Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

On the Discrimination and Consistency for Exemplar-Free Class Incremental Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read DCNet claims exemplary-free class incremental learning can match replay-based methods by keeping class embeddings mutually orthogonal and equally tight across tasks, and reports state-of-the-art accuracy on CIFAR-100, Tiny-ImageNet, and…

desk verdict Strong empirical EF-CIL method, but the theory section is an upper bound that doesn't support the causal claim, and the fair comparison to HAT+CSI is buried in the ablation. read the letter →

arxiv 2501.15454 v1 pith:5MTCKC24 submitted 2025-01-26 cs.CV

Tianqi Wang , Jingcai Guo , Depeng Li , Zhi Chen This is my paper · ORCID
classification cs.CV
keywords exemplar-freeclassincrementallearningtask-idpredictionout-of-distributiondetectionhypersphericalembeddingorthogonalprototypesintra-classaggregationcatastrophicforgettingTIL+OODframework
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 addresses exemplar-free class incremental learning, where no old samples may be stored or replayed. It adopts the TIL+OOD decomposition—train per-task experts and decide task identity by out-of-distribution detection—and argues that in the exemplar-free setting the bottleneck is task-id prediction, because tasks cannot interact through replay. It claims that preserving a discriminative and consistent feature space, specifically large inter-class separation and aligned intra-class aggregation, is what enables interaction, and proposes DCNet to realize it. If correct, EF-CIL can match replay-based methods without storing exemplars, which matters for privacy and memory-constrained deployment.

What carries the argument

The load-bearing mechanism is an incremental generator of mutually orthogonal unit basis vectors on the unit hypersphere (IOE), trained with a von Mises-Fisher-style softmax loss that pulls each class's normalized embeddings toward its assigned basis vector, combined with Dynamic Aggregation Compensation (DAC). DAC measures intra-class aggregation as the average cosine similarity between sample embeddings and their class basis vector, then adaptively scales the temperature of a supervised contrastive loss so that later, less plastic tasks are compressed as tightly as earlier tasks. Hard attention masks from HAT protect old-task weights, while the orthogonal basis vectors and the aggregation statistic carry information between tasks without replaying any old samples.

What would settle it

Train the IOE objective on synthetic Gaussian mixtures where the OOD mean vectors are fixed and the IND prototype separation is varied; if the empirical OOD gap $D$ and downstream task-id accuracy do not increase with separation, the design principle is falsified. Alternatively, replace IOE's orthogonal basis with random fixed unit vectors matched in number and norm and measure $A_{inc}$: if accuracy does not drop, orthogonality is not the operative cause of the reported gains.

Watch

Extended reading notes

Core claim

The central claim is that the failure of exemplar-free class incremental learning under the TIL+OOD framework comes from task isolation in feature space, and the remedy is to make within-task class embeddings mutually orthogonal on the unit hypersphere while equalizing intra-class aggregation across tasks. The paper argues through Theorem 1 that larger inter-class prototype distance and a smaller shared covariance enlarge the expected gap between in-distribution and out-of-distribution energy scores, thereby improving task-id prediction. On benchmarks, DCNet reports 75.84/71.52 Ainc on CIFAR-100 Split-10/20, 57.00/50.05 on Tiny-ImageNet, and 76.82/69.12 on ImageNet-Subset, beating all exemplar-free baselines and nearly matching or exceeding methods that replay 2000 stored samples.

Load-bearing premise

The argument assumes that the upper bound in Theorem 1 can be read as a causal statement: making in-distribution prototypes farther apart while tightening their shared covariance will itself enlarge the true OOD detection gap, even though the bound also contains an unmeasured OOD-to-nearest-IND term and is not a lower bound.

Editorial extensions

If this is right

  • If correct, exemplar-free continual learning no longer requires storing old samples; DCNet's reported gains approach or beat exemplar-based TIL+OOD baselines on CIFAR-100 and Tiny-ImageNet.
  • Increasing inter-class separation and aligning intra-class aggregation across tasks should improve task-id prediction, which is the main bottleneck of the TIL+OOD framework in the exemplar-free setting.
  • The method's improvements are largest on the harder ImageNet-Subset task, suggesting that geometric compensation matters more as model plasticity drops over longer incremental sequences.
  • The IOE and DAC components are designed to be complementary: orthogonality defines where new class embeddings should live, while DAC corrects for the diffuse embeddings that arise from reduced plasticity.
  • DCNet provides an end-to-end exemplar-free alternative to replay-based interaction methods such as MORE, ROW, and TPL within the same TIL+OOD family.

Reading between the lines

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

  • Theorem 1 is only an upper bound, so the paper's causal reading—more inter-class separation forces a better OOD gap—is an interpretive step; deriving a matching lower bound in terms of inter-class separation would make the design principle fully grounded.
  • The orthogonal-basis construction could transfer to other continual settings such as blurry task boundaries or online learning, where the fixed basis vectors could serve as a stable communication channel between tasks.
  • DAC's scalar aggregation heuristic could be replaced by direct concentration-parameter estimation for the von Mises-Fisher distribution or by per-class uncertainty weighting; a testable extension is whether calibrated per-class temperatures outperform the single adaptive temperature.
  • Because the basis vectors are generated incrementally in task order, the method may be sensitive to class order; testing multiple random class orders would reveal whether the benefit comes from orthogonality itself or from the specific ordering of the basis vectors.
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. The paper addresses exemplar-free class incremental learning (EF-CIL) within the TIL+OOD framework, where task-id prediction is treated as an out-of-distribution (OOD) detection problem. The authors provide a theoretical analysis intended to show that preserving inter-class separation and intra-class aggregation improves OOD detection in an incremental sequence. Based on this analysis, they propose DCNet, which combines Incremental Orthogonal Embedding (IOE) to place class embeddings on near-orthogonal basis vectors on the hypersphere and Dynamic Aggregation Compensation (DAC) to adaptively adjust the degree of intra-class clustering. Experiments on CIFAR-100, Tiny-ImageNet, and ImageNet-Subset report substantially higher accuracy than existing EF-CIL methods, and competitive performance against exemplar-based methods without replay.

Significance. If the theoretical mechanism were established, the paper would make a meaningful conceptual contribution by connecting feature-space geometry to task-id prediction in EF-CIL. The empirical results are strong and consistently favor DCNet across three datasets and two split protocols, with code and detailed hyperparameters provided, which supports reproducibility. However, the central theoretical claim as stated does not follow from Theorem 1: the theorem yields an upper bound, not a lower bound, and the proof has a technical gap involving the t-dependence of the nearest-prototype index. The method may still be a valuable empirical contribution, but the paper's claim to provide a theoretical foundation for its design principle is not supported. Thus the significance is primarily empirical at this stage.

major comments (3)
  1. [Section 3.2, Eq. (6) and following text] The sentence immediately after Eq. (6), 'It can be found that the performance of TIL+OOD approaches can be improved by increasing the inter-class difference of IND prototypes,' reads Theorem 1 as a causal monotonic statement. Theorem 1 is an upper bound on D, not a lower bound. Enlarging the controllable inter-class separation term (1/2)Σ_i d_M(µ_in,i0, µ_in,i) need not increase D, because the first term (k/(2T))Σ_t d_M(µ_out,t, µ_in,i0) is unmeasured and may increase as the IND prototypes move apart. Without a lower bound or an additional assumption controlling the OOD-to-nearest-IND term, the design principle invoked in Sections 3.2, 3.3, and 4.2 is not derived from this theorem. This is load-bearing because the paper claims the method is theory-guided.
  2. [Appendix A, Proof of Theorem 1] The proof of Theorem 1 defines i0 = arg min_i d_M(µ_out,t, µ_in,i), which depends on t, but the displayed result in Eq. (6) uses a single i0 in the second term: (1/2)Σ_i d_M(µ_in,i0, µ_in,i). Summing Eq. (5) over t and i gives (1/(2T))Σ_t Σ_i d_M(µ_in,i0(t), µ_in,i), not the stated expression. To obtain the theorem as written, one would need i0 to be independent of t, e.g., the prototype closest to all OOD means or to the overall OOD mixture, which is neither stated nor proved. Therefore Theorem 1 is not established as stated. A corrected statement or an explicit extra assumption is required before the bound can be used as the basis for the subsequent design arguments.
  3. [Section 4.2 and 4.3] The IOE and DAC components are presented as realizing the theoretical conclusions of Section 3.2. Given that the theoretical conclusions are not supported (see major comments above), the causal attribution 'BCNet preserves discrimative and consistent feature space' via the stated theorem is not justified. The empirical ablation in Table 3 does show that adding IOE and DAC improves over HAT+CSI, which is evidence for the method's practical value, but it does not validate the theoretical mechanism. I recommend either revising the theoretical section to provide a valid supporting argument (e.g., a lower bound on D under a controlled OOD distribution, or a formal statement about relative improvement) or explicitly repositioning the method as an empirically motivated design with the theorem as a heuristic inspiration rather than a proof.
minor comments (4)
  1. [Section 1, Contributions] The first contribution bullet contains a typo: 'TIL+ODD' should be 'TIL+OOD'.
  2. [Appendix A, Lemma 2 heading] The text before Lemma 2 says 'we prove the following lemma, which elucidates the factors contributing to the lower bound,' but the lemma is used to establish an upper bound on D. This is a mislabeling that could confuse readers.
  3. [Section 5.1, Training details] The hyperparameter 'cosine value of 0.1 is used to approximate orthogonality between vectors' is vague; it is unclear whether this is a threshold on absolute cosine similarity or a target value in the optimization in Eq. (9), and how the cosine constraint interacts with the penalty terms in Eq. (9) should be clarified.
  4. [Section 5.3, Figure 4] Figure 4 shows 'average inter-class Mahalanobis distance across the incremental sequence,' but the text does not specify how this distance is computed for the HAT+CSI baseline versus DCNet, nor whether the Mahalanobis distance is computed using the shared covariance estimate from the feature space. Adding this detail would improve interpretability.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: DCNet's empirical gains are benchmarked externally, and its theoretical bound functions as a design heuristic rather than a self-fulfilling prediction.

full rationale

The derivation chain is not circular. Theorem 1 (Eq. 6) upper-bounds the IND/OOD score gap in terms of Mahalanobis distances; IOE maximizes the inter-class separation term appearing in that bound, but this is a hypothesis-driven design choice, not a definition of success in terms of the model's own fitted parameters. The bound is not used to fit a constant and then report that constant as a prediction. The paper's central empirical claims are checked against external EF-CIL and exemplar-based baselines on CIFAR-100, Tiny-ImageNet, and ImageNet-Subset, so the method is independently falsifiable. The self-citations to Li et al. (2024a, 2024b) appear only in generic statements about HSIC regularization and adaptive hyperparameters and are not load-bearing for DCNet's mechanism. The Appendix itself acknowledges that 'within the context of EF-CIL, mu_out,t is not estimable' and then introduces inter-class separation as a proxy, which is transparent rather than circular. The skeptic's concern that Eq. (6) is an upper bound with an uncontrolled OOD-to-nearest-IND term, and that the proof appears to treat the nearest-prototype index i0 as constant across tasks when it may be t-dependent, is a correctness and rigor criticism of the theoretical justification, not a circularity: nothing in the paper reduces the experimental outcome to an input by construction. Score 2 reflects only the presence of minor, non-load-bearing self-citations.

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

The method's free parameters are conventional loss temperatures, weights, and embedding dimensions, plus one threshold. The central derivation depends on the Gaussian-mixture model and, critically, on an unstated inference from an upper bound to a causal design rule. No new physical entities are introduced.

free parameters (6)
  • tau_IOE = 0.05
    Temperature in the hyperspherical vMF loss Eq. (11); hand-set across experiments.
  • tau_0 = 0.2
    Initial temperature for DAC contrastive loss Eq. (12); hand-set.
  • lambda = 1.0
    Weight of LDAC in Eq. (14); selected after ablation over {0.5, 1.0, 2.0} on the same benchmark, not a separate validation set.
  • lambda_HAT = 1.5 first task, 1.0 later
    Weight of LHAT; adopted from prior configuration in Kim et al. 2022a.
  • basis vector dimension = 256
    Dimension of the hyperspherical embedding and basis vectors; hand-set.
  • orthogonality cosine threshold = 0.1
    Threshold used to approximate orthogonality in the basis generation step; hand-set.
assumptions (5)
  • domain assumption Gaussian mixture model for IND and OOD features
    Section 3.2 models in-distribution data as k Gaussian components and out-of-distribution data as T Gaussian means with shared covariance; real image embeddings need not follow this, so the bound is conditional on this model.
  • standard math Standard information inequalities (total variation, KL, Pinsker)
    Used in Appendix A to prove Lemma 1 and Theorem 1; these are accepted background results.
  • ad hoc to paper Upper bound can be used as a design principle
    The paper implicitly assumes that increasing the inter-class separation term in the upper bound improves the actual OOD gap D; this is not proven and is the main logical gap.
  • domain assumption Task selection via argmax over separately trained classifiers works when outputs are comparable
    Section 3.3 relies on Eq. (8) and on the claim that aligned aggregation makes output magnitudes comparable, but no formal consistency guarantee is given.
  • standard math Orthogonal basis vectors can be generated for all classes in 256 dimensions
    With 256 dimensions and at most 100 classes, exact orthogonality is possible in principle; the paper does not give the generation algorithm but this is not a risky assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Discrimination and Consistency for Exemplar-Free Class Incremental Learning." pith.science (2026). https://pith.science/paper/5MTCKC24

@misc{pith2026250115454,
  author       = {Pith},
  title        = {Pith review of: On the Discrimination and Consistency for Exemplar-Free Class Incremental Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5MTCKC24}},
  note         = {Machine review of arXiv:2501.15454}
}
read the original abstract

Exemplar-free class incremental learning (EF-CIL) is a nontrivial task that requires continuously enriching model capability with new classes while maintaining previously learned knowledge without storing and replaying any old class exemplars. An emerging theory-guided framework for CIL trains task-specific models for a shared network, shifting the pressure of forgetting to task-id prediction. In EF-CIL, task-id prediction is more challenging due to the lack of inter-task interaction (e.g., replays of exemplars). To address this issue, we conduct a theoretical analysis of the importance and feasibility of preserving a discriminative and consistent feature space, upon which we propose a novel method termed DCNet. Concretely, it progressively maps class representations into a hyperspherical space, in which different classes are orthogonally distributed to achieve ample inter-class separation. Meanwhile, it also introduces compensatory training to adaptively adjust supervision intensity, thereby aligning the degree of intra-class aggregation. Extensive experiments and theoretical analysis verified the superiority of the proposed DCNet.

Figures

Figures reproduced from arXiv: 2501.15454 by the authors.

Figure 1
Figure 1. Overview of DCNet. Upon the arrival of task t, DCNet optimizes the learnable part of the backbone and creates a new OOD [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. t-SNE visualization of the embedding space, where a color [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Changes in the de￾gree of aggregation ω (t) . 8 1 0 1 2 M a h a l a n o b i s D i s t a n c e [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Progressive Homeostatic and Plastic Prompt Tuning for Audio-Visual Multi-Task Incremental Learning

    cs.AI 2025-07 reject novelty 4.0 of 10

    A three-stage prompt-tuning method for audio-visual multi-task incremental learning is proposed, reporting state-of-the-art results on AVE, AVVP, AVS, and AVQA, with caveats about its evaluation metric and ablations.

Reference graph

Works this paper leans on

41 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    Conditional channel gated networks for task-aware continual learning

    [Abati et al., 2020] Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, and Babak Ehteshami Bejnordi. Conditional channel gated networks for task-aware continual learning. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3931–3940,

    work page 2020

  2. [3]

    Expanding hyperspherical space for few-shot class- incremental learning

    [Deng and Xiang, 2024] Yao Deng and Xiang Xiang. Expanding hyperspherical space for few-shot class- incremental learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages 1967–1976,

    work page 2024

  3. [6]

    Deep residual learning for image recognition

    [He et al., 2016] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778,

    work page 2016

  4. [13]

    Harnessing Neural Unit Dynamics for Effective and Scalable Class-Incremental Learning

    [Li et al., 2024a] Depeng Li, Tianqi Wang, Junwei Chen, Wei Dai, and Zhigang Zeng. Harnessing neural unit dy- namics for effective and scalable class-incremental learn- ing. arXiv preprint arXiv:2406.02428,

    work page Pith review arXiv

  5. [14]

    Class incremental learning via likelihood ratio based task prediction

    [Lin et al., 2024] Haowei Lin, Yijia Shao, Weinan Qian, Ningxin Pan, Yiduo Guo, and Bing Liu. Class incremental learning via likelihood ratio based task prediction. InInter- national Conference on Learning Representations,

    work page 2024

  6. [15]

    Task-adaptive saliency guidance for exemplar-free class incremental learning

    [Liu et al., 2024] Xialei Liu, Jiang-Tian Zhai, Andrew D Bagdanov, Ke Li, and Ming-Ming Cheng. Task-adaptive saliency guidance for exemplar-free class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 23954– 23963,

    work page 2024

  7. [16]

    Learning with mixture of prototypes for out-of-distribution detection

    [Lu et al., 2024] Haodong Lu, Dong Gong, Shuo Wang, Ja- son Xue, Lina Yao, and Kristen Moore. Learning with mixture of prototypes for out-of-distribution detection. In International Conference on Learning Representations ,

    work page 2024

  8. [17]

    Bagdanov

    [Magistri et al., 2024] Simone Magistri, Tomaso Trinci, Al- bin Soutif, Joost van de Weijer, and Andrew D. Bagdanov. Elastic feature consolidation for cold start exemplar-free incremental learning. In International Conference on Learning Representations,

    work page 2024

Show all 41 references
  1. [18]

    Directional statistics

    [Mardia and Jupp, 2009] Kanti V Mardia and Peter E Jupp. Directional statistics. John Wiley & Sons,

    2009

  2. [19]

    Catastrophic interference in connectionist networks: The sequential learning problem

    [McCloskey and Cohen, 1989] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of Learning and Motivation , volume 24, pages 109–165. Elsevier,

    1989

  3. [21]

    How to exploit hyperspherical embeddings for out-of-distribution detection? In International Confer- ence on Learning Representations,

    [Ming et al., 2023] Yifei Ming, Yiyou Sun, Ousmane Dia, and Yixuan Li. How to exploit hyperspherical embeddings for out-of-distribution detection? In International Confer- ence on Learning Representations,

    2023

  4. [22]

    Provable guarantees for understanding out-of-distribution detection

    [Morteza and Li, 2022] Peyman Morteza and Yixuan Li. Provable guarantees for understanding out-of-distribution detection. In Proceedings of the AAAI Conference on Ar- tificial Intelligence, pages 7831–7840,

    2022

  5. [23]

    Fetril: Feature translation for exemplar-free class-incremental learning

    [Petit et al., 2023] Gr´egoire Petit, Adrian Popescu, Hugo Schindler, David Picard, and Bertrand Delezoide. Fetril: Feature translation for exemplar-free class-incremental learning. In Proceedings of the IEEE/CVF Winter Con- ference on Applications of Computer Vision, pages 39...

    2023

  6. [24]

    itaml: An incremental task-agnostic meta- learning approach

    [Rajasegaran et al., 2020] Jathushan Rajasegaran, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Mubarak Shah. itaml: An incremental task-agnostic meta- learning approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13588–13597,

    2020

  7. [25]

    icarl: Incremental classifier and representation learning

    [Rebuffi et al., 2017] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Pro- ceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001–2010,

    2017

  8. [26]

    Imagenet large scale visual recogni- tion challenge

    [Russakovsky et al., 2015] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recogni- tion challenge. International journal of computer vision , 1...

    2015

  9. [27]

    Divide and not for- get: Ensemble of selectively trained experts in continual learning

    [Rype´s´c et al., 2023] Grzegorz Rype ´s´c, Sebastian Cygert, Valeriya Khan, Tomasz Trzcinski, Bartosz Michał Zieli´nski, and Bartłomiej Twardowski. Divide and not for- get: Ensemble of selectively trained experts in continual learning. In International Conference on Learning ...

    2023

  10. [28]

    Adaptive hyperparame- ter optimization for continual learning scenarios

    [Semola et al., 2024] Rudy Semola, Julio Hurtado, Vincenzo Lomonaco, and Davide Bacciu. Adaptive hyperparame- ter optimization for continual learning scenarios. arXiv preprint arXiv:2403.07015,

    2024 arXiv

  11. [29]

    Overcoming catastrophic forgetting with hard attention to the task

    [Serr`a et al., 2018] Joan Serr `a, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In Interna- tional Conference on Machine Learning , pages 4548–

    2018

  12. [30]

    Continual learning with hypernetworks

    [V on Oswaldet al., 2019] Johannes V on Oswald, Christian Henning, Benjamin F Grewe, and Jo ˜ao Sacramento. Continual learning with hypernetworks. arXiv preprint arXiv:1906.00695,

    2019 arXiv

  13. [32]

    BEEF: Bi-compatible class-incremental learning via energy-based expansion and fusion

    [Wang et al., 2023] Fu-Yun Wang, Da-Wei Zhou, Liu Liu, Han-Jia Ye, Yatao Bian, De-Chuan Zhan, and Peilin Zhao. BEEF: Bi-compatible class-incremental learning via energy-based expansion and fusion. In International Conference on Learning Representations,

    2023

  14. [33]

    DER: Dynamically expandable representation for class incremental learning

    [Yan et al., 2021] Shipeng Yan, Jiangwei Xie, and Xum- ing He. DER: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3014–3023,

    2021

  15. [35]

    A model or 603 exemplars: Towards memory-efficient class-incremental learning

    [Zhou et al., 2023] Da-Wei Zhou, Qi-Wei Wang, Han-Jia Ye, and De-Chuan Zhan. A model or 603 exemplars: Towards memory-efficient class-incremental learning. In Interna- tional Conference on Learning Representations,

    2023

  16. [36]

    Prototype augmentation and self-supervision for incremental learning

    [Zhu et al., 2021] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and Cheng-Lin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 5871–5880,

    2021

  17. [37]

    Self-organizing pathway expansion for non-exemplar class-incremental learning

    [Zhu et al., 2023] Kai Zhu, Kecheng Zheng, Ruili Feng, Deli Zhao, Yang Cao, and Zheng-Jun Zha. Self-organizing pathway expansion for non-exemplar class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 19204–19213,

    2023

  18. [38]

    Acil: Analytic class-incremental learning with ab- solute memorization and privacy protection

    [Zhuang et al., 2022] Huiping Zhuang, Zhenyu Weng, Hongxin Wei, Renchunzi Xie, Kar-Ann Toh, and Zhiping Lin. Acil: Analytic class-incremental learning with ab- solute memorization and privacy protection. Advances in Neural Information Processing Systems, 35:11602–11614,

    2022

  19. [39]

    1: # Training Time 2: for t = 1, 2,

    Algorithm 1 DCNet Training and Test Algorithm Input: Datasets {D(t)}T t=1; Encoder f; Initial temperature τ (0); Epoch for IOE epochIOE ; Epoch for DAC epochDAC, etc. 1: # Training Time 2: for t = 1, 2, . . . , Tdo 3: Generate orthogonal unit basis vectors µ(t) by Eq. (9) 4: f...

    2018

  20. [40]

    All methods employed a ResNet-18 network trained from scratch as the backbone, without leveraging any pre-trained models

    C Additional implementation details C.1 Hyper-parameter For all baselines, we utilize either the results reported in the latest state-of-the-art baseline or reproduce the experiments using the hyperparameters provided in the original source code to ensure a fair comparison. Al...

    2024

  21. [1989]

    Diffclass: Diffusion-based class incremental learning

    [Meng et al., 2025] Zichong Meng, Jie Zhang, Changdi Yang, Zheng Zhan, Pu Zhao, and Yanzhi Wang. Diffclass: Diffusion-based class incremental learning. In European Conference on Computer Vision, pages 142–159. Springer,

    2025

  22. [2000]

    For the CIFAR-100 and Tiny-ImageNet datasets, we train the backbone for 700 epochs using LARS [You et al., 2017] with an initial learning rate of 0.1 and a batch size of

    For our approach, we follow the setup of the previous work [Kim et al., 2022a; Kim et al., 2022c]. For the CIFAR-100 and Tiny-ImageNet datasets, we train the backbone for 700 epochs using LARS [You et al., 2017] with an initial learning rate of 0.1 and a batch size of

    2017

  23. [2009]

    Tiny imagenet visual recognition challenge

    [Le and Yang, 2015] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3,

    2015

  24. [2015]

    Learning without forgetting

    [Li and Hoiem, 2017] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pat- tern Analysis and Machine Intelligence , 40(12):2935– 2947,

    2017

  25. [2016]

    Posterior meta-replay for continual learning

    [Henning et al., 2021] Christian Henning, Maria Cervera, Francesco D’Angelo, Johannes V on Oswald, Regina Tra- ber, Benjamin Ehret, Seijin Kobayashi, Benjamin F Grewe, and Joao Sacramento. Posterior meta-replay for continual learning. Advances in Neural Information Pro- cessin...

    2021

  26. [2017]

    Learning multiple layers of features from tiny im- ages

    [Krizhevsky et al., 2009] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny im- ages. Handbook of Systemic Autoimmune Diseases,

    2009

  27. [2019]

    FOSTER: Feature boosting and compression for class-incremental learning

    [Wang et al., 2022] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. FOSTER: Feature boosting and compression for class-incremental learning. In European Conference on Computer Vision, pages 398–414. Springer,

    2022

  28. [2020]

    Dark experience for general continual learning: a strong, sim- ple baseline

    [Buzzega et al., 2020] Pietro Buzzega, Matteo Boschini, An- gelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, sim- ple baseline. Advances in Neural Information Processing Systems, 33:15920–15930,

    2020

  29. [2021]

    Large batch training of convolutional networks

    [You et al., 2017] Yang You, Igor Gitman, and Boris Gins- burg. Large batch training of convolutional networks. arXiv preprint arXiv:1708.03888,

    2017 arXiv

  30. [2022]

    Learnability and algorithm for con- tinual learning

    [Kim et al., 2023] Gyuhak Kim, Changnan Xiao, Tatsuya Konishi, and Bing Liu. Learnability and algorithm for con- tinual learning. In International Conference on Machine Learning, pages 16877–16896. PMLR,

    2023

  31. [2023]

    Overcom- ing catastrophic forgetting in neural networks

    [Kirkpatrick et al., 2017] James Kirkpatrick, Razvan Pas- canu, Neil Rabinowitz, Joel Veness, Guillaume Des- jardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcom- ing catastrophic forgetting in neural networks. Proceed- in...

    2017

  32. [2024]

    Exemplar-free continual representation learning via learnable drift compensation

    [Gomez-Villa et al., 2025] Alex Gomez-Villa, Dipam Goswami, Kai Wang, Andrew D Bagdanov, Bartlomiej Twardowski, and Joost van de Weijer. Exemplar-free continual representation learning via learnable drift compensation. In European Conference on Computer Vision, pages 473–490. ...

    2025

  33. [2025]

    Resurrecting old classes with new data for exemplar-free continual learn- ing

    [Goswami et al., 2024] Dipam Goswami, Albin Soutif- Cormerais, Yuyang Liu, Sandesh Kamath, Bart Twar- dowski, Joost van de Weijer, et al. Resurrecting old classes with new data for exemplar-free continual learn- ing. In Proceedings of the IEEE/CVF Conference on Computer Vision...

    2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.