arxiv: 2604.02765 · v1 · submitted 2026-04-03 · 💻 cs.LG

Towards Realistic Class-Incremental Learning with Free-Flow Increments

Zhiming Xu , Baile Xu , Jian Zhao , Furao Shen , Suorong Yang This is my paper

Pith reviewed 2026-05-13 19:49 UTC · model grok-4.3

classification 💻 cs.LG

keywords class-incremental learningcontinual learningfree-flow incrementsclass-wise mean objectivereplay methodscatastrophic forgettingdynamic weight alignment

0 comments

The pith

A class-wise mean objective stabilizes incremental learning when new classes arrive in unpredictable batches.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard class-incremental learning assumes fixed-size tasks where each step adds the same number of new classes. Real data streams rarely follow this pattern, so the paper formalizes free-flow class-incremental learning in which the number of new classes per arrival can vary arbitrarily. Under this setting many existing methods degrade because their loss weighting favors classes with more samples and their distillation or contrastive terms become unstable with small increments. The authors replace frequency-weighted losses with a class-wise mean objective that supplies uniform supervision to every class, then add targeted fixes such as limiting distillation to replayed samples, scaling contrastive and transfer losses, and introducing dynamic intervention weight alignment to counter unstable batch statistics. These changes restore performance across representative CIL approaches without requiring architecture modifications.

Core claim

The paper claims that a model-agnostic framework built around a class-wise mean objective and a small set of method-wise adjustments enables stable class-incremental learning when classes arrive in highly variable numbers. The class-wise mean objective discards sample-frequency weighting in favor of uniform aggregation of class-conditional supervision. Additional adjustments constrain distillation to replayed data, normalize the scale of contrastive and knowledge-transfer losses, and apply dynamic intervention weight alignment to prevent over-correction from small-increment statistics. Experiments show that standard baselines suffer clear drops under free-flow conditions while the proposed策略

What carries the argument

The class-wise mean objective that replaces sample-frequency weighted loss with uniformly aggregated class-conditional supervision.

If this is right

Existing CIL methods exhibit clear performance degradation once increment sizes become variable.
The class-wise mean objective plus the three adjustments produce consistent gains across multiple representative CIL paradigms.
The framework operates without architecture changes and remains compatible with replay, regularization, and architecture-based methods.
Constraining distillation to replayed data and applying dynamic weight alignment specifically counter instabilities from small class increments.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same uniform-supervision principle may transfer to other streaming settings where data batches arrive with irregular class or label distributions.
Improved stability from the class-wise mean could allow smaller replay buffers while maintaining performance, a possibility the paper does not test.
Real-world deployments such as continuous image labeling from user uploads would be natural testbeds for the free-flow formulation.

Load-bearing premise

Uniform class-conditional supervision together with the listed adjustments will stabilize learning across arbitrary increment sizes without creating new instabilities or requiring model changes.

What would settle it

Run the proposed framework and a standard replay baseline on CIFAR-100 while drawing the number of new classes per step from a wide random distribution and measure whether the accuracy gap between them remains closed or re-opens.

Figures

Figures reproduced from arXiv: 2604.02765 by Baile Xu, Furao Shen, Jian Zhao, Suorong Yang, Zhiming Xu.

**Figure 1.** Figure 1: Illustration of FFCIL. (a) Unlike equal-size tasks, FFCIL allows variable perstep class increments. (b) Existing CIL methods experience a substantial accuracy drop under FFCIL, even with the same number of classes and learning stages. Existing CIL paradigms can be broadly categorized into replay-based methods [30, 32] that store or generate representative samples, regularization-based and distillation-ba… view at source ↗

**Figure 2.** Figure 2: The proposed strategies for FFCIL. Class-wise mean loss enforces classinvariant updates, mitigating instability caused by free-flow class exposure. Replay-only distillation excludes new-class samples, reducing sensitivity to free-flow class arrivals. Objectives whose magnitudes depend on the sample or the activated class space are scale-normalized. The dynamic weight alignment scheme regulates calibration… view at source ↗

**Figure 3.** Figure 3: BiC confusion matrices on CIFAR-100 for equal-split CIL, Free-Flow with original method, and Free-Flow with our framework. identical to the number of tasks in the standard benchmark. For each dataset, we run the following experiments: baselines under standard CIL with equal splits (Equ.T), FFCIL using the original method (FF.org), and the variant equipped with our framework (FF.ours). These results are sum… view at source ↗

**Figure 4.** Figure 4: Impact of FFCIL step schedules on CIFAR-100: (a) iCaRL and (b) DER under ascending, descending, and highly fluctuating schedules. 1 2 3 4 5 6 7 8 9 10 Learning Step 35 45 55 65 75 Accuracy (%) original ours (a) DER 1 2 3 4 5 6 7 8 9 10 Learning Step 0 25 50 75 100 Accuracy (%) original ours (b) TagFex [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

**Figure 5.** Figure 5: Step-wise accuracy on CIFAR-100 under an extreme FFCIL schedule, with 90 classes introduced initially, followed by 1–2 classes per step. the second step that contains small class increments. Notably, TagFex exhibits a near-collapse behavior, with accuracy degrading to around 1%. In contrast, our method effectively mitigates this issue and maintains stable performance under such extreme step schedules. 5.5 … view at source ↗

**Figure 6.** Figure 6: Training-time study of each component. (a) CWM loss on three baselines. (b) Other components (Replay.Dist: replay-only distillation on iCaRL; TagFex.CWM: TagFex with CWM). (c) Weight-alignment time with and w/o DIWA. to a slight reduction in runtime, while scale normalization introduces only a negligible time increase. These results indicate that our proposed framework does not introduce additional comput… view at source ↗

read the original abstract

Class-incremental learning (CIL) is typically evaluated under predefined schedules with equal-sized tasks, leaving more realistic and complex cases unexplored. However, a practical CIL system should learns immediately when any number of new classes arrive, without forcing fixed-size tasks. We formalize this setting as Free-Flow Class-Incremental Learning (FFCIL), where data arrives as a more realistic stream with a highly variable number of unseen classes each step. It will make many existing CIL methods brittle and lead to clear performance degradation. We propose a model-agnostic framework for robust CIL learning under free-flow arrivals. It comprises a class-wise mean (CWM) objective that replaces sample frequency weighted loss with uniformly aggregated class-conditional supervision, thereby stabilizing the learning signal across free-flow class increments, as well as method-wise adjustments that improve robustness for representative CIL paradigms. Specifically, we constrain distillation to replayed data, normalize the scale of contrastive and knowledge transfer losses, and introduce Dynamic Intervention Weight Alignment (DIWA) to prevent over-adjustment caused by unstable statistics from small class increments. Experiments confirm a clear performance degradation across various CIL baselines under FFCIL, while our strategies yield consistent gains.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper names the free-flow arrival setting for class-incremental learning and shows that a few lightweight changes to the loss and distillation steps reduce the damage from uneven increments.

read the letter

The core contribution is the formalization of FFCIL, where new classes arrive in batches of arbitrary size rather than the tidy equal-sized tasks common in benchmarks. The authors demonstrate that standard CIL methods lose ground under this regime, which matches the practical mismatch between lab setups and real streams. They then introduce a class-wise mean objective that replaces frequency-weighted supervision with uniform per-class aggregation, plus three targeted adjustments: distillation limited to replay data, loss-scale normalization, and Dynamic Intervention Weight Alignment to dampen unstable statistics from small increments. These pieces are presented as model-agnostic and produce consistent recovery across representative baselines in the reported experiments. The changes are simple enough that they could be adopted without architecture redesigns. One limitation is that the abstract and available description give limited visibility into the exact range of increment sizes tested or the statistical tests used, so it is still unclear how far the gains extend when variability becomes extreme. The work stays within the existing CIL literature rather than introducing new theory or proofs. Readers who care about deploying continual learners in open environments will find the benchmark framing and the concrete fixes useful to try. The paper is coherent on its own terms and supplies enough evidence to merit referee time, even if revisions will be needed for fuller experimental detail.

Referee Report

2 major / 3 minor

Summary. The paper formalizes Free-Flow Class-Incremental Learning (FFCIL) as a realistic CIL setting in which classes arrive in streams with highly variable (unconstrained) numbers per increment. It claims that existing CIL methods suffer clear performance degradation under this regime and proposes a model-agnostic framework whose core is a Class-Wise Mean (CWM) objective that replaces frequency-weighted supervision with uniform class-conditional aggregation, together with three targeted adjustments (distillation restricted to replay buffers, loss-scale normalization, and Dynamic Intervention Weight Alignment (DIWA)) intended to stabilize training under small or unstable increments. Experiments are said to show both baseline degradation and consistent gains from the proposed components.

Significance. If the empirical claims hold, the work usefully shifts CIL evaluation toward more realistic arrival patterns and supplies lightweight, paradigm-agnostic fixes that do not require architectural changes. The model-agnostic framing and the explicit identification of instability sources (unstable class statistics, replay-only distillation) are constructive contributions that could be adopted by multiple existing CIL families.

major comments (2)

[Abstract / Experiments] Abstract and Experiments section: the central claim that baselines exhibit 'clear performance degradation' and that the proposed strategies yield 'consistent gains' is presented without any description of the datasets, the generative process for variable increment sizes, the number of random seeds, or statistical tests. These details are load-bearing for the empirical validation that underpins the entire contribution.
[§3.2] §3.2 (CWM objective): the claim that uniform class-conditional aggregation is strictly more stable than frequency-weighted loss under arbitrary increment sizes is asserted but not accompanied by a derivation or bound showing that the variance of the gradient signal is reduced; the paper therefore relies entirely on the (undetailed) experiments to support this key modeling choice.

minor comments (3)

[§3.1] Notation for the CWM loss should be introduced with an explicit equation rather than prose description so that readers can immediately compare it to standard cross-entropy or replay losses.
[§3.3] The description of DIWA would benefit from a short algorithmic box or pseudocode, especially the rule used to compute the dynamic intervention weight from the observed class statistics.
[Figures] Figure captions should state the exact FFCIL increment schedule used for each plot so that the visual results can be reproduced without consulting the main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the insightful comments. We address the major comments below and will make revisions to improve the clarity and completeness of the empirical validation and theoretical motivation.

read point-by-point responses

Referee: [Abstract / Experiments] Abstract and Experiments section: the central claim that baselines exhibit 'clear performance degradation' and that the proposed strategies yield 'consistent gains' is presented without any description of the datasets, the generative process for variable increment sizes, the number of random seeds, or statistical tests. These details are load-bearing for the empirical validation that underpins the entire contribution.

Authors: We agree that providing these details is essential. Although the Experiments section describes the datasets (CIFAR-100, Tiny-ImageNet, ImageNet-100) and the FFCIL setup with variable class arrivals, we will revise the abstract to briefly note the experimental protocol. Additionally, we will add explicit information on the generative process for increment sizes (random sampling from a distribution with high variance), the use of 5 random seeds, and include statistical tests such as standard deviation reporting and t-tests where appropriate in the revised manuscript. revision: yes
Referee: [§3.2] §3.2 (CWM objective): the claim that uniform class-conditional aggregation is strictly more stable than frequency-weighted loss under arbitrary increment sizes is asserted but not accompanied by a derivation or bound showing that the variance of the gradient signal is reduced; the paper therefore relies entirely on the (undetailed) experiments to support this key modeling choice.

Authors: The motivation for CWM is that frequency-weighted losses can lead to biased gradients when increment sizes vary greatly, as small classes may be underrepresented. We support this through extensive experiments and ablations. We will revise §3.2 to include a brief intuitive explanation and a simple analysis of how uniform aggregation reduces the impact of frequency imbalance on gradient variance, without a full formal bound, as deriving such a bound rigorously would require assumptions on data distributions that may not hold generally in CIL. revision: partial

Circularity Check

0 steps flagged

No significant circularity; framework is self-contained

full rationale

The paper defines FFCIL as a new setting, introduces the CWM objective as a direct replacement for frequency-weighted loss via uniform class-conditional aggregation, and proposes targeted adjustments (distillation constraint, loss normalization, DIWA) as model-agnostic heuristics. These steps are presented as novel contributions supported by empirical degradation of baselines and consistent gains, without any equation reducing by construction to a fitted parameter, self-citation chain, or renamed known result. The argument relies on external validation rather than internal redefinition.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard continual learning components such as replay buffers and distillation being available, plus the assumption that the new objective and adjustments will generalize without new instabilities.

axioms (1)

domain assumption Standard continual learning assumptions including access to replay buffers and distillation mechanisms from prior tasks.
The framework builds directly on representative CIL paradigms that rely on these mechanisms.

pith-pipeline@v0.9.0 · 5519 in / 1284 out tokens · 63156 ms · 2026-05-13T19:49:56.941687+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

class-wise mean (CWM) objective that replaces sample frequency weighted loss with uniformly aggregated class-conditional supervision
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean absolute_floor_iff_bare_distinguishability unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Dynamic Intervention Weight Alignment (DIWA) to prevent over-adjustment caused by unstable statistics from small class increments

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages · 1 internal anchor

[1]

Advances in Neural Information Processing Systems37, 7608–7630 (2024)

Bian, A., Li, W., Yuan, H., Wang, M., Zhao, Z., Lu, A., Ji, P., Feng, T., et al.: Make continual learning stronger via c-flat. Advances in Neural Information Processing Systems37, 7608–7630 (2024)

work page 2024
[2]

In: Proceedings of the European conference on computer vision (ECCV)

Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for in- cremental learning: Understanding forgetting and intransigence. In: Proceedings of the European conference on computer vision (ECCV). pp. 532–547 (2018)

work page 2018
[3]

IEEE Transactions on Knowledge and Data Engineering35(7), 7263–7277 (2022)

Chen, H., Wang, Y., Hu, Q.: Multi-granularity regularized re-balancing for class in- cremental learning. IEEE Transactions on Knowledge and Data Engineering35(7), 7263–7277 (2022)

work page 2022
[4]

Advances in Neural Information Processing Systems35, 6345–6357 (2022)

Chen, S., Gong, C., Li, J., Yang, J., Niu, G., Sugiyama, M.: Learning contrastive embedding in low-dimensional space. Advances in Neural Information Processing Systems35, 6345–6357 (2022)

work page 2022
[5]

In: International Confer- ence on Machine Learning

Chen, S., Niu, G., Gong, C., Li, J., Yang, J., Sugiyama, M.: Large-margin con- trastive learning with distance polarization regularizer. In: International Confer- ence on Machine Learning. pp. 1673–1683. PMLR (2021)

work page 2021
[6]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Cui, Y., Liu, L., Yu, Z., Huang, G., Hong, X.: Few-shot audio-visual class- incremental learning with temporal prompting and regularization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 16118–16126 (2025)

work page 2025
[7]

IEEE transactions on pattern analysis and machine intelligence 44(7), 3366–3385 (2021)

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

work page 2021
[8]

In: 2009 IEEE conference on computer vision and pattern recognition

Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large- scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248–255. Ieee (2009)

work page 2009
[9]

IEEE Transactions on Pattern Analysis and Machine Intelligence46(4), 2054–2070 (2023)

Dong, J., Li, H., Cong, Y., Sun, G., Zhang, Y., Van Gool, L.: No one left behind: Real-world federated class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence46(4), 2054–2070 (2023)

work page 2054
[10]

In: European Conference on Computer Vision

Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled out- puts distillation for small-tasks incremental learning. In: European Conference on Computer Vision. pp. 86–102. Springer (2020)

work page 2020
[11]

In: ICASSP 2025-2025 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP)

Fu, F., Lu, Z.: Enhancing data-free class-incremental learning via image-centric dual distillation. In: ICASSP 2025-2025 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). pp. 1–5. IEEE (2025)

work page 2025
[12]

ACM Computing Surveys (CSUR)50(2), 1–36 (2017)

Gomes, H.M., Barddal, J.P., Enembreck, F., Bifet, A.: A survey on ensemble learn- ing for data stream classification. ACM Computing Surveys (CSUR)50(2), 1–36 (2017)

work page 2017
[13]

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

He, C., Wang, R., Chen, X.: A tale of two cils: The connections between class incremental learning and class imbalanced learning, and beyond. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3559–3569 (2021)

work page 2021
[14]

In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition

He, J.: Gradient reweighting: Towards imbalanced class-incremental learning. In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition. pp. 16668–16677 (2024)

work page 2024
[15]

In: Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence

Hong, C., Jin, Y., Kang, Z., Chen, Y., Li, M., Lu, Y., Wang, H.: Dynamically anchored prompting for task-imbalanced continual learning. In: Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence. pp. 4127– 4135 (2024) Towards Realistic Class-Incremental Learning with Free-Flow Increments 17

work page 2024
[16]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Huang, L., Zeng, Y., Yang, C., An, Z., Diao, B., Xu, Y.: etag: Class-incremental learning via embedding distillation and task-oriented generation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 12591–12599 (2024)

work page 2024
[17]

Kim, S., Hwang, D., Woo, S., Singh, R.: Does prior data matter? exploring joint traininginthecontextoffew-shotclass-incrementallearning.In:Proceedingsofthe IEEE/CVF International Conference on Computer Vision. pp. 5185–5194 (2025)

work page 2025
[18]

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

Korycki, L., Krawczyk, B.: Class-incremental experience replay for continual learn- ing under concept drift. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3649–3658 (2021)

work page 2021
[19]

Technical report (2009)

Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. Technical report (2009)

work page 2009
[20]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision

Lai, X., Lin, L., Chen, W., Yu, Y.: A tiny change, a giant leap: Long-tailed class- incremental learning via geometric prototype alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1444–1453 (2025)

work page 2025
[21]

IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

Li, Y., Wang, H., Qi, Y., Liu, W., Li, R.: Re-fed+: A better replay strategy for fed- erated incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

work page 2025
[22]

Ad- vances in Neural Information Processing Systems36, 11151–11167 (2024)

Liang, Y.S., Li, W.J.: Loss decoupling for task-agnostic continual learning. Ad- vances in Neural Information Processing Systems36, 11151–11167 (2024)

work page 2024
[23]

In: European Conference on Computer Vision

Liu, X., Hu, Y.S., Cao, X.S., Bagdanov, A.D., Li, K., Cheng, M.M.: Long-tailed class incremental learning. In: European Conference on Computer Vision. pp. 495–

work page
[24]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Luo, Z., Liu, Y., Schiele, B., Sun, Q.: Class-incremental exemplar compression for class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11371–11380 (2023)

work page 2023
[25]

arXiv preprint arXiv:2502.08181 (2025)

Ma’sum, M.A., Pratama, M., Skrjanc, I.: Latest advancements towards catas- trophic forgetting under data scarcity: A comprehensive survey on few-shot class incremental learning. arXiv preprint arXiv:2502.08181 (2025)

work page arXiv 2025
[26]

In: International Conference on Learning Representations (2018)

Nguyen, C.V., Li, Y., Bui, T.D., Turner, R.E.: Variational continual learning. In: International Conference on Learning Representations (2018)

work page 2018
[27]

Machine Learning114(3), 1–20 (2025)

Qi, Z.H., Zhou, D.W., Yao, Y., Ye, H.J., Zhan, D.C.: Adaptive adapter routing for long-tailed class-incremental learning. Machine Learning114(3), 1–20 (2025)

work page 2025
[28]

In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

Raghavan, S., He, J., Zhu, F.: Online class-incremental learning for real-world food image classification. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 8195–8204 (2024)

work page 2024
[29]

In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition

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

work page 2001
[30]

Advances in neural information processing systems32(2019)

Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., Wayne, G.: Experience replay for continual learning. Advances in neural information processing systems32(2019)

work page 2019
[31]

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

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

work page 2020
[32]

Neural Networks184, 107053 (2025)

Wang, C., Jiang, J., Hu, X., Liu, X., Ji, X.: Enhancing consistency and mitigat- ing bias: A data replay approach for incremental learning. Neural Networks184, 107053 (2025)

work page 2025
[33]

In: The eleventh international conference on learning representations (2022) 18 Zhiming Xu et al

Wang, F.Y., Zhou, D.W., Liu, L., Ye, H.J., Bian, Y., Zhan, D.C., Zhao, P.: Beef: Bi-compatible class-incremental learning via energy-based expansion and fusion. In: The eleventh international conference on learning representations (2022) 18 Zhiming Xu et al

work page 2022
[34]

Advances in Neural Information Processing Systems36, 15060–15076 (2023)

Wang, Q.W., Zhou, D.W., Zhang, Y.K., Zhan, D.C., Ye, H.J.: Few-shot class- incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems36, 15060–15076 (2023)

work page 2023
[35]

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

Wu, Y., Chen, Y., Wang, L., Ye, Y., Liu, Z., Guo, Y., Fu, Y.: Large scale incre- mental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 374–382 (2019)

work page 2019
[36]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Xu, S., Meng, G., Nie, X., Ni, B., Fan, B., Xiang, S.: Defying imbalanced forgetting in class incremental learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 16211–16219 (2024)

work page 2024
[37]

Neural Networks p

Xu, Z., Yang, S., Xu, B., Shen, F., Zhao, J.: Dual prototypes for adaptive pre- trained model in class-incremental learning. Neural Networks p. 108389 (2025)

work page 2025
[38]

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

Yan, S., Xie, J., He, X.: 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)

work page 2021
[39]

In: European Conference on Computer Vision

Yang, S., Shen, F., Zhao, J.: Entaugment: Entropy-driven adaptive data augmen- tation framework for image classification. In: European Conference on Computer Vision. pp. 197–214. Springer (2024)

work page 2024
[40]

In: ICASSP 2024-2024 IEEE International Conference onAcoustics,SpeechandSignalProcessing(ICASSP).pp.5915–5919.IEEE(2024)

Yang, Y., Ren, D., Peng, C., Huo, J., Li, W., Gao, Y.: Dynamic replay training for class-incremental learning. In: ICASSP 2024-2024 IEEE International Conference onAcoustics,SpeechandSignalProcessing(ICASSP).pp.5915–5919.IEEE(2024)

work page 2024
[41]

IEEE transactions on pattern analysis and machine intelligence 42(7), 1698–1712 (2019)

Ye, H.J., Zhan, D.C., Li, N., Jiang, Y.: Learning multiple local metrics: Global con- sideration helps. IEEE transactions on pattern analysis and machine intelligence 42(7), 1698–1712 (2019)

work page 2019
[42]

A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark

Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djo- longa, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867 (2019)

work page internal anchor Pith review arXiv 1910
[43]

IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

Zhang, J., Liu, L., Silven, O., Pietikäinen, M., Hu, D.: Few-shot class-incremental learning for classification and object detection: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

work page 2025
[44]

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

Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T.: Maintaining discrimination and fairness in class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13208–13217 (2020)

work page 2020
[45]

In: Proceedings of the Computer Vision and Pattern Recognition Conference

Zheng, B., Zhou, D.W., Ye, H.J., Zhan, D.C.: Task-agnostic guided feature expan- sion for class-incremental learning. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 10099–10109 (2025)

work page 2025
[46]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Zhou, D.W., Sun, H.L., Ye, H.J., Zhan, D.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)

work page 2024
[47]

Zhou, D.W., Wang, F.Y., Ye, H.J., Zhan, D.C.: Pycil: a python toolbox for class- incremental learning (2023)

work page 2023
[48]

IEEE Transactions on Pattern Analysis and Machine Intelli- gence (2024)

Zhou,D.W.,Wang,Q.W.,Qi,Z.H.,Ye,H.J.,Zhan,D.C.,Liu,Z.:Class-incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelli- gence (2024)

work page 2024
[49]

A model or 603 exemplars: Towards memory-efficient class-incremental learning.arXiv preprint arXiv:2205.13218, 2022

Zhou, D.W., Wang, Q.W., Ye, H.J., Zhan, D.C.: A model or 603 ex- emplars: Towards memory-efficient class-incremental learning. arXiv preprint arXiv:2205.13218 (2022)

work page arXiv 2022