Pith. sign in

REVIEW 4 major objections 4 minor 38 references

Fresh-CL: Feature Realignment through Experts on Hypersphere in Continual Learning

T0 review · 4 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 Fresh-CL claims that fixed, maximally separated hypersphere targets combined with a router that selects per-task expert projections reduce feature drift and outperform the strongest continual-learning baseline on 11 datasets.

desk verdict Fresh-CL is a plausible MOA extension with fixed ETF targets and expert freezing that reports consistent gains, but a load-bearing label-mapping ambiguity and an unfair compute comparison undercut the headline claims. read the letter →

arxiv 2501.02198 v2 pith:JMKXWVJV submitted 2025-01-04 cs.LG cs.CV

Zhongyi Zhou , Yaxin Peng , Pin Yi , Minjie Zhu , Chaomin Shen This is my paper · ORCID
classification cs.LGcs.CV
keywords continuallearningcatastrophicforgettingneuralcollapsesimplexequiangulartightframemixtureofexpertsmulti-taskincrementalhyperspherefeatureseparationfine-grainedclassification
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

Continual learning models forget because features of old and new domains become entangled as tasks arrive. Fresh-CL attacks this by giving every class a pre-assigned 'pseudo-target' on a hypersphere: a fixed set of equiangular vectors with maximal pairwise separation, so classes are pulled toward distinct directions. The paper argues that a single such target frame is not enough, because the learned projection to it drifts when a new domain arrives. It therefore adds a mixture of experts, with a per-task router sending each input to the top two expert projections, each carrying its own equiangular target frame; frequently used experts are frozen to preserve old domains. On 11 datasets under multi-task incremental learning, the method reports 2 percent higher final accuracy than the strongest baseline (and 1.1 percent in the few-shot version), with its clearest gains on fine-grained datasets.

What carries the argument

The load-bearing object is the simplex equiangular tight frame (ETF): a set of $K$ unit vectors in $\mathbb{R}^d$ with equal pairwise inner products, which are the maximally separated points on a hypersphere. Fresh-CL uses a fixed ETF as pseudo-targets for a regression loss, and treats each expert as a separate projection onto its own ETF. A task-specific router computes gating weights $G_t = \text{Softmax}(\text{Topk}(R_t(\hat{\mu}_i)))$ over experts, so each task is mapped onto the subspaces whose target frames best fit its domain; after a task, the most-used experts are frozen, which the paper credits with preventing representation shift.

What would settle it

A reader could count the cumulative number of classes across the 11-task sequence; if it exceeds the feature dimension $d$ (768 for the ViT-B/16 encoder used), then the fixed ETF in Eq. (2) cannot assign a unique target to each class. Running the method with a strict global label-to-target mapping and checking whether early-task accuracy drops would settle whether the reported separation is achieved by the ETF geometry itself or by a local label remapping that weakens inter-task separation.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a fixed simplex equiangular tight frame (ETF) can serve as a stable reference grid for class features in continual learning only if the model is given multiple such frames and a way to choose among them per task. Each expert in Fresh-CL is a projection layer tied to a distinct ETF, and a lightweight router selects the top two experts for each sample. The dot-product regression loss $L_{DR} = \frac{1}{2}(\hat{w}_{y_i}^T \hat{\mu}_i - 1)^2$ pulls the normalized feature toward the pseudo-target indexed by the class label. The paper reports that this setup improves feature separation within and across tasks and yields final-accuracy gains of 2.0 percent (full-shot) and 1.1 percent (few-shot) over the strongest comparison method on the 11-dataset MTIL benchmark, with the largest improvements on fine-grained datasets.

Load-bearing premise

The method assumes that one fixed d-dimensional equiangular frame can hold a distinct pseudo-target for every class encountered across all tasks, with the class label directly selecting the target column.

Editorial extensions

If this is right

  • If the 2-percent gain holds, rehearsal-free continual learning can be improved without storing old data, merely by fixing target geometry and routing new tasks to disjoint expert projections.
  • The freezing of frequently used experts after each task implies a growing but sparse set of frozen projections; the model's parameter footprint stays moderate (the paper reports 56.22 MB trainable), making the approach attractive for memory-limited continual learning.
  • The largest gains on fine-grained datasets suggest the method specifically helps when classes within a domain are easy to confuse, because the equiangular targets force them apart even when the backbone features are similar.
  • The per-task router and frozen experts provide a direct mechanism against catastrophic forgetting: old domains keep their projection heads unchanged, so previously learned features are not overwritten by gradient updates from new tasks.

Reading between the lines

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

  • The paper does not specify how class labels are mapped to ETF columns when the total class count can exceed the feature dimension; if labels are re-indexed within each task, the claimed cross-task separation may come from the router and freezing rather than from a single global ETF.
  • Because the backbone is frozen, Fresh-CL is essentially a study of how to arrange classification heads and targets on top of fixed features; a natural test is whether the same gains appear when the backbone is trainable or when tasks share classes.
  • With 22 experts and top-2 routing, the method resembles a sparse linear probe ensemble; comparing against an ensemble of ordinary (non-ETF) per-task linear probes would isolate how much of the improvement comes from the equiangular target geometry itself.
  • The authors freeze experts by frequency of use, which may create a routing imbalance where early tasks dominate expert occupancy; an explicit load-balancing loss could change both accuracy and the forgetting profile.
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 paper proposes Fresh-CL, a rehearsal-free method for multi-task incremental learning (MTIL) built on a frozen CLIP ViT-B/16 backbone. Each expert is a projection layer trained against a fixed, predefined simplex equiangular tight frame (ETF) pseudo-target through a dot-regression loss, and a per-task router selects the top-2 experts for each sample; frequently used experts are frozen after each task but remain selectable. The authors report last-task accuracy over 11 datasets and claim a 2.0 percentage-point average improvement over the strongest baseline in full-shot MTIL and 1.1 points in few-shot MTIL, with an ablation over the number of experts.

Significance. If substantiated, Fresh-CL would be an attractive rehearsal-free, parameter-efficient MTIL baseline: the ETF pseudo-targets are parameter-free, the sparse top-2 routing limits trainable parameters to 56.22MB, and the paper provides side-by-side comparisons across 11 datasets. The idea of using multiple fixed ETF subspaces selected by a router is distinctive and worth investigating. However, the central empirical claim is currently clouded by a compute mismatch, an unresolved target-indexing/capacity ambiguity, and an ablation that contradicts the stated motivation, so the result is not yet established.

major comments (4)
  1. [Section III-A, Eq. (2), Fig. 2 caption] The indexing of the pseudo-target by the class label y_i is not well-defined in the 11-task MTIL setting. With K=d and d=512 for the standard CLIP ViT-B/16 embedding (or d=768 penultimate), an ETF in R^d has at most d+1 mutually equiangular columns, but the 11 tasks in Table I contain about 1,201 classes in total. If y_i is a global class index, Eq. (2) indexes columns beyond K for most classes; if y_i is a per-task or per-dataset label, then classes from different tasks share the same pseudo-target column, and L_DR pulls their normalized features to the same point, contradicting the claimed inter-task separation and the Fig. 2 caption statement that pseudo-targets are predefined for each class of all tasks. The paper must specify the label-to-column mapping, the label space of each expert's ETF, and how K=d can supply a unique target for every class; as written, this is a load-bearing ambiguity.
  2. [Section IV-A and Table I] The headline '+2.0%' improvement is obtained by Fresh-CL(3k), while the implementation details in Section IV-A state that the method is trained for 1k iterations per task. The 1k row gives +1.5% over MOA, and no baseline is shown with 3k iterations. The 2.0% claim is therefore confounded by a threefold increase in optimization steps. The authors should either headline the compute-matched 1k comparison or provide 3k results for the baselines.
  3. [Table III] The ablation row labeled 'Fixed ETF w/o MoE' reports an average last accuracy of 53.5%, which is 23.8 points below Continual-FT (77.3%) and 30.9 points below the 11-expert variant. This directly contradicts the paper's motivation that inducing neural collapse through an ETF improves feature separation. If the single-ETF loss is this harmful on its own, the improvement of the full method cannot be attributed to 'combining ETF and MoE' as stated. The authors need to explain what is trained in that row and why the core ETF component is so detrimental when used alone.
  4. [Section IV-B, Tables I and II] All results are reported as single numbers with no error bars and no stated number of seeds. The claimed gains are small averages (1.5, 2.0, and 1.1 percentage points) over a strong CLIP-based baseline and may be within run-to-run variation. Reporting mean and standard deviation over at least three independent runs is necessary to support the central empirical claim.
minor comments (4)
  1. [Section II-B, Eq. (1)] The formula defining the ETF is only meaningful for K ≥ 2 and does not state whether the vectors w_i are normalized; please state the convention explicitly.
  2. [Section III-B, Eqs. (3)-(4)] The index i is used both for the expert and for the input feature ƒÂµ_i; rename the feature of sample x_k (e.g., ƒÂµ_k) to avoid confusion between the expert summation index and the sample index.
  3. [Table I] The table caption should define Fresh-CL(1k) and Fresh-CL(3k) and state which row corresponds to the headline claim in the abstract, since Section IV-A only mentions 1k iterations.
  4. [Section IV-C] The sentence 'a probability of number of experts/11 is applied to freeze experts after the training stage for each task' is unclear; please state the exact freezing rule and how it depends on the expert count.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ETF geometry, DR loss, and MoE routing are defined independently of the reported accuracies, and the performance claims are empirical comparisons against external baselines.

full rationale

The paper's derivation chain is self-contained in the relevant sense. The ETF pseudo-targets are fixed mathematical objects defined by Eq. (1), the dot-regression loss in Eq. (2) is a training objective, and the expert routing in Eqs. (3)-(4) defines a model class. No parameter is fitted to the reported 2% or 1.1% improvements; those numbers are measured against external baselines in Tables I and II. The paper cites prior ETF and MoE work, but none of the load-bearing claims rest on a self-citation by the authors, and the ETF separation property is an established geometric fact rather than an imported uniqueness theorem. The only substantive concern visible in the text is the K=d capacity issue relative to the cumulative number of classes, which is an injectivity/correctness risk, not a circularity: if the targets are not injective the method may fail or require remapping, but the loss and the claimed improvements are not equivalent to their inputs by construction. Therefore the paper does not exhibit self-definitional, fitted-as-prediction, or citation-smuggled circularity.

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

All reported numbers come from training runs rather than derivation. The central claim depends on: (1) frozen CLIP features being alignable to fixed ETF targets in every domain, (2) K=d providing a unique target column for every class, (3) the AlexNet task identifier recovering the correct task, and (4) expert freezing preserving prior projections. No new entities are postulated. The free parameters are experimental hyperparameters chosen by hand.

free parameters (4)
  • Number of experts NE = 22
    Chosen by hand in Section IV-A; ablation in Table III shows expert count strongly changes accuracy from 58.24 with 2 experts to 84.4 with 11 experts, so this choice materially affects the headline result.
  • Top-k experts per example = 2
    Chosen by hand in Section IV-A; gating selects the top 2 experts and no sensitivity analysis is reported.
  • Training iterations per task = 1,000 (3,000 for Fresh-CL(3k))
    Set in Section IV-A; the headline 2.0% gain in Table I uses 3,000 iterations, while the 1,000-iteration variant gains 1.5%, so compute differs from baselines.
  • ETF width K = d (768 implied by ViT-B/16)
    Set in Section III-A as equal to the feature dimension; this choice determines how many target vectors exist and is not validated against the roughly 1,200 total classes in Table I.
assumptions (4)
  • domain assumption A frozen CLIP ViT-B/16 encoder provides features that can be aligned to fixed ETF targets across all 11 domains.
    Section IV-A freezes the backbone; the method's success depends on CLIP features being suitable for dot-regression to predefined targets in every domain.
  • ad hoc to paper Each class can be assigned to a column of a fixed d-column ETF, and K=d supplies enough columns for all classes encountered.
    Section III-A sets K=d; with 11 datasets totaling more classes than d, the paper neither proves nor states how unique columns are allocated.
  • domain assumption The incrementally trained AlexNet auto-task identifier recovers the correct task at inference.
    Section IV-A follows [33] and trains an AlexNet task identifier; if the pseudo task id is wrong, the wrong router gating is used.
  • domain assumption Freezing the top-k most used experts after each task preserves prior projections and prevents interference.
    Section III-B; no analysis of capacity or interference is provided, and this is load-bearing for the forgetting claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fresh-CL: Feature Realignment through Experts on Hypersphere in Continual Learning." pith.science (2026). https://pith.science/paper/JMKXWVJV

@misc{pith2026250102198,
  author       = {Pith},
  title        = {Pith review of: Fresh-CL: Feature Realignment through Experts on Hypersphere in Continual Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JMKXWVJV}},
  note         = {Machine review of arXiv:2501.02198}
}
read the original abstract

Continual Learning enables models to learn and adapt to new tasks while retaining prior knowledge. Introducing new tasks, however, can naturally lead to feature entanglement across tasks, limiting the model's capability to distinguish between new domain data. In this work, we propose a method called Feature Realignment through Experts on hyperSpHere in Continual Learning (Fresh-CL). By leveraging predefined and fixed simplex equiangular tight frame (ETF) classifiers on a hypersphere, our model improves feature separation both intra and inter tasks. However, the projection to a simplex ETF shifts with new tasks, disrupting structured feature representation of previous tasks and degrading performance. Therefore, we propose a dynamic extension of ETF through mixture of experts, enabling adaptive projections onto diverse subspaces to enhance feature representation. Experiments on 11 datasets demonstrate a 2% improvement in accuracy compared to the strongest baseline, particularly in fine-grained datasets, confirming the efficacy of combining ETF and MoE to improve feature distinction in continual learning scenarios.

Figures

Figures reproduced from arXiv: 2501.02198 by the authors.

Figure 1
Figure 1. t-SNE visualization of Aircraft dataset features. (a) Features extracted [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall framework of Fresh-CL. When training [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 31 canonical work pages

  1. [1]

    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 , pages 139-149, 2022

    work page 2022

  2. [2]

    Dualprompt: Complementary prompting for rehearsal-free continual learning

    Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-YuLee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648, 2022

    work page 2022

  3. [3]

    Generative feature replay for class-incremental learning

    Xialei Liu, Chenshen Wu, Mikel Menta, Luis Herranz, Bogdan Rad- ucanu, Andrew DBagdanov, Shangling Jui, and Joost van de Weijer. Generative feature replay for class-incremental learning. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 226-227, 2020

    work page 2020

  4. [4]

    Memory-efficient incremental learning through feature adaptation

    Ahmet Iscen, Jeffrey Zhang, Svetlana Lazebnik, and Cordelia Schmid. Memory-efficient incremental learning through feature adaptation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVI 16 , pages 699-715, 2020

    work page 2020

  5. [5]

    Self- sustaining representation expansion for non-exemplar class-incremental learning

    Kai Zhu, Wei Zhai, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Self- sustaining representation expansion for non-exemplar class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9296-9305, 2022

    work page 2022

  6. [6]

    An unconstrained layer-peeled perspective on neural collapse

    Wenlong Ji, Yiping Lu, Yiliang Zhang, Zhun Deng, and Weijie Su. An unconstrained layer-peeled perspective on neural collapse. In Interna- tional Conference on Learning Representations , 2021

    work page 2021

  7. [7]

    Neural collapse in deep homogeneous classifiers and the role of weight decay

    Akshay Rangamani and Andrzej Banburski-Fahey. Neural collapse in deep homogeneous classifiers and the role of weight decay. In International Conference on Acoustics, Speech, and Signal Processing . pages 4243-4247, 2022

    work page 2022

  8. [8]

    Prevalence of neural collapse during the terminal phase of deep learning training

    Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceed- ings of the National Academy of Sciences, 117(40):24652-24663, 2020

    work page 2020

Show all 38 references
  1. [9]

    PCA versus LDA

    Aleix M Martinez and Avinash C Kak. PCA versus LDA. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(2):228- 233, 2001

    2001

  2. [10]

    OrCo: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning

    Ahmed N, Kukleva A, Schiele B. OrCo: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28762-28771, 2024

    2024

  3. [11]

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

    Y Yang, H Yuan, X Li, Z Lin, P Torr, and D Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. arXiv preprint arXiv:2302.03004 , 2023

    2023 arXiv

  4. [12]

    Learning equi-angular representations for online continual learn- ing

    Minhyuk Seo, Hyunseo Koh, Wonje Jeung, Minjae Lee, San Kim, Han- kook Lee, Sungjun Cho, Sungik Choi, Hyunwoo Kim, and Jonghyun Choi. Learning equi-angular representations for online continual learn- ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern ...

    2024

  5. [13]

    DER: Dynamically expandable representation for class incremental learning

    Shipeng Yan, Jiangwei Xie, and Xuming He. DER: Dynamically expandable representation for class incremental learning. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3014-3023, 2021

    2021

  6. [14]

    Il2m: Class incremental learning with dual memory

    Eden Belouadah and Adrian Popescu. Il2m: Class incremental learning with dual memory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 583-592, 2019

    2019

  7. [15]

    Class- incremental learning using diffusion model for distillation and replay

    Quentin Jodelet, Xin Liu, Yin Jun Phua, and Tsuyoshi Murata. Class- incremental learning using diffusion model for distillation and replay. In Proceedings of the IEEE/CVF International Conference on Com- puter Vision, pages 3425-3433, 2023

    2023

  8. [16]

    PackNet: Adding multiple tasks to a single network by iterative pruning

    Arun Mallya and Svetlana Lazebnik. PackNet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 7765- 7773, 2018

    2018

  9. [17]

    In defense of the learning without forgetting for task incremental learning

    Guy Oren and Lior Wolf. In defense of the learning without forgetting for task incremental learning. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision , pages 2209-2218, 2021

    2021

  10. [19]

    Adaptive mixtures of local experts

    Robert A Jacbs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural Computation , 3(1):79-87, 1991

    1991

  11. [20]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017

    2017 arXiv

  12. [21]

    Fine-grained visual classification of aircraft

    S Maji, E Rahtu, J Kannal, Blaschko M, Vedaldi A. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 . 2013

    2013 arXiv

  13. [22]

    Automated flower classification over a large number of classes

    M-E Nilsback, A Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing , pages 722-729, 2018

    2008

  14. [23]

    Bossard, M

    L. Bossard, M. Guillaumin, L Van Gool. Food-101 – Mining discrim- inative components with random forests. In European Conference on Computer Vision, pages 446-461, 2014

    2014

  15. [24]

    Vedaldi, A

    OM Parkhi, A. Vedaldi, A. Zisserman, CV Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition , pages 3498-3505, 2012

    2012

  16. [25]

    Krause, M

    J. Krause, M. Stark, J Deng J, L. Fei-Fei. 3D object representations for fine-grained categorization. In 4th IEEE Workshop on 3D Repre- sentation and Recognition, at ICCV 2013 (3DRR-13) , 2013

    2013

  17. [26]

    Fei-Fei, R

    L. Fei-Fei, R. Fergus, P. Perona. Learning generative visual models from few training examples: an incremental Bayesian approach tested on 101 object categories. In Computer Vision and Pattern Recognition Workshop, 2004

    2004

  18. [27]

    Dytox: Transformers for continual learning with dynamic token expansion

    Arthur Douillard, Alexandre Ram ´e, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer vision and Pattern Recognition , pages 9285-9295, 2022

    2022

  19. [28]

    Cimpoi, S

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 3606-3613, 2014

    2014

  20. [29]

    Helber, B

    P. Helber, B. Bischke, A. Dengel, D. Borth. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. arXiv preprint arXiv:1709.00029 . 2017

    2017 arXiv

  21. [30]

    LeCun, C

    Y . LeCun, C. Cortes, C. Burges. MNIST handwritten digit database. ATT Labs[Online]. Available from: http://yann.lecun.com/exdb/mnist. 2010

    2010

  22. [31]

    J. Xiao, J. Hays, KA Ehinger, A. Oliva, A. Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Com- puter Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010

    2010

  23. [32]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 , 2020

    2010 arXiv

  24. [33]

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

    2024

  25. [34]

    Overcoming catastrophic forgetting by incremental moment matching

    Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming catastrophic forgetting by incremental moment matching. In Advances in Neural Information Processing Systems, pages 4655-4665, 2017

    2017

  26. [35]

    Don’t stop learning: Towards continual learning for the CLIP model

    Yuxuan Ding, Lingqiao Liu, Chunna Tian, Jingyuan Yang, and Haox- uan Ding. Don’t stop learning: Towards continual learning for the CLIP model. arXiv preprint arXiv:2207.09248 , 2022

    2022 arXiv

  27. [36]

    Robust fine-tuning of zero-shot models

    Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Si- mon Kornblith, Rebecca Roelofs Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Visio...

    2022

  28. [37]

    Preventing zero-shot transfer degradation in continual learning of vision-language models

    Zangwei Zheng, Mingyuan Ma, Kai Wang, Ziheng Qin, Xiangyu Yue, and Yang You. Preventing zero-shot transfer degradation in continual learning of vision-language models. arXiv preprint arXiv:2303.06628 , 2023

    2023 arXiv

  29. [38]

    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 conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017

    2001

  30. [39]

    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 International Conference on Machine Learning, pa...

    2021

Pith tools

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