Pith. sign in

REVIEW 4 major objections 4 minor 65 references

Divide-and-Conquer for Enhancing Unlabeled Learning, Stability, and Plasticity in Semi-supervised Continual Learning

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read USP claims that semi-supervised continual learning improves most when its three objectives—unlabeled learning, memory stability, and learning plasticity—are solved by one coupled divide-and-conquer framework, reporting gains up to 5.94% in

desk verdict A plausible divide-and-conquer recipe for semi-supervised continual learning, but the visible snapshot omits the core method and main results, so the headline 5.94% gain cannot be checked from this version. read the letter →

arxiv 2508.05316 v1 pith:SGAGDPAX submitted 2025-08-07 cs.LG cs.CV

Yue Duan , Taicai Chen , Lei Qi , Yinghuan Shi This is my paper
classification cs.LGcs.CV
keywords semi-supervisedcontinuallearningunlabeledmemorystabilityplasticitypseudo-labelingequiangulartightframeknowledgedistillationclass-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

USP, the method this paper proposes, is built on the claim that semi-supervised continual learning fails when its three objectives are handled separately. The framework combines three coupled strategies: reserving feature-space directions for future classes, pseudo-labeling both confident and uncertain unlabeled examples, and anchoring old-class means during distillation. The paper reports that the combination outperforms earlier semi-supervised continual learning methods, with final-task accuracy gains up to 5.94 percentage points on standard benchmarks. A sympathetic reader would take the central claim to be that joint design of unlabeled learning, memory stability, and learning plasticity is the deciding factor, not any single technique.

What carries the argument

Equiangular tight frame (ETF): a geometry of class prototypes in which all pairwise angles are equal and as large as possible; here it is used to pre-allocate feature directions for future classes. Around it sit two coupled algorithms: Divide-and-Conquer Pseudo-labeling (DCP), which sends high-confidence unlabeled samples down a hard-label path and low-confidence samples down a soft-label path, and Class-mean-anchored Unlabeled Distillation (CUD), which reuses DCP's outputs to anchor unlabeled data to class means during distillation. The ETF supplies plasticity, DCP supplies unlabeled supervision, and CUD converts that supervision into stability.

What would settle it

Construct a class-incremental stream that has more classes than the number of feature slots Feature Space Reservation pre-allocates, while keeping everything else identical. If accuracy on the final, overflow classes collapses relative to a method with no reservation, the claimed plasticity mechanism is the limiting factor; if it does not, the reservation is not what carries the gain.

Watch

Extended reading notes

Core claim

The central discovery is that the three failure modes can be turned into one another's solutions. Feature Space Reservation reorganizes the features of already-seen classes into an equiangular tight frame, where class directions are evenly separated; the empty directions in that frame are reserved slots for classes that will arrive later, so new classes can be absorbed without overwriting old geometry. Divide-and-Conquer Pseudo-labeling splits unlabeled data by confidence and assigns pseudo-labels to both groups, including low-confidence samples that standard methods discard. Class-mean-anchored Unlabeled Distillation then takes those pseudo-labels and uses them to keep unlabeled examples al

Load-bearing premise

Everything rests on knowing an upper bound on the total number of classes ahead of time, because the reserved feature slots are fixed when training starts; if the stream later exceeds that bound, the reserved slots may no longer correspond to real future classes, and the plasticity benefit can vanish.

Editorial extensions

If this is right

  • Low-confidence unlabeled data can be turned into usable training signal rather than filtered out, increasing the effective supervision in each task.
  • Feature-space reservation for future classes and class-mean anchoring for old classes are complementary, so plasticity gains do not have to come from sacrificing old-class memory.
  • The method's gains appear both with a small replay buffer and with a buffer large enough to store every labeled sample, suggesting the benefit is not an artifact of buffer pressure alone.
  • Because CUD consumes DCP's pseudo-labels, the unlabeled-learning and memory-stability modules are interdependent; improvements in pseudo-label quality should directly improve forgetting resistance.

Reading between the lines

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

  • Inference: if the slot-count requirement of Feature Space Reservation could be relaxed to an upper bound or made adaptive, the same geometry could carry the method into open-world or task-free streams, where the paper does not claim to operate.
  • Inference: the tight coupling between DCP and CUD implies that pseudo-label noise can propagate into the distillation anchor; a direct measurement of how low-confidence label error flows into final accuracy would be a natural stress test, not one the paper performs.
  • Inference: the 'up to 5.94%' figure is an upper bound across datasets; the paper does not establish where in the spectrum of labeled-to-unlabeled ratios the gain peaks, so a sweep over annotation budgets would clarify when the framework pays off most.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The manuscript proposes USP, a divide-and-conquer framework for semi-supervised continual learning (SSCL), with three components: Feature Space Reservation (FSR) for learning plasticity, Divide-and-Conquer Pseudo-labeling (DCP) for unlabeled learning, and Class-mean-anchored Unlabeled Distillation (CUD) for memory stability. The abstract claims improved performance over prior SSCL methods, with gains up to 5.94% in last accuracy, evaluated on benchmarks such as CIFAR and ImageNet variants. The text supplied for review consists of the abstract, introduction, references, and two experimental tables at the end: a buffer-size sweep (Table 11) and a 20K-buffer comparison of iCaRL&Fix and DER&Fix with and without USP (Table 12). No method section, loss equations, pseudo-code, benchmark protocol, standard deviations, or comparison with prior SSCL methods is present in the supplied manuscript.

Significance. If the claimed results are reproducible, USP would be a useful contribution to SSCL by jointly addressing stability, plasticity, and unlabeled learning rather than optimizing them separately. The public code link and the large-buffer experiments in Table 12 provide some initial evidence that adding USP can improve two CL baselines. However, because the method is not formally described and the main comparisons and ablations are absent, the current text cannot establish the central claim. The significance is therefore conditional: the contribution is potentially valuable but unverified in the submitted snapshot.

major comments (4)
  1. [After Section 1] The manuscript contains no method section. The text goes from the Introduction directly to references, so FSR, DCP, and CUD are only named, not defined. There are no equations, no algorithm, no pseudo-code, and no hyperparameters such as the DCP confidence thresholds, the number of reserved feature slots in FSR, or the distillation weight in CUD. This is load-bearing: the paper's central claim is that these mechanisms jointly improve SSCL, and without their formal definitions the result cannot be reproduced, attributed, or checked.
  2. [Table 12] Table 12 compares iCaRL&Fix and DER&Fix with and without USP under a 20K buffer. These are CL baselines augmented by FixMatch, not the prior SSCL methods referenced in the abstract and introduction (e.g., ORdisco and others). No table showing USP against actual SSCL baselines is visible. Therefore the abstract's central claim that USP 'outperforms prior SSCL methods' is unsupported by the supplied evidence. The +USP gains could in principle come from the added semi-supervised consistency or distillation alone, since no component-level ablation is shown.
  3. [Tables 11 and 12] The experimental results report single numbers without standard deviations, number of seeds, or significance tests. In Table 12, the ImageNet100-100 average accuracy gain for DER&Fix is only 0.15 percentage points (62.17 vs 62.02), which is within plausible run-to-run variability. The claimed headline gain of 5.94% is not located in the visible tables, and none of the visible gains can be assessed for statistical reliability. Multiple seeds and error bars are necessary before the improvement claims can be accepted.
  4. [Table 11 and FSR description] Table 11 is titled an ablation study but it only varies the buffer size; it is not an ablation of the three proposed components. No results are shown with FSR, DCP, or CUD removed or individually ablated. Relatedly, the abstract's description of FSR as shaping old classes into an equiangular tight frame presupposes that the total number of classes is known and that reserved slots correspond to actual future classes. The supplied text does not state this assumption, discuss its validity when the task sequence is open-ended, or test sensitivity to the number of reserved slots. These are load-bearing gaps for the plasticity claim.
minor comments (4)
  1. [General layout] Tables 11 and 12 appear after the references without section headings or captions explaining where they belong. The manuscript should have a proper experimental section with those tables in the main body.
  2. [Table 11] The table is captioned 'Ablation studies on memory buffer size,' but buffer-size variation is a sensitivity analysis, not an ablation of the method's components. The terminology should be corrected.
  3. [Benchmark naming] Benchmark names such as CIFAR10-30, CIFAR10-150, CIFAR100-125, and ImageNet100-100 are not defined. The number of tasks, classes per task, labeled/unlabeled split, unlabeled data availability, buffer-management strategy, and evaluation protocol must be specified for the tables to be interpretable.
  4. [Abstract] The abstract cites 'gains up to 5.94% in the last accuracy,' but this number does not appear in the visible tables. The largest visible last-accuracy gains are different values. Please make the source of the claimed number explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: empirical gains are measured outcomes, not derivationally forced by inputs.

full rationale

The available manuscript is an empirical semi-supervised continual learning paper. Its central claim—USP outperforms prior SSCL methods—is supported in the visible text only by Tables 11 and 12, which report measured accuracy values for trained models with and without USP on top of iCaRL&Fix and DER&Fix. These numbers are experimental outcomes, not quantities derived by construction from inputs; no loss equation, pseudo-labeling rule, or distillation objective is shown that would permit a claim such as 'X is defined in terms of Y' to be checked. The abstract's three components (FSR, DCP, CUD) are described qualitatively, and no equation ties the reserved feature frame, pseudo-label assignment, or class-mean anchoring to the reported last accuracy in a way that would make the result definitional. The paper's self-citations (e.g., refs. [9], [15]–[17]) appear in related-work context and are not used to justify the central empirical claim. There is no uniqueness theorem imported from the authors' prior work, no ansatz smuggled in via citation, and no renamed known result exhibited. Even though the manuscript appears truncated (no visible method section or main comparison tables), absence of methodological detail is a completeness/reproducibility concern, not circularity. No specific reduction from output to input can be quoted, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 3 invented entities

USP introduces no physical or natural entities, but three algorithmic constructs that each carry implicit assumptions and design choices. The free parameters (thresholds, slot counts, weights) are not specified in the available text, and the axioms are standard for SSCL or borrowed from neural-collapse literature.

free parameters (3)
  • Confidence thresholds in DCP
    Divide-and-Conquer Pseudo-labeling splits data into high and low confidence sets, requiring thresholds that are presumably chosen by hand or tuned on validation data. Specific values are not given in the abstract or appendix.
  • Number of reserved feature slots in FSR
    Feature Space Reservation needs to know how many future-class slots to pre-allocate, likely set to the total number of classes in the benchmark. This is a design parameter.
  • Distillation weight in CUD
    Class-mean-anchored Unlabeled Distillation likely includes a weighting term balancing distillation loss and task loss, not specified in the available text.
assumptions (4)
  • domain assumption Class-incremental setup with non-overlapping classes across tasks
    The SSCL setting in Fig. 1 assumes each new task introduces distinct classes with no overlap with earlier tasks. This is standard for class-incremental learning and is a boundary condition for the method.
  • domain assumption Labeled and unlabeled data within a task come from the same label distribution
    Pseudo-labeling and distillation implicitly assume that unlabeled examples are drawn from the same classes as labeled examples in that task, otherwise pseudo-labels would be meaningless.
  • domain assumption Equiangular tight frame (ETF) geometry improves feature discrimination and forward transfer
    FSR relies on the neural-collapse characteristic that an ETF classifier/feature layout is optimal for discrimination. This is borrowed from prior work (e.g., [57]) and is adopted without independent validation in SSCL.
  • domain assumption Memory buffer stores labeled exemplars and unlabeled data can be used for distillation
    CUD assumes access to a stable set of class means and to unlabeled data for distillation, which presumes the replay buffer and feature extractor are maintained appropriately.
invented entities (3)
  • Feature Space Reservation (FSR) reserved slots independent evidence
    purpose: Pre-allocated positions in the feature space for future classes, created by shaping old classes into an equiangular tight frame.
    These slots are algorithmic constructs evaluated indirectly on benchmark accuracy; no external evidence beyond the reported experiments is provided.
  • Divide-and-Conquer Pseudo-labeling (DCP) confidence groups independent evidence
    purpose: Separate treatment of high and low confidence unlabeled examples to generate pseudo-labels for both.
    This is a method design; its success is observable through downstream accuracy, so it has a falsifiable handle in the benchmark results.
  • Class-mean-anchored Unlabeled Distillation (CUD) anchors independent evidence
    purpose: Stable class means used to anchor unlabeled data for distillation to reduce forgetting.
    The anchors are computed from the model and exemplars; their utility is testable via accuracy and forgetting metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Divide-and-Conquer for Enhancing Unlabeled Learning, Stability, and Plasticity in Semi-supervised Continual Learning." pith.science (2026). https://pith.science/paper/SGAGDPAX

@misc{pith2026250805316,
  author       = {Pith},
  title        = {Pith review of: Divide-and-Conquer for Enhancing Unlabeled Learning, Stability, and Plasticity in Semi-supervised Continual Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SGAGDPAX}},
  note         = {Machine review of arXiv:2508.05316}
}
read the original abstract

Semi-supervised continual learning (SSCL) seeks to leverage both labeled and unlabeled data in a sequential learning setup, aiming to reduce annotation costs while managing continual data arrival. SSCL introduces complex challenges, including ensuring effective unlabeled learning (UL), while balancing memory stability (MS) and learning plasticity (LP). Previous SSCL efforts have typically focused on isolated aspects of the three, while this work presents USP, a divide-and-conquer framework designed to synergistically enhance these three aspects: (1) Feature Space Reservation (FSR) strategy for LP, which constructs reserved feature locations for future classes by shaping old classes into an equiangular tight frame; (2) Divide-and-Conquer Pseudo-labeling (DCP) approach for UL, which assigns reliable pseudo-labels across both high- and low-confidence unlabeled data; and (3) Class-mean-anchored Unlabeled Distillation (CUD) for MS, which reuses DCP's outputs to anchor unlabeled data to stable class means for distillation to prevent forgetting. Comprehensive evaluations show USP outperforms prior SSCL methods, with gains up to 5.94% in the last accuracy, validating its effectiveness. The code is available at https://github.com/NJUyued/USP4SSCL.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 62 canonical work pages

  1. [1]

    Prototype-sample relation distillation: towards replay-free continual learning

    Nader Asadi, MohammadReza Davari, Sudhir Mudur, Rahaf Aljundi, and Eugene Belilovsky. Prototype-sample relation distillation: towards replay-free continual learning. In Inter- national Conference on Machine Learning, 2023. 2

    work page 2023

  2. [2]

    Beyond Supervised Continual Learning: a Review

    Benedikt Bagus, Alexander Gepperth, and Timoth ´ee Lesort. Beyond supervised continual learning: a review. arXiv preprint arXiv:2208.14307, 2022. 1

    work page Pith review arXiv 2022

  3. [3]

    Rainbow memory: Continual learning with a memory of diverse samples

    Jihwan Bang, Heesu Kim, YoungJoon Yoo, Jung-Woo Ha, and Jonghyun Choi. Rainbow memory: Continual learning with a memory of diverse samples. InIEEE/CVF Conference on Computer Cision and Pattern Recognition, 2021. 2

    work page 2021

  4. [4]

    Continual semi-supervised learning through contrastive interpolation consistency

    Matteo Boschini, Pietro Buzzega, Lorenzo Bonicelli, Angelo Porrello, and Simone Calderara. Continual semi-supervised learning through contrastive interpolation consistency. Pat- tern Recognition Letters, 162:9–14, 2022. 3, 6

    work page 2022

  5. [5]

    Hy- pernetworks for continual semi-supervised learning

    Dhanajit Brahma, Vinay Kumar Verma, and Piyush Rai. Hy- pernetworks for continual semi-supervised learning. arXiv preprint arXiv:2110.01856, 2021. 1, 3

    arXiv 2021

  6. [6]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In IEEE/CVF International Conference on Computer Vision ,

    work page

  7. [7]

    Efficient lifelong learning with a- gem

    Arslan Chaudhry, Marc’Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a- gem. In International Conference on Learning Representa- tions, 2018. 6, 1

    work page 2018

  8. [8]

    Softmatch: Addressing the quantity-quality tradeoff in semi- supervised learning

    Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. Softmatch: Addressing the quantity-quality tradeoff in semi- supervised learning. In International Conference on Learn- ing Representations, 2023. 2

    work page 2023

Show all 65 references
  1. [9]

    Pg-lbo: Enhancing high-dimensional bayesian optimization with pseudo-label and gaussian process guid- ance

    Taicai Chen, Yue Duan, Dong Li, Lei Qi, Yinghuan Shi, and Yang Gao. Pg-lbo: Enhancing high-dimensional bayesian optimization with pseudo-label and gaussian process guid- ance. In AAAI Conference on Artificial Intelligence , 2024. 2

    2024

  2. [10]

    Boosting semi- supervised learning by exploiting all unlabeled data

    Yuhao Chen, Xin Tan, Borui Zhao, Zhaowei Chen, Ren- jie Song, Jiajun Liang, and Xuequan Lu. Boosting semi- supervised learning by exploiting all unlabeled data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2

    2023

  3. [11]

    Semi-supervised few-shot class-incremental learning

    Yawen Cui, Wuti Xiong, Mohammad Tavakolian, and Li Liu. Semi-supervised few-shot class-incremental learning. In IEEE International Conference on Image Processing, 2021. 7

    2021

  4. [12]

    Uncertainty-guided semi- supervised few-shot class-incremental learning with knowl- edge distillation

    Yawen Cui, Wanxia Deng, Xin Xu, Zhen Liu, Zhong Liu, Matti Pietik ¨ainen, and Li Liu. Uncertainty-guided semi- supervised few-shot class-incremental learning with knowl- edge distillation. IEEE Transactions on Multimedia , 25: 6422–6435, 2022. 7

    2022

  5. [13]

    Uncertainty-aware distillation for semi-supervised few-shot class-incremental learning

    Yawen Cui, Wanxia Deng, Haoyu Chen, and Li Liu. Uncertainty-aware distillation for semi-supervised few-shot class-incremental learning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 1, 4, 6, 7

    2023

  6. [14]

    Continual pro- totype evolution: Learning online from non-stationary data streams

    Matthias De Lange and Tinne Tuytelaars. Continual pro- totype evolution: Learning online from non-stationary data streams. In IEEE/CVF International Conference on Com- puter Vision, 2021. 2

    2021

  7. [15]

    Towards semi-supervised learning with non- random missing labels

    Yue Duan, Zhen Zhao, Lei Qi, Luping Zhou, Lei Wang, and Yinghuan Shi. Towards semi-supervised learning with non- random missing labels. In IEEE/CVF International Confer- ence on Computer Vision, 2023. 2

    2023

  8. [16]

    Mutexmatch: Semi- supervised learning with mutex-based consistency regular- ization

    Yue Duan, Zhen Zhao, Lei Qi, Lei Wang, Luping Zhou, Yinghuan Shi, and Yang Gao. Mutexmatch: Semi- supervised learning with mutex-based consistency regular- ization. IEEE Transactions on Neural Networks and Learn- ing Systems, 35(6):8441–8455, 2024. 2

    2024

  9. [17]

    Roll with the punches: Expansion and shrinkage of soft label selection for semi-supervised fine- grained learning

    Yue Duan, Zhen Zhao, Lei Qi, Luping Zhou, Lei Wang, and Yinghuan Shi. Roll with the punches: Expansion and shrinkage of soft label selection for semi-supervised fine- grained learning. In AAAI Conference on Artificial Intelli- gence, 2024. 2

    2024

  10. [18]

    Dy- namic sub-graph distillation for robust semi-supervised con- tinual learning

    Yan Fan, Yu Wang, Pengfei Zhu, and Qinghua Hu. Dy- namic sub-graph distillation for robust semi-supervised con- tinual learning. In AAAI Conference on Artificial Intelli- gence, 2024. 1, 2, 3, 4, 6

    2024

  11. [19]

    Ddgr: Continual learning with deep diffusion-based generative replay

    Rui Gao and Weiwei Liu. Ddgr: Continual learning with deep diffusion-based generative replay. In International Conference on Machine Learning, 2023. 2

    2023

  12. [20]

    A survey on semi-supervised learning for delayed partially la- belled data streams

    Heitor Murilo Gomes, Maciej Grzenda, Rodrigo Mello, Jesse Read, Minh Huong Le Nguyen, and Albert Bifet. A survey on semi-supervised learning for delayed partially la- belled data streams. ACM Computing Surveys, 55(4):1–42,

  13. [21]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Con- ference on Computer Vision and Pattern Recognition, 2016. 6

    2016

  14. [22]

    Class-incremental learning using diffusion model for distillation and replay

    Quentin Jodelet, Xin Liu, Yin Jun Phua, and Tsuyoshi Mu- rata. Class-incremental learning using diffusion model for distillation and replay. In IEEE/CVF International Confer- ence on Computer Vision, 2023. 2

    2023

  15. [23]

    Class- incremental learning by knowledge distillation with adaptive feature consolidation

    Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class- incremental learning by knowledge distillation with adaptive feature consolidation. In IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2022. 2

    2022

  16. [24]

    A soft nearest-neighbor framework for con- tinual semi-supervised learning

    Zhiqi Kang, Enrico Fini, Moin Nabi, Elisa Ricci, and Kar- teek Alahari. A soft nearest-neighbor framework for con- tinual semi-supervised learning. In IEEE/CVF International Conference on Computer Vision, 2023. 6, 2, 3

    2023

  17. [25]

    Achieving a better stability-plasticity trade-off via auxiliary networks in continual learning

    Sanghwan Kim, Lorenzo Noci, Antonio Orvieto, and Thomas Hofmann. Achieving a better stability-plasticity trade-off via auxiliary networks in continual learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1

    2023

  18. [26]

    Bal- ancing stability and plasticity through advanced null space in continual learning

    Yajing Kong, Liu Liu, Zhen Wang, and Dacheng Tao. Bal- ancing stability and plasticity through advanced null space in continual learning. In European Conference on Computer Vision, 2022. 1

    2022

  19. [27]

    Learning multiple layers of features from tiny images

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

    2009

  20. [28]

    Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks

    Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on Challenges in Representation Learning, In- ternational Conference on Machine Learning, 2013. 2

    2013

  21. [29]

    Do pre-trained models benefit equally in continual learning? In IEEE/CVF Winter Conference on Applications of Computer Vision, 2023

    Kuan-Ying Lee, Yuanyi Zhong, and Yu-Xiong Wang. Do pre-trained models benefit equally in continual learning? In IEEE/CVF Winter Conference on Applications of Computer Vision, 2023. 3

    2023

  22. [30]

    Learn to grow: A continual structure learn- ing framework for overcoming catastrophic forgetting

    Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to grow: A continual structure learn- ing framework for overcoming catastrophic forgetting. In International Conference on Machine Learning, 2019. 2

    2019

  23. [31]

    Model behavior preserving for class-incremental learning

    Yu Liu, Xiaopeng Hong, Xiaoyu Tao, Songlin Dong, Jin- gang Shi, and Yihong Gong. Model behavior preserving for class-incremental learning. IEEE Transactions on Neural Networks and Learning Systems , 34(10):7529–7540, 2022. 2

    2022

  24. [32]

    Augmented geometric distillation for data-free incremental person reid

    Yichen Lu, Mei Wang, and Weihong Deng. Augmented geometric distillation for data-free incremental person reid. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2

    2022

  25. [33]

    Learning to predict gradients for semi-supervised con- tinual learning

    Yan Luo, Yongkang Wong, Mohan Kankanhalli, and Qi Zhao. Learning to predict gradients for semi-supervised con- tinual learning. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1, 3

    2024

  26. [34]

    Metric learning for large scale image clas- sification: Generalizing to new classes at near-zero cost

    Thomas Mensink, Jakob Verbeek, Florent Perronnin, and Gabriela Csurka. Metric learning for large scale image clas- sification: Generalizing to new classes at near-zero cost. In European Conference on Computer Vision, 2012. 4

    2012

  27. [35]

    Virtual adversarial training: a regularization method for supervised and semi-supervised learning

    Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(8):1979–1993, 2018. 2

    1979

  28. [36]

    An overview of deep semi-supervised learning

    Yassine Ouali, C ´eline Hudelot, and Myriam Tami. An overview of deep semi-supervised learning. arXiv preprint arXiv:2006.05278, 2020. 1

    2006 arXiv

  29. [37]

    Class- incremental learning for action recognition in videos

    Jaeyoo Park, Minsoo Kang, and Bohyung Han. Class- incremental learning for action recognition in videos. In IEEE/CVF International Conference on Computer Vision ,

  30. [38]

    Class-incremental learn- ing with pre-allocated fixed classifiers

    Federico Pernici, Matteo Bruni, Claudio Baecchi, Francesco Turchini, and Alberto Del Bimbo. Class-incremental learn- ing with pre-allocated fixed classifiers. InInternational Con- ference on Pattern Recognition, 2021. 4

    2021

  31. [39]

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

    Gr ´egoire Petit, Adrian Popescu, Hugo Schindler, David Pi- card, and Bertrand Delezoide. Fetril: Feature translation for exemplar-free class-incremental learning. In IEEE/CVF Winter Conference on Applications of Computer Vision ,

  32. [40]

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

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

  33. [41]

    icarl: Incremental clas- sifier and representation learning

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental clas- sifier and representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2, 3, 4, 6, 1

    2017

  34. [42]

    Memory-efficient semi-supervised continual learning: The world is its own replay buffer

    James Smith, Jonathan Balloch, Yen-Chang Hsu, and Zsolt Kira. Memory-efficient semi-supervised continual learning: The world is its own replay buffer. In International Joint Conference on Neural Networks, 2021. 1, 2, 3

    2021

  35. [43]

    Always be dreaming: A new approach for data-free class-incremental learning

    James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In IEEE/CVF International Conference on Computer Vision ,

  36. [44]

    Coda-prompt: Contin- ual decomposed attention-based prompting for rehearsal-free continual learning

    James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Contin- ual decomposed attention-based prompting for rehearsal-free continual learning. In IEEE/CVF Conference ...

    2023

  37. [45]

    Fixmatch: Simplifying semi-supervised learning with consistency and confidence

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems , 202...

    2020

  38. [46]

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems, 2017. 2

    2017

  39. [47]

    Con- trastive multiview coding

    Yonglong Tian, Dilip Krishnan, and Phillip Isola. Con- trastive multiview coding. In European Conference on Com- puter Vision, 2020. 5

    2020

  40. [48]

    Foster: Feature boosting and compression for class- incremental learning

    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, 2022. 2

    2022

  41. [49]

    Ordisco: Effective and efficient usage of incremental unlabeled data for semi-supervised continual learning

    Liyuan Wang, Kuo Yang, Chongxuan Li, Lanqing Hong, Zhenguo Li, and Jun Zhu. Ordisco: Effective and efficient usage of incremental unlabeled data for semi-supervised continual learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1, 3, 6

    2021

  42. [50]

    A comprehensive survey of continual learning: theory, method and application

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1

    2024

  43. [51]

    Freematch: Self-adaptive thresholding for semi-supervised learning

    Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, et al. Freematch: Self-adaptive thresholding for semi-supervised learning. In International Conference on Learning Representations, 2022. 2

    2022

  44. [52]

    Learning to prompt for continual learning

    Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jen- nifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2

    2022

  45. [53]

    Continual learning: A review of techniques, challenges and future directions

    Buddhi Wickramasinghe, Gobinda Saha, and Kaushik Roy. Continual learning: A review of techniques, challenges and future directions. IEEE Transactions on Artificial Intelli- gence, 2023. 1

    2023

  46. [54]

    Self-training with noisy student improves imagenet clas- sification

    Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet clas- sification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2

    2020

  47. [55]

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

    Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynam- ically expandable representation for class incremental learn- ing. In IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, 2021. 2, 3, 6, 1

    2021

  48. [56]

    A survey on deep semi-supervised learning

    Xiangli Yang, Zixing Song, Irwin King, and Zenglin Xu. A survey on deep semi-supervised learning. IEEE Trans- actions on Knowledge and Data Engineering , 35(9):8934– 8954, 2022. 1

    2022

  49. [57]

    Neural collapse inspired feature- classifier alignment for few-shot class-incremental learning

    Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature- classifier alignment for few-shot class-incremental learning. In International Conference on Learning Representations ,

  50. [58]

    Few-shot incremental learning with contin- ually evolved classifiers

    Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with contin- ually evolved classifiers. In IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2021. 7

    2021

  51. [59]

    Class-incremental learning via deep model consolidation

    Junting Zhang, Jie Zhang, Shalini Ghosh, Dawei Li, Serafet- tin Tasci, Larry Heck, Heming Zhang, and C-C Jay Kuo. Class-incremental learning via deep model consolidation. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2020. 2

    2020

  52. [60]

    Memory-efficient class-incremental learning for image classification

    Hanbin Zhao, Hui Wang, Yongjian Fu, Fei Wu, and Xi Li. Memory-efficient class-incremental learning for image classification. IEEE Transactions on Neural Networks and Learning Systems, 33(10):5966–5977, 2021. 2

    2021

  53. [61]

    Simmatch: Semi-supervised learning with similarity matching

    Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2022. 2

    2022

  54. [62]

    Simmatchv2: Semi- supervised learning with graph consistency

    Mingkai Zheng, Shan You, Lang Huang, Chen Luo, Fei Wang, Chen Qian, and Chang Xu. Simmatchv2: Semi- supervised learning with graph consistency. In IEEE/CVF International Conference on Computer Vision, 2023. 2

    2023

  55. [63]

    Forward compatible few- shot class-incremental learning

    Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shil- iang Pu, and De-Chuan Zhan. Forward compatible few- shot class-incremental learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 4

    2022

  56. [64]

    Class-incremental learning: A survey

    Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De- Chuan Zhan, and Ziwei Liu. Class-incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2

    2024

  57. [65]

    Self-promoted prototype refinement for few-shot class- incremental learning

    Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class- incremental learning. In IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2021. 7 Table 11. Ablation studies on memory buffer size of exemplar set...

    2021

Pith tools

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