Pith. sign in

REVIEW 4 major objections 5 minor 39 references

DeCLIP: Decoupled Prompting for Multi-Label Class-Incremental Learning with CLIP

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A replay-free framework with per-class positive and negative prompts plus an adaptive temperature schedule is claimed to be the first multi-label class-incremental learner to exceed 80% mAP and 70% F1.

desk verdict A legitimate prompt-based MLCIL advance with a real but addressable evaluation problem: the headline margins depend on PCD hyperparameters tuned on the same test benchmarks, and the submission has a title/abstract mismatch. read the letter →

arxiv 2509.23335 v3 pith:TTVBQYVR submitted 2025-09-27 cs.CV

Kaile Du , Zihan Ye , Junzhou Xie , Yixi Shen , Yuyang Li , Fuyuan Hu , Ling Shao , Guangcan Liu
show 2 more authors
Joost van de Weijer Fan Lyu
This is my paper · ORCID
classification cs.CV
keywords multi-labelclass-incrementallearningCLIPpromptcatastrophicforgettingfalsepositivesreplay-freeconfidencecalibrationDual-DecoupledPrompting
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

Multi-label class-incremental learning (MLCIL) models must recognize an ever-growing set of co-occurring categories without storing old samples. The paper argues that CLIP-based prompting fails in this setting for two reasons: shared prompts entangle the semantics of co-occurring classes, and task-level partial labeling makes the model confidently predict absent classes (false positives). To fix this, it introduces Dual-Decoupled Prompting (DDP), which gives each class its own positive and negative prompts in both text and vision branches, reframing recognition as per-class binary decisions, and adds Progressive Confidence Decoupling (PCD), an inference-time temperature schedule that grows with the cumulative class count. On MS-COCO and PASCAL VOC, DDP reports consistent gains and is the first replay-free MLCIL method to exceed 80% mean average precision and 70% F1 on the standard B40-C10 benchmark, suggesting that forgetting and false positives can be controlled without exemplar memory.

What carries the argument

The key object is the per-class positive–negative prompt pair P_c = {P_c+, P_c-} applied in both modalities, which converts multi-label recognition into a set of binary presence/absence judgments via cosine similarities s_c+ and s_c- fed through a binary softmax. PCD is the temperature schedule τ(t) = 1 + (τ_max − 1)·((|C_1:t| − |C_1|)/(|C_1:T| − |C_1|))^γ, which raises confidence sharpness as tasks accumulate, suppressing false positives at inference without touching representations. Interlayer prompting attaches these prompts to the last five MSA layers and shares them across layers, giving efficiency and better semantics.

What would settle it

Run the B40-C10 benchmark with the PCD schedule computed only from classes seen so far (or with a single fixed temperature) and check whether the last mAP and FPR still beat the strongest previous methods; if the gap collapses or FPR returns to roughly 25%, the claimed suppression is an artifact of the schedule's dependence on the full sequence.

Watch

Extended reading notes

Core claim

The central claim is that the two pathologies of MLCIL — semantic confusion from co-occurring labels and true-negative/false-positive confusion from partial labeling — can be addressed through prompting alone, without replay or selectors. DDP assigns a one-to-one prompt set to every class: a positive prompt that signals presence and a negative prompt that signals absence, in both the text and visual encoders. Similarities are combined in a binary softmax whose temperature τ(t) is raised progressively as more classes are seen, acting as a curriculum that suppresses over-confidence in unobserved negatives. Past prompts are frozen as knowledge anchors, so new tasks add parameters rather than ov

Load-bearing premise

The method depends on knowing the total number of classes across all future tasks to define the PCD schedule, and on τ_max and γ chosen per benchmark; if the task horizon is unknown or the schedule does not transfer, the reported false-positive suppression and performance margins may not hold.

Editorial extensions

If this is right

  • If DDP's results hold, replay-free MLCIL becomes a practical option: each new class adds a small set of prompt parameters and old prompts are never modified, so privacy-friendly continual learning needs no stored exemplars.
  • PCD implies that false-positive suppression can be achieved at inference time via a global temperature schedule rather than by changing the training loss, decoupling representation learning from confidence calibration.
  • The one-to-one prompting design suggests the label space can be expanded additively with a fixed per-class parameter budget, making the method scalable to large or open-ended class sets.
  • Because the method requires no selector or auxiliary classification head, it reduces computational overhead relative to prompt-selection approaches, which could matter for deployment on edge devices.
  • The reported FPR reduction (25.4% → 2.8%) indicates that most false positives in this setting are a confidence-calibration artifact rather than a representation limitation.

Reading between the lines

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

  • If the total class count is unknown in advance, the PCD schedule cannot be computed as written; one could test an online variant that estimates the horizon or adapts τ based on observed task statistics, which the paper does not address.
  • The temperature schedule resembles standard confidence-calibration techniques applied progressively; its success on two benchmarks hints that a class-order-dependent schedule might yield further gains, though the paper reports a single global schedule.
  • The per-class binary reformulation may also benefit single-label CIL or open-world recognition: the negative prompts encode 'absence' semantics that could help reject unknown classes, but the paper does not explore those settings.
  • Since τ_max and γ are tuned per benchmark on the same test sets, the reported margins may be optimistic; a reader should check whether the advantage persists under a fixed schedule held out across benchmarks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper (titled DeCLIP in the metadata, DDP in the body) proposes a replay-free, parameter-efficient prompt-based framework for multi-label class-incremental learning (MLCIL). The method assigns each class a one-to-one pair of positive and negative prompts in both text and visual encoders, computing positive/negative similarities and applying a confidence-adjusted binary softmax. A Progressive Confidence Decoupling (PCD) module schedules a temperature tau(t) that grows with the cumulative number of seen classes, suppressing false positives at inference. Experiments on MS-COCO and PASCAL VOC, including long-sequence and high-frequency increment settings, report consistent gains over prior methods and claim that DDP is the first replay-free MLCIL method to surpass 80% mAP and 70% F1 on MS-COCO B40-C10.

Significance. If the reported results are reproducible, the paper makes a useful contribution: it identifies two distinct failure modes in prompt-based MLCIL (semantic confusion and true-negative/false-positive confusion), proposes a simple class-specific prompting mechanism, and validates it with broad ablations across many baselines. The selector-free, interlayer-prompting design is pragmatically attractive, and the positive/negative prompt decomposition is a principled way to turn multi-label recognition into per-class binary tasks. The central claims, however, depend on the PCD schedule, whose hyperparameters are selected on the evaluation benchmarks themselves, and on single-run numbers; these issues must be resolved before the SOTA claim is reliable.

major comments (4)
  1. [Section 3.2.2, Eq. (3), Appendix A] The PCD schedule has two free hyperparameters, tau_max and gamma, and Appendix A (Figures A.1 and A.2) selects them by measuring Last/Avg CF1/OF1 on the same benchmark configurations used for the main results (VOC B4-C2 and COCO B0-C5). The main text does not report the tau_max and gamma values used for the headline B40-C10 and B0-C10 scenarios. Because PCD directly modifies prediction confidences at inference, choosing these values on the test set can inflate all reported gains, including the claimed 'first to surpass 80% mAP / 70% F1' result. The abstract's assertion of 'without specific tuning' is not supported by the appendix, which shows per-dataset tuning. Please provide a validation-based selection procedure (or nested cross-validation), report the chosen values for every scenario, and show sensitivity of the main metrics to tau_max and gamma.
  2. [Eq. (3)] The temperature schedule requires |C_{1:T}|, the total number of classes across all future tasks. This is an oracle quantity in an incremental stream whose total length is not known in advance. The paper frames MLCIL as a continuously expanding label space, yet PCD cannot be applied to a genuinely open-ended sequence without knowing the final class count. A concrete adaptation should be given (e.g., schedule based only on observed classes, or a domain-agnostic default), and the effect of misspecifying |C_{1:T}| should be measured. As written, the central PCD mechanism is only defined for benchmarks with a known, finite task sequence.
  3. [Section 3.2.2 and Table 5] The paper claims PCD reduces the false-positive rate from 25.4% to 2.8%, but the definition of FPR is not given. In multi-label classification, FPR depends on a decision threshold, and no threshold or operating point is specified. It is also unclear whether this FPR is computed on the final-task test set, on all tasks, or on the union of test sets. Without this information, the FPR suppression claim cannot be verified or compared with other methods. Please define FPR, state the thresholding rule, and report the corresponding precision/recall or the ROC curve if threshold dependence is material.
  4. [Tables 1, 2, A.1, A.2] All reported results are single-run numbers with no standard deviations, confidence intervals, or significance tests. The headline claims are based on margins of 1–6 points over strong baselines, and the PCD hyperparameters are tuned on the same benchmarks. Under these conditions, the 'first to surpass 80% mAP and 70% F1' assertion may not be robust to run-to-run variation. Please provide multiple seeds (at least 3–5) with mean±std, or otherwise demonstrate that the reported improvements are larger than the noise level.
minor comments (5)
  1. [Title / Abstract / Body] The arXiv title and abstract call the method DeCLIP, while the full text, equations, and tables call it DDP (Dual-Decoupled Prompting). Please unify the name throughout.
  2. [Eq. (1) and Section 3.2.1] The construction of the negative text prompt P_c^- is not described. Is it a literal 'not c' prompt or a learned embedding? Please clarify how the absence cue is injected into the text encoder.
  3. [Figure 4(b)] The comparison with asymmetric loss (AL) is only qualitative. Please provide a quantitative description of the AL baseline (loss formulation, hyperparameters, and the numbers behind Figure 4(b)) or move this comparison to a table.
  4. [Tables 3–5 captions] The ablation tables do not state the scenario in their captions. The text says they are on VOC B4-C2, but the captions should include this information for self-containment.
  5. [Tables 1–2] Several cells contain dashes with no explanation (e.g., some OF1 and Avg entries). If a method was not evaluated for that metric, say so explicitly rather than leaving an unannotated dash.

Circularity Check

1 steps flagged · score 6.0 of 10

PCD's claimed gains are selected on the same test metrics used for the headline results, making the main comparison partially circular.

  1. fitted input called prediction [Appendix A (Hyperparameter Analysis), Eqs. (2)-(3), and Section 4.2 / Table 1]
    "Figure A.1 and Figure A.2 illustrate the influence of τmax and γ on PCD. On VOC B4-C2, we choose τmax=7 and γ=0.2, which lead to clear improvements in both Last and Avg. CF1/OF1. On COCO B0-C5, we choose τmax=3 and γ=0.7, which also enhance both Last and Avg. metrics."

    The PCD schedule (Eq. 3) is controlled by τmax and γ, which directly modulate the binary softmax in Eq. 2 and therefore all reported metrics. The appendix selects these two values by inspecting Last/Avg CF1/OF1 on the same test sets whose results are later reported as evidence that DDP 'consistently outperforms prior methods' and 'is the first replay-free MLCIL approach to exceed 80% mAP and 70% F1'. The headline gains are thus not independent predictions of a fixed, tuning-free method; they are, in part, the selection criterion used to set the inference-time tempering schedule. This is a fitted input presented as a generic improvement, so the claimed margins are partially constructed by test-set choice.

full rationale

The semantic-decoupling prompt design and the comparison protocol are not circular: DDP is tested on external benchmarks (MS-COCO, PASCAL VOC, NUS-WIDEseq) against published baselines, and the class-specific positive/negative prompting is a genuine architectural contribution whose ablations compare internal variants. The principal circularity is confined to the PCD component: its hyperparameters τmax and γ are selected by measuring the very Last/Avg CF1/OF1 metrics used for the main SOTA claims, and Eq. (3) requires the total number of classes |C1:T|, which is future knowledge in a true incremental stream. These issues do not make the whole framework definitionally equivalent to its inputs (the prompt architecture and BCE training are independent), but they undercut the claimed parameter-free adaptivity and mean the reported >80% mAP / >70% F1 margins are at least partly a result of test-set selection rather than an out-of-sample prediction. No load-bearing self-citation chain or uniqueness-imported-from-authors pattern is present; prior work by the same group is cited mainly as baselines.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central empirical result rests on standard benchmark assumptions plus a handful of hand-tuned hyperparameters. No new physical or conceptual entities are introduced; the prompts are parameters, not entities.

free parameters (5)
  • tau_max = 7 on VOC B4-C2, 3 on COCO B0-C5
    Ceiling of the progressive temperature schedule; selected by grid evaluation on the same benchmark (Figure A.1).
  • gamma = 0.2 on VOC B4-C2, 0.7 on COCO B0-C5
    Exponent controlling the growth of tau(t); selected on the same benchmark (Figure A.2).
  • prompt insertion depth = last 5 layers
    The authors tested different depths (Figure 4c) and chose the last five layers on MS-COCO B0-C10; this choice is reflected in all results.
  • prompt length = 16
    Sequence length of prompts, set to 16 for all experiments (Section 4.1).
  • training epochs = 20
    Number of epochs for prompt training (Section 4.1).
assumptions (4)
  • domain assumption Pre-trained CLIP (ViT-B/16) provides a shared embedding space where frozen encoders plus lightweight prompts suffice for multi-label recognition.
    The whole method relies on CLIP features being informative; no fine-tuning of the backbone is performed (Section 3.2).
  • domain assumption The total number of classes across all tasks, |C_{1:T}|, is known in advance and used to set the PCD schedule.
    Eq. 3 normalizes by |C_{1:T}|; this is standard in benchmark CIL but fails in open-ended streams.
  • domain assumption Treating all unannotated classes within the current task as negatives is an acceptable learning signal despite future-task objects being present in the images.
    BCE loss (Eq. 5) uses current-task labels only; the paper acknowledges TN-FP confusion and relies on PCD to correct it rather than a more principled loss.
  • ad hoc to paper The monotonic increase of temperature with class count is the right curriculum for confidence calibration.
    This is the defining hypothesis of PCD; it is motivated heuristically and validated only on the two benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeCLIP: Decoupled Prompting for Multi-Label Class-Incremental Learning with CLIP." pith.science (2026). https://pith.science/paper/TTVBQYVR

@misc{pith2026250923335,
  author       = {Pith},
  title        = {Pith review of: DeCLIP: Decoupled Prompting for Multi-Label Class-Incremental Learning with CLIP},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TTVBQYVR}},
  note         = {Machine review of arXiv:2509.23335}
}
read the original abstract

Multi-label class-incremental learning (MLCIL) continuously expands the label space while recognizing multiple co-occurring categories, making catastrophic forgetting a central challenge. Recent class-incremental learning methods have increasingly adopted CLIP as their backbone. However, we find that applying CLIP to MLCIL exhibits two critical issues: entanglement of class-specific cues in shared visual representations and high false-positive rates (FPR) under task-level partial labeling. We propose DeCLIP, a replay-free and parameter-efficient framework for CLIP-based MLCIL. DeCLIP uses Decoupled Prompting to learn class-specific positive and negative prompts in both visual and textual modalities, enabling class-conditioned vision-language matching and reducing representation entanglement. Only new-category prompts are optimized, previous prompts remain unchanged, preserving prior knowledge and mitigating catastrophic forgetting without replay. DeCLIP further incorporates Adaptive Similarity Tempering, an inference-time strategy that adapts similarity-tempering strength to the incremental configuration, suppressing false positives without specific tuning. Experiments on MS-COCO, PASCAL VOC, and the real-world NUS-WIDEseq benchmark demonstrate consistent improvements over prior methods with a few trainable parameters.

Figures

Figures reproduced from arXiv: 2509.23335 by the authors.

Figure 1
Figure 1. Comparison of prompt-based methods: two multi-label (ML) confusions and decoupling. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall pipeline of DDP. Training stage (bottom): prompts from past tasks are frozen [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Evidence for Decoupling in DDP. (a) t-SNE of class-specific prompts, showing clearer [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: (a) Illustration of the PCD mapping functions under different values of [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Comparison of prompt-based CIL methods in terms of last and average performance. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The results of multi-task MLCIL on challenging VOC B4-C2 and COCO B0-C5, where a [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Visualization of DDP using the last task model. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 3 linked inside Pith

  1. [1]

    Curriculum learning

    Yoshua Bengio, J \'e r \^o me Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the International Conference on Machine Learning, pp.\ 41--48, 2009

    2009

  2. [2]

    Dark experience for general continual learning: a strong, simple baseline

    Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. In Proceedings of the Advances in Neural Information Processing Systems, pp.\ 15920--15930, 2020

    2020

  3. [3]

    Rebalancing batch normalization for exemplar-based class-incremental learning

    Sungmin Cha, Sungjun Cho, Dasol Hwang, Sunwon Hong, Moontae Lee, and Taesup Moon. Rebalancing batch normalization for exemplar-based class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 20127--20136, 2023

    2023

  4. [4]

    Towards calibrated multi-label deep neural networks

    Jiacheng Cheng and Nuno Vasconcelos. Towards calibrated multi-label deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 27589--27599, 2024

    2024

  5. [5]

    Less is more: Summarizing patch tokens for efficient multi-label class-incremental learning

    Thomas De Min, Massimiliano Mancini, St \'e phane Lathuili \`e re, Subhankar Roy, and Elisa Ricci. Less is more: Summarizing patch tokens for efficient multi-label class-incremental learning. arXiv preprint arXiv:2405.15633, 2024

    arXiv 2024

  6. [6]

    Knowledge restore and transfer for multi-label class-incremental learning

    Songlin Dong, Haoyu Luo, Yuhang He, Xing Wei, Jie Cheng, and Yihong Gong. Knowledge restore and transfer for multi-label class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 18711--18720, 2023

    2023

  7. [7]

    Podnet: Pooled outputs distillation for small-tasks incremental learning

    Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proceedings of the European Conference on Computer Vision, pp.\ 86--102, 2020

    2020

  8. [8]

    Multi-label continual learning using augmented graph convolutional network

    Kaile Du, Fan Lyu, Linyan Li, Fuyuan Hu, Wei Feng, Fenglei Xu, Xuefeng Xi, and Hanjing Cheng. Multi-label continual learning using augmented graph convolutional network. IEEE Transactions on Multimedia, 26: 0 2978--2992, 2024 a

    2024

Show all 39 references
  1. [9]

    Confidence self-calibration for multi-label class-incremental learning

    Kaile Du, Yifan Zhou, Fan Lyu, Yuyang Li, Chen Lu, and Guangcan Liu. Confidence self-calibration for multi-label class-incremental learning. In European Conference on Computer Vision, pp.\ 234--252, 2024 b

    2024

  2. [10]

    Rebalancing multi-label class-incremental learning

    Kaile Du, Yifan Zhou, Fan Lyu, Yuyang Li, Junzhou Xie, Yixi Shen, Fuyuan Hu, and Guangcan Liu. Rebalancing multi-label class-incremental learning. In Proceedings of the AAAI conference on artificial intelligence, volume 39, pp.\ 16372--16380, 2025

    2025

  3. [11]

    The pascal visual object classes (voc) challenge

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 2010

    2010

  4. [12]

    Dualcoop++: Fast and effective adaptation to multi-label recognition with limited annotations

    Ping Hu, Ximeng Sun, Stan Sclaroff, and Kate Saenko. Dualcoop++: Fast and effective adaptation to multi-label recognition with limited annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46 0 (5): 0 3450--3462, 2023

    2023

  5. [13]

    Mind the gap: Preserving and compensating for the modality gap in clip-based continual learning

    Linlan Huang, Xusheng Cao, Haori Lu, Yifan Meng, Fei Yang, and Xialei Liu. Mind the gap: Preserving and compensating for the modality gap in clip-based continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025

    2025

  6. [14]

    Ovor: Oneprompt with virtual outlier regularization for rehearsal-free class-incremental learning

    Wei-Cheng Huang, Chun-Fu Chen, and Hsiang Hsu. Ovor: Oneprompt with virtual outlier regularization for rehearsal-free class-incremental learning. In Proceedings of the International Conference on Learning Representations, 2024

    2024

  7. [15]

    Imbalanced continual learning with partitioning reservoir sampling

    Chris Dongjoo Kim, Jinseo Jeong, and Gunhee Kim. Imbalanced continual learning with partitioning reservoir sampling. In Proceedings of the European Conference on Computer Vision, pp.\ 411--428, 2020

    2020

  8. [16]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations, 2015

    2015

  9. [17]

    Overcoming catastrophic forgetting in neural networks

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. National Academy of Sciences, 114 0 (13): 0 ...

    2017

  10. [18]

    Learning without forgetting

    Z Li and D Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40 0 (12): 0 2935--2947, 2017

    2017

  11. [19]

    Optimizing class distribution in memory for multi-label online continual learning

    Yan-Shuo Liang and Wu-Jun Li. Optimizing class distribution in memory for multi-label online continual learning. arXiv preprint arXiv:2209.11469, 2022

    2022 arXiv

  12. [20]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll \'a r, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, pp.\ 740--755, 2014

    2014

  13. [21]

    D3former: Debiased dual distilled transformer for incremental learning

    Abdelrahman Mohamed, Rushali Grandhe, KJ Joseph, Salman Khan, and Fahad Khan. D3former: Debiased dual distilled transformer for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 2420--2429, 2023

    2023

  14. [22]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on M...

    2021

  15. [23]

    icarl: Incremental classifier and representation learning

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 2001--2010, 2017

    2001

  16. [24]

    Experience replay for continual learning

    David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. In Proceedings of the Advances in Neural Information Processing Systems, pp.\ 350--360, 2019

    2019

  17. [25]

    Progress & compress: A scalable framework for continual learning

    Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In Proceedings of the International Conference on Machine Learning, pp.\ 4528--...

    2018

  18. [26]

    Coda-prompt: Continual 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: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEE...

    2023

  19. [27]

    Topology-preserving class-incremental learning

    Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XIX 16, pp.\ 254--270. Springer, 2020

    2020

  20. [28]

    Cut out and replay: A simple yet versatile strategy for multi-label online continual learning

    Xinrui Wang, Shao-yuan Li, Jiaqiang Zhang, and Songcan Chen. Cut out and replay: A simple yet versatile strategy for multi-label online continual learning. In Proceedings of the International Conference on Machine Learning, 2025

    2025

  21. [29]

    Dualprompt: Complementary prompting for rehearsal-free continual learning

    Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pp.\ 631--648. Springe...

    2022

  22. [30]

    Learning to prompt for continual learning

    Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 139--149, 2022 b

    2022

  23. [31]

    Large scale incremental learning

    Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 374--382, 2019

    2019

  24. [32]

    Specifying what you know or not for multi-label class-incremental learning

    Aoting Zhang, Dongbao Yang, Chang Liu, Xiaopeng Hong, and Yu Zhou. Specifying what you know or not for multi-label class-incremental learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 22345--22353, 2025

    2025

  25. [33]

    Dynamic prompt adjustment for multi-label class-incremental learning

    Haifeng Zhao, Yuguang Jin, and Leilei Ma. Dynamic prompt adjustment for multi-label class-incremental learning. arXiv preprint arXiv:2501.00340, 2024

    2024 arXiv

  26. [34]

    Few-shot class-incremental learning via class-aware bilateral distillation

    Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class-incremental learning via class-aware bilateral distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11838--11847, 2023

    2023

  27. [35]

    Few-shot class-incremental learning by sampling multi-phase tasks

    Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Few-shot class-incremental learning by sampling multi-phase tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    2022

  28. [36]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  29. [37]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  30. [38]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  31. [39]

    =5' 7&o6ؒv 9 Ӵe ̙ ͸0k 5;|H݇d VRR R8u!5uU n 5 eQJnr /- Գ[P ? X>Z 1W^ M@nl t!->NteUZ I R 﫰:3 skQ Tx

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.