer without forgetting. In: European Conference on Computer Vision, Springer, pp 692--709

    work page 2022

  5. [5]

    In: Proceedings of the IEEE/CVF international conference on computer vision, pp 9650--9660

    Caron M, Touvron H, Misra I, J \'e gou H, Mairal J, Bojanowski P, Joulin A (2021) Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision, pp 9650--9660

    work page 2021

  6. [6]

    In: Proceedings of the European conference on computer vision (ECCV), pp 532--547

    Chaudhry A, Dokania PK, Ajanthan T, Torr PH (2018) Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: Proceedings of the European conference on computer vision (ECCV), pp 532--547

    work page 2018

  7. [7]

    Advances in Neural Information Processing Systems 35:16664--16678

    Chen S, Ge C, Tong Z, Wang J, Song Y, Wang J, Luo P (2022) Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35:16664--16678

    work page 2022

  8. [8]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 23595--23604

    Cheng D, Xu Z, Jiang X, Wang N, Li D, Gao X (2024) Disentangled prompt representation for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 23595--23604

    work page 2024

Show all 74 references
  1. [9]

    IEEE transactions on pattern analysis and machine intelligence 44(7):3366--3385

    De Lange M, Aljundi R, Masana M, Parisot S, Jia X, Leonardis A, Slabaugh G, Tuytelaars T (2021) A continual learning survey: Defying forgetting in classification tasks. IEEE transactions on pattern analysis and machine intelligence 44(7):3366--3385

    2021

  2. [10]

    arXiv preprint arXiv:201011929

    Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, Zhai X, Unterthiner T, Dehghani M, Minderer M, Heigold G, Gelly S, et al (2020) An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:201011929

    2020

  3. [11]

    Advances in Neural Information Processing Systems 35:10629--10642

    Ermis B, Zappella G, Wistuba M, Rawal A, Archambeau C (2022) Memory efficient continual learning with transformers. Advances in Neural Information Processing Systems 35:10629--10642

    2022

  4. [12]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 11483--11493

    Gao Q, Zhao C, Sun Y, Xi T, Zhang G, Ghanem B, Zhang J (2023) A unified continual learning framework with general parameter-efficient tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 11483--11493

    2023

  5. [13]

    In: European Conference on Computer Vision, Springer, pp 89--106

    Gao X, Dong S, He Y, Wang Q, Gong Y (2024 a ) Beyond prompt learning: Continual adapter for efficient rehearsal-free continual learning. In: European Conference on Computer Vision, Springer, pp 89--106

    2024

  6. [14]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 28463--28473

    Gao Z, Cen J, Chang X (2024 b ) Consistent prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 28463--28473

    2024

  7. [15]

    International Journal of Computer Vision pp 1--20

    He L, Cheng D, Wang N, Gao X (2024) Exploring homogeneous and heterogeneous consistent label associations for unsupervised visible-infrared person reid. International Journal of Computer Vision pp 1--20

    2024

  8. [16]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 8340--8349

    Hendrycks D, Basart S, Mu N, Kadavath S, Wang F, Dorundo E, Desai R, Zhu T, Parajuli S, Guo M, et al (2021 a ) The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, ...

    2021

  9. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 15262--15271

    Hendrycks D, Zhao K, Basart S, Steinhardt J, Song D (2021 b ) Natural adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 15262--15271

    2021

  10. [18]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 831--839

    Hou S, Pan X, Loy CC, Wang Z, Lin D (2019) Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 831--839

    2019

  11. [19]

    In: International Conference on Machine Learning, PMLR, pp 2790--2799

    Houlsby N, Giurgiu A, Jastrzebski S, Morrone B, De Laroussilhe Q, Gesmundo A, Attariyan M, Gelly S (2019) Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning, PMLR, pp 2790--2799

    2019

  12. [20]

    ICLR 1(2):3

    Hu EJ, Shen Y, Wallis P, Allen-Zhu Z, Li Y, Wang S, Wang L, Chen W, et al (2022) Lora: Low-rank adaptation of large language models. ICLR 1(2):3

    2022

  13. [21]

    In: International Conference on Learning Representations

    Huang WC, Chen CF, Hsu H (2024) OVOR : OnePrompt with virtual outlier regularization for rehearsal-free class-incremental learning. In: International Conference on Learning Representations

    2024

  14. [22]

    In: European Conference on Computer Vision, Springer, pp 709--727

    Jia M, Tang L, Chen BC, Cardie C, Belongie S, Hariharan B, Lim SN (2022) Visual prompt tuning. In: European Conference on Computer Vision, Springer, pp 709--727

    2022

  15. [23]

    In: European Conference on Computer Vision, Springer, pp 163--179

    Kim Y, Li Y, Panda P (2024) One-stage prompt-based continual learning. In: European Conference on Computer Vision, Springer, pp 163--179

    2024

  16. [24]

    Proceedings of the national academy of sciences 114(13):3521--3526

    Kirkpatrick J, Pascanu R, Rabinowitz N, Veness J, Desjardins G, Rusu AA, Milan K, Quan J, Ramalho T, Grabska-Barwinska A, et al (2017) Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13):3521--3526

    2017

  17. [25]

    Placeholder Journal

    Krizhevsky A, Hinton G, et al (2009) Learning multiple layers of features from tiny images. Placeholder Journal

    2009

  18. [26]

    arXiv preprint arXiv:210408691

    Lester B, Al-Rfou R, Constant N (2021) The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:210408691

    2021

  19. [27]

    arXiv preprint arXiv:210100190

    Li XL, Liang P (2021) Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:210100190

    2021

  20. [28]

    IEEE transactions on pattern analysis and machine intelligence 40(12):2935--2947

    Li Z, Hoiem D (2017) Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12):2935--2947

    2017

  21. [29]

    Advances in Neural Information Processing Systems 35:109--123

    Lian D, Zhou D, Feng J, Wang X (2022) Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems 35:109--123

    2022

  22. [30]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 23638--23647

    Liang YS, Li WJ (2024) Inflora: Interference-free low-rank adaptation for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 23638--23647

    2024

  23. [31]

    arXiv preprint arXiv:240605658

    Lu Y, Zhang S, Cheng D, Xing Y, Wang N, Wang P, Zhang Y (2024) Visual prompt tuning in null space for continual learning. arXiv preprint arXiv:240605658

    2024

  24. [32]

    In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp 7765--7773

    Mallya A, Lazebnik S (2018) Packnet: Adding multiple tasks to a single network by iterative pruning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp 7765--7773

    2018

  25. [33]

    In: Proceedings of the European conference on computer vision (ECCV), pp 67--82

    Mallya A, Davis D, Lazebnik S (2018) Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In: Proceedings of the European conference on computer vision (ECCV), pp 67--82

    2018

  26. [34]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 45(5):5513--5533

    Masana M, Liu X, Twardowski B, Menta M, Bagdanov AD, Van De Weijer J (2022) Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(5):5513--5533

    2022

  27. [35]

    In: Psychology of Learning and Motivation, vol 24, Elsevier, pp 109--165

    McCloskey M, Cohen NJ (1989) Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of Learning and Motivation, vol 24, Elsevier, pp 109--165

    1989

  28. [36]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 11321--11329

    Ostapenko O, Puscas M, Klein T, Jahnichen P, Nabi M (2019) Learning to remember: A synaptic plasticity driven framework for continual learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 11321--11329

    2019

  29. [37]

    In: European Conference on Computer Vision, Springer, pp 382--397

    Peng C, Zhao K, Wang T, Li M, Lovell BC (2022) Few-shot class-incremental learning from an open-set perspective. In: European Conference on Computer Vision, Springer, pp 382--397

    2022

  30. [38]

    In: Proceedings of the IEEE/CVF international conference on computer vision, pp 1406--1415

    Peng X, Bai Q, Xia X, Huang Z, Saenko K, Wang B (2019) Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision, pp 1406--1415

    2019

  31. [39]

    In: International Conference on Learning Representations

    Qiao J, Zhang Z, Tan X, Chen C, Qu Y, Peng Y, Xie Y (2024) Prompt Gradient Projection for Continual Learning . In: International Conference on Learning Representations

    2024

  32. [40]

    In: International conference on machine learning, PMLR, pp 8748--8763

    Radford A, Kim JW, Hallacy C, Ramesh A, Goh G, Agarwal S, Sastry G, Askell A, Mishkin P, Clark J, et al (2021) Learning transferable visual models from natural language supervision. In: International conference on machine learning, PMLR, pp 8748--8763

    2021

  33. [41]

    Psychological review 97(2):285

    Ratcliff R (1990) Connectionist models of recognition memory: Constraints imposed by learning and forgetting functions. Psychological review 97(2):285

    1990

  34. [42]

    In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp 2001--2010

    Rebuffi SA, Kolesnikov A, Sperl G, Lampert CH (2017) icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp 2001--2010

    2017

  35. [43]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 23616--23626

    Roy A, Moulick R, Verma VK, Ghosh S, Das A (2024) Convolutional prompting meets language models for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 23616--23626

    2024

  36. [44]

    In: International conference on machine learning, PMLR, pp 4548--4557

    Serra J, Suris D, Miron M, Karatzoglou A (2018) Overcoming catastrophic forgetting with hard attention to the task. In: International conference on machine learning, PMLR, pp 4548--4557

    2018

  37. [45]

    Advances in neural information processing systems 30

    Shin H, Lee JK, Kim J, Kim J (2017) Continual learning with deep generative replay. Advances in neural information processing systems 30

    2017

  38. [46]

    arXiv preprint arXiv:230406027

    Smith JS, Hsu YC, Zhang L, Hua T, Kira Z, Shen Y, Jin H (2023 a ) Continual diffusion: Continual customization of text-to-image diffusion with c-lora. arXiv preprint arXiv:230406027

    2023

  39. [47]

    In: IEEE / CVF Conference on Computer Vision and Pattern Recognition , pp 11909--11919, doi:10.1109/CVPR52729.2023.01146

    Smith JS, Karlinsky L, Gutta V, Cascante-Bonilla P, Kim D, Arbelle A, Panda R, Feris R, Kira Z (2023 b ) CODA-Prompt : COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning . In: IEEE / CVF Conference on Computer Vision and Pattern Recognition , ...

    2023

  40. [48]

    In: International Conference on Machine Learning (ICML) Workshops, vol 2

    Soutif-Cormerais A, Masana M, Van de Weijer J, Twardowski B (2021) On the importance of cross-task features for class-incremental learning. In: International Conference on Machine Learning (ICML) Workshops, vol 2

    2021

  41. [49]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 23252--23262

    Tan Y, Zhou Q, Xiang X, Wang K, Wu Y, Li Y (2024) Semantically-shifted incremental adapter-tuning is a continual vitransformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 23252--23262

    2024

  42. [50]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 1706--1716

    Tang YM, Peng YX, Zheng WS (2023) When prompt-based incremental learning does not meet strong pretraining. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 1706--1716

    2023

  43. [51]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 12183--12192

    Tao X, Hong X, Chang X, Dong S, Wei X, Gong Y (2020) Few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 12183--12192

    2020

  44. [52]

    Placeholder Journal

    Wah C, Branson S, Welinder P, Perona P, Belongie S (2011) The caltech-ucsd birds-200-2011 dataset. Placeholder Journal

    2011

  45. [53]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 5265--5274

    Wang H, Wang Y, Zhou Z, Ji X, Gong D, Zhou J, Li Z, Liu W (2018) Cosface: Large margin cosine loss for deep face recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 5265--5274

    2018

  46. [54]

    Advances in Neural Information Processing Systems 36

    Wang L, Xie J, Zhang X, Huang M, Su H, Zhu J (2024) Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality. Advances in Neural Information Processing Systems 36

    2024

  47. [55]

    In: European Conference on Computer Vision, Springer, pp 631--648

    Wang Z, Zhang Z, Ebrahimi S, Sun R, Zhang H, Lee CY, Ren X, Su G, Perot V, Dy J, et al (2022 a ) Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision, Springer, pp 631--648

    2022

  48. [56]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 139--149

    Wang Z, Zhang Z, Lee CY, Zhang H, Sun R, Ren X, Su G, Perot V, Dy J, Pfister T (2022 b ) Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 139--149

    2022

  49. [57]

    In: European Conference on Computer Vision, Springer, pp 205--222

    Xiang X, Tan Y, Wan Q, Ma J, Yuille A, Hager GD (2022) Coarse-to-fine incremental few-shot learning. In: European Conference on Computer Vision, Springer, pp 205--222

    2022

  50. [58]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 3014--3023

    Yan S, Xie J, He X (2021) Der: Dynamically expandable representation for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 3014--3023

    2021

  51. [59]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 23219--23230

    Yu J, Zhuge Y, Zhang L, Hu P, Wang D, Lu H, He Y (2024) Boosting continual learning of vision-language models via mixture-of-experts adapters. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 23219--23230

    2024

  52. [60]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 6982--6991

    Yu L, Twardowski B, Liu X, Herranz L, Wang K, Cheng Y, Jui S, Weijer Jvd (2020) Semantic drift compensation for class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 6982--6991

    2020

  53. [61]

    In: International conference on machine learning, PMLR, pp 3987--3995

    Zenke F, Poole B, Ganguli S (2017) Continual learning through synaptic intelligence. In: International conference on machine learning, PMLR, pp 3987--3995

    2017

  54. [62]

    arXiv preprint arXiv:240608830

    Zhang D, Li Y, Cheng D, Wang N, Han J (2024) Center-sensitive kernel optimization for efficient on-device incremental learning. arXiv preprint arXiv:240608830

    2024

  55. [63]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 19148--19158

    Zhang G, Wang L, Kang G, Chen L, Wei Y (2023) Slca: Slow learner with classifier alignment for continual learning on a pre-trained model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 19148--19158

    2023

  56. [64]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 8552--8562

    Zhang R, Guo Z, Zhang W, Li K, Miao X, Cui B, Qiao Y, Gao P, Li H (2022) Pointclip: Point cloud understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 8552--8562

    2022

  57. [65]

    arXiv preprint arXiv:230203648

    Zhou DW, Wang QW, Qi ZH, Ye HJ, Zhan DC, Liu Z (2023 a ) Deep class-incremental learning: A survey. arXiv preprint arXiv:230203648

    2023

  58. [66]

    International Journal of Computer Vision pp 1--21

    Zhou DW, Cai ZW, Ye HJ, Zhan DC, Liu Z (2024 a ) Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need. International Journal of Computer Vision pp 1--21

    2024

  59. [67]

    International Journal of Computer Vision pp 1--21

    Zhou DW, Cai ZW, Ye HJ, Zhan DC, Liu Z (2024 b ) Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need. International Journal of Computer Vision pp 1--21

    2024

  60. [68]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 23554--23564

    Zhou DW, Sun HL, Ye HJ, Zhan DC (2024 c ) Expandable subspace ensemble for pre-trained model-based class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 23554--23564

    2024

  61. [69]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 16816--16825

    Zhou K, Yang J, Loy CC, Liu Z (2022 a ) Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 16816--16825

    2022

  62. [70]

    International Journal of Computer Vision 130(9):2337--2348

    Zhou K, Yang J, Loy CC, Liu Z (2022 b ) Learning to prompt for vision-language models. International Journal of Computer Vision 130(9):2337--2348

    2022

  63. [71]

    Neural Processing Letters pp 1--17

    Zhou Q, Xiang X, Ma J (2023 b ) Hierarchical task-incremental learning with feature-space initialization inspired by neural collapse. Neural Processing Letters pp 1--17

    2023

  64. [72]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 5871--5880

    Zhu F, Zhang XY, Wang C, Yin F, Liu CL (2021) Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 5871--5880

    2021

  65. [73]

    , " * write output.state after.block = add.period write newline

    ENTRY address archive author booktitle chapter doi edition editor eid eprint howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mi...

  66. [74]

    write newline

    " write newline "" before.all 'output.state := FUNCTION add.period duplicate empty 'skip "." * add.blank if FUNCTION if.digit duplicate "0" = swap duplicate "1" = swap duplicate "2" = swap duplicate "3" = swap duplicate "4" = swap duplicate "5" = swap duplicate "6" = swap dupl...

Pith tools

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