Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

RegCL: Continual Adaptation of Segment Anything Model via Model Merging

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read RegCL merges LoRA adapter weights into a single SAM so the model can keep learning new segmentation domains without replaying old data.

desk verdict RegCL is a legitimate but incremental reuse of RegMean for SAM LoRA adapters; the empirical gains look real, but the LoRA representation in the merge formula is underspecified and needs fixing before the parameter-efficiency claims hold. read the letter →

arxiv 2507.12297 v1 pith:6V4WT45H submitted 2025-07-16 cs.LG cs.CV

Yuan-Chen Shu , Zhiwei Lin , Yongtao Wang This is my paper · ORCID
classification cs.LGcs.CV
keywords continuallearningmodelmergingSegmentAnythingLoRAcatastrophicforgettingdomain-incrementalnon-replayimagesegmentation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes RegCL, a continual-learning method that adapts SAM to a stream of image-segmentation domains without replaying old data or growing the model. RegCL trains a fresh LoRA adapter on each new domain and then merges it into the current adapter weights using a closed-form update derived from RegMean, which minimizes the difference between the merged model's predictions and each domain-specific model's predictions. The update is split into a new-knowledge term from the current task and a historical term accumulated from previous tasks, carried by inner-product statistics $C_t = X_t^\top X_t$. Across medical, camouflaged-object, and shadow datasets, the merged model keeps most old-task performance while adapting to new tasks, reaching 91.6% of the mIoU score of a model jointly fine-tuned on all five datasets. The contribution is a parameter-efficient, non-replay continual-learning recipe for foundation-model fine-tuning.

What carries the argument

The central object is the incremental merge update $\bar{W}_t = (P_t + C_t)^{-1}(P_t \bar{W}_{t-1} + C_t W_t)$, together with the inner-product accumulator $P_t = \sum_{i<t} C_i$ and the task statistics $C_t = X_t^\top X_t$. The formula separates the merged weights into a historical term $P_t \bar{W}_{t-1}$ and a new-knowledge term $C_t W_t$, reweighted by $(P_t + C_t)^{-1}$, so that knowledge from all past tasks enters through accumulated feature correlations rather than through stored data. It is applied to the linear LoRA layers; for weights in nonlinear layers, the paper adopts the simpler running average of Eq. (6).

What would settle it

Run the five-domain sequence in a different order, for example COD→ISIC→ISTD→CAMO→Kvasir, and compare final merged weights and average accuracy (ACC); RegCL's order-independence claim predicts the same final model, so any material change in ACC under reordering would refute the central merging mechanism.

Watch

Extended reading notes

Core claim

RegCL's central claim is that RegMean's closed-form model-merging solution can be converted into an online continual-learning rule for SAM's LoRA adapters: after training a task-specific adapter $W_t$, the merged weights update as $\bar{W}_t = (P_t + C_t)^{-1}(P_t \bar{W}_{t-1} + C_t W_t)$, where $C_t$ is the inner-product matrix of that task's encoder features and $P_t$ accumulates all previous $C_i$. This rule follows from minimizing the squared prediction discrepancy between the merged model and each individually trained domain model, and it keeps the final model the same size as a single LoRA adapter while storing no historical images. The paper demonstrates on Kvasir, CAMO, ISTD, ISIC, and COD that this schedule suppresses catastrophic forgetting, achieving backward transfer of $-0.028$ mIoU and average accuracy of $0.751$ mIoU, and that it outperforms sequential LoRA tuning and existing continual-learning baselines on the same architecture and training recipe.

Load-bearing premise

The load-bearing premise is that fusing LoRA weights with the linear-model formula and averaging the nonlinear layers keeps the full SAM network predicting almost like every domain-specific model, an exact guarantee that is proven only for linear models.

Editorial extensions

If this is right

  • A single SAM with one LoRA adapter can serve a stream of segmentation domains, with old-task performance dropping only a few points (backward transfer $-0.028$ mIoU) instead of collapsing as in sequential LoRA fine-tuning.
  • Deployment cost stays constant as domains accumulate, and no historical images or labels need to be stored, which matters when storage or data-privacy constraints forbid replay.
  • Because the merge depends only on the commutative sum of inner-product matrices, the final model should be independent of the order in which tasks arrive.
  • RegCL composes with rehearsal: adding 300 replay samples per dataset raises average accuracy by $0.058$ mIoU, so the merging rule can be a component of memory-based continual learning rather than a replacement for it.

Reading between the lines

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

  • Beyond the paper's claims, the same new-knowledge/historical decomposition should transfer to other parameter-efficient adapters whose weight update is linear in the features, not just LoRA, since the RegMean solve itself only assumes linearity of the layer.
  • Beyond the paper's claims, the inner-product matrices $C_t$ are computed from encoder features and are shared instead of raw data; whether those matrices leak private information about the training images is not analyzed, so privacy guarantees beyond 'no raw data stored' remain open.
  • Beyond the paper's claims, the memory overhead of storing full $C_t$ matrices grows with the square of the layer input dimension, so for very large encoders a low-rank or diagonal approximation of the accumulator is a natural stress test of the method's scalability.
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

5 major / 6 minor

Summary. The paper proposes RegCL, a non-replay continual learning method for adapting SAM to a sequence of domains using LoRA adapters. For each task, a task-specific LoRA adapter is trained while SAM is frozen, and the method computes per-layer input inner-product matrices C_t. The adapter weights are then merged incrementally with the accumulated merged weights via Eq. (4) for linear layers and via a simple average, Eq. (6), for nonlinear layers. The authors claim that this process retains previous-domain knowledge without storing historical data, keeps model size constant, and is order-independent. Experiments on five segmentation datasets (Kvasir, CAMO, ISTD, ISIC, COD) report that RegCL outperforms sequential LoRA and several continual learning baselines on average accuracy, backward transfer, and forward transfer, and that adding a small replay buffer further improves performance.

Significance. Continual adaptation of segmentation foundation models is a timely and practical problem, and the idea of reusing RegMean's closed-form merging in an incremental, replay-free setting is simple and potentially useful. The reported gains over sequential LoRA (about 5.5 points mIoU average accuracy) are non-trivial, and the paper identifies a real gap in current adapter-based SAM adaptation. However, the method is largely an application of an existing merging algorithm to a new setting, and the central correctness and evaluation gaps described below make the current claims stronger than the evidence provided.

major comments (5)
  1. [§3.2.2, Eq. (4)] The merging operation is underspecified for the LoRA parameterization. LoRA restricts the incremental weight to ΔW = BA with rank r, but Eq. (4) is the closed-form minimizer of the regression objective for an unrestricted weight matrix W, as in Eqs. (1)–(2). The manuscript never states whether W_t in Eq. (4) denotes the LoRA factors (A_t, B_t) or the dense delta matrix ΔW_t. If it denotes the factors, then (P_t + C_t)^{-1}(P_t W̄_{t-1} + C_t W_t) is not a well-defined operation on LoRA parameters and does not minimize the stated objective. If it denotes the dense delta, the merged matrix is a weighted sum of rank-r matrices and is generically full rank after two tasks; representing it as a single LoRA adapter of rank r requires a low-rank projection for which no formula or error analysis is provided. This ambiguity also undermines the 'Low Memory Overhead' property in §3.2.4, since storing a dense d_out × d_in delta per layer is not the same as storing LoRA factors. The authors should specify the exact representation of W_t and either derive a rank-constrained solution or provide an explicit projection step with error analysis.
  2. [§3.2.2, Eq. (6)] The simple average for 'nonlinear layers' is introduced without derivation and without identifying which weights in the LoRA-based fine-tuning setup are nonlinear. Since Eq. (4) is applied to linear LoRA layers and Eq. (6) to the remaining weights, the final merged model is not the minimizer of a single well-defined prediction-discrepancy objective over the whole network. The central claim in the Introduction and in §3.2.2 that RegCL 'minimizes prediction discrepancies between the merged model and each of the domain-specific models' is therefore only heuristic for the nonlinear parts. The authors should either prove an approximation guarantee for Eq. (6), restrict the claim to the linear layers, or empirically test the sensitivity of the final performance to alternative nonlinear merging strategies.
  3. [§3.2.2, Eqs. (4)–(5)] The iterative update does not solve the global RegMean objective exactly. At step t, Eq. (4) is the exact minimizer of the discrepancy between the new merged model and the pair (W̄_{t-1}, W_t) weighted by (P_t, C_t), but W̄_{t-1} is itself a previously merged approximation of the historical task-specific models W_1, ..., W_{t-1}. Thus the objective minimized at step t is not the same as the original objective of matching each individual W_i on its own data, and no bound on the accumulated approximation error is provided. The claim of closed-form optimality inherited from RegMean should be stated only for the two-model step, or else supported by an error analysis that controls the drift introduced by repeated merging.
  4. [§3.2.4 and §4.3] The 'Order Independent' property is asserted but not demonstrated. The commutativity argument holds only if each task-specific W_t is trained independently from the same initialization W0 and the C_t are fixed; these assumptions are not stated in the property and are not verified. Section 4 reports a single task order (Kvasir→CAMO→ISTD→ISIC→COD), so the order-independence claim has no empirical support. Moreover, the replay variant in §4.5 fine-tunes the merged weights W̄_t with replay samples and D_t, which breaks the independence assumption and invalidates the order-independence argument for that variant. The authors should either test multiple task orders and state the assumptions explicitly, or remove the order-independence claim.
  5. [§4.3, Tables 1–3] The main empirical claim rests on single-run results with no error bars, no reported number of seeds, and no code release. In addition, the SPPA baseline attains an average mIoU of 0.282, far below the zero-shot SAM baseline (0.647 in Table 2), which suggests that the baseline is not properly configured for this setting; if so, the comparison does not support the claim that RegCL 'surpasses all other continual learning models.' The authors should provide multiple seeds with standard deviations, verify that all baselines are reasonably tuned, and ideally include results with alternative task orders to support the order-independence claim.
minor comments (6)
  1. [Abstract and §1] There are minor grammar issues, e.g., 'these methods are specific developed for specific domains' and 'If used on other domains may lead to performance degradation'; these should be corrected.
  2. [§3.2.2] The notation for the inner product accumulator is malformed: 'P_t = P_{t-1} i=1 C_i' should be written as P_t = Σ_{i=1}^{t-1} C_i.
  3. [§3.2.2, Eq. (6)] Equation (6) is missing parentheses; it should read W̄_t = (1/t)((t-1) W̄_{t-1} + W_t).
  4. [Table 3] There is a typo in the table caption: 'RegCL+Repaly' should be 'RegCL+Replay'.
  5. [§4.6] The visualization section refers to 'Figure 3', but the corresponding figure with RGB images, ground truth, SAM predictions, and RegCL predictions appears to be Figure 4; the cross-reference should be corrected.
  6. [§2.2] The description of RegMean says 'the output of the merged model should be as close as possible to the output of the merged model'; this should be 'the output of each individual model'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RegCL applies the external RegMean closed-form solution sequentially; its predictions are independent of the method's fit.

full rationale

The paper's derivation chain is a direct reuse of the externally published RegMean closed-form solution, applied incrementally in a continual-learning setting. Equation (4) is obtained by treating the current task weights and the accumulated historical weights as the two models in RegMean's two-model solution, and the historical accumulator P_t is defined as the sum of inner-product matrices. This is algebra, not self-reference: the merged weights minimize a prediction-discrepancy objective on training inputs, and evaluation is then performed on held-out test data. No parameter is fitted to the reported metrics and then reported as a prediction, and no load-bearing claim is justified only by a self-citation. The paper's use of RegMean, LoRA, and SAM is based on external prior work, and those citations provide independent support rather than circular grounding. The skeptical concern that Eq. (4) is not fully specified for LoRA's factorized or low-rank parameters is a correctness or implementation gap, not a circularity, because it does not make the derived result equivalent to its own inputs. Overall, the central result is not forced by definition or by the paper's own assumptions; it is an adaptation of an existing method to a new setting.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method inherits RegMean's assumptions and adds an unproven layer-wise approximation for deep nonlinear models. No new entities are introduced.

free parameters (3)
  • LoRA rank
    Not reported in the paper; it controls the size of the merged LoRA modules and of the inner product matrices C_t.
  • Loss weight 10 on dice loss = 10
    Chosen by hand in Eq. (7); part of the fine-tuning loss for each task.
  • Replay samples per task (Section 4.5) = 300
    Arbitrary choice for the replay extension; not part of the core method.
assumptions (4)
  • domain assumption RegMean closed-form merge (Eq. 2) is optimal for linear layers with squared loss.
    Adopted from prior work [13] as the foundation of the merge objective.
  • ad hoc to paper Layer-wise merging with Gram matrices approximates the full-network objective.
    The paper applies RegMean to individual LoRA layers without proving that layer-wise solutions optimize the whole model's prediction discrepancy.
  • domain assumption Task-specific LoRA models trained from the same random initialization are compatible for merging.
    Required for the merged weights to produce meaningful segmentations; stated implicitly in the setup.
  • ad hoc to paper For nonlinear layers, simple averaging preserves performance.
    Introduced in Eq. (6) without derivation; no experiment isolates its effect.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RegCL: Continual Adaptation of Segment Anything Model via Model Merging." pith.science (2026). https://pith.science/paper/6V4WT45H

@misc{pith2026250712297,
  author       = {Pith},
  title        = {Pith review of: RegCL: Continual Adaptation of Segment Anything Model via Model Merging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6V4WT45H}},
  note         = {Machine review of arXiv:2507.12297}
}
read the original abstract

To address the performance limitations of the Segment Anything Model (SAM) in specific domains, existing works primarily adopt adapter-based one-step adaptation paradigms. However, some of these methods are specific developed for specific domains. If used on other domains may lead to performance degradation. This issue of catastrophic forgetting severely limits the model's scalability. To address this issue, this paper proposes RegCL, a novel non-replay continual learning (CL) framework designed for efficient multi-domain knowledge integration through model merging. Specifically, RegCL incorporates the model merging algorithm into the continual learning paradigm by merging the parameters of SAM's adaptation modules (e.g., LoRA modules) trained on different domains. The merging process is guided by weight optimization, which minimizes prediction discrepancies between the merged model and each of the domain-specific models. RegCL effectively consolidates multi-domain knowledge while maintaining parameter efficiency, i.e., the model size remains constant regardless of the number of tasks, and no historical data storage is required. Experimental results demonstrate that RegCL achieves favorable continual learning performance across multiple downstream datasets, validating its effectiveness in dynamic scenarios.

Figures

Figures reproduced from arXiv: 2507.12297 by the authors.

Figure 1
Figure 1. Illustration of RegCL for continual learning. RegCL merges weights from independent fine-tuned models in a continual learning setting. known for its exceptional zero-shot segmentation capabil￾ities across various natural image datasets. It can pro￾duce diverse and detailed segmentation masks based on user prompts, e.g., points and bounding boxes. Despite its strong performance with natural images, recent stud￾ies re… view at source ↗
Figure 2
Figure 2. The overall pipeline of RegCL. After SAM is fine-tuned on a new task with LoRA modules. RegCL then computes the feature inner product Ci and updates an inner product accumulator Pi used to merge current model weights Wt with previous weights Wt−1. The merged weights Wt are incrementally updated across tasks, enabling knowledge retention while adapting to new tasks. For merging details, please refer to Eq. (4)– (6). … view at source ↗
Figure 3
Figure 3. Performance for each dataset during the continual learning process. From left to right, we report the accuracy of Independent fine-tuning, Sequential fine-tuning, and the proposed RegCL. The results for already learned tasks are presented in red, and the performance of unlearned tasks is denoted as blue. The matrices illustrate task-specific accuracy (%) for evaluated tasks (columns) after training on subsequent tas… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization results of the segmentation mask. From left to right, we show the results on medical segmentation (Kvasir-SEG and ISIC), camouflaged objects segmentation (CAMO and COD-10K), and shadow object segmentation (ISTD). For each row, we show the input RGB images…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. DA-MergeLoRA: Hypernetwork-Based LoRA Merging for Few-Shot Test-Time Domain Adaptation

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A hypernetwork generates per-column merging weights to combine source LoRA modules on CLIP, achieving state-of-the-art few-shot test-time domain adaptation.

Reference graph

Works this paper leans on

35 extracted references · 26 canonical work pages · cited by 1 Pith paper

  1. [1]

    On the opportunities and risks of foundation models

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Alt- man, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1

    arXiv 2021

  2. [2]

    Dark experience for general continual learning: a strong, simple baseline.NeurIPS, 2020

    Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline.NeurIPS, 2020. 8

    work page 2020

  3. [3]

    Adaptformer: Adapting vision transformers for scalable visual recognition

    Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 3

    work page 2022

  4. [4]

    Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the interna- tional skin imaging collaboration (isic)

    Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the interna- tional skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 5

    arXiv 2018

  5. [5]

    A continual learning survey: Defying for- getting in classification tasks

    Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ale ˇs Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying for- getting in classification tasks. IEEE TPAMI, 2021. 2

    work page 2021

  6. [6]

    Continual learning with tiny episodic memories

    P Dokania, P Torr, and M Ranzato. Continual learning with tiny episodic memories. InWorkshop on Multi-Task and Life- long Reinforcement Learning, 2019. 8

    work page 2019

  7. [7]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 3

    work page 2021

  8. [8]

    Concealed object detection

    Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE TPAMI, 2021. 5

    work page 2021

Show all 35 references
  1. [9]

    Parameter-efficient transfer learning for nlp

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, 2019. 3

    2019

  2. [10]

    Re-evaluating continual learning scenarios: A categorization and case for strong baselines

    Yen-Chang Hsu, Yen-Cheng Liu, Anita Ramasamy, and Zsolt Kira. Re-evaluating continual learning scenarios: A categorization and case for strong baselines. arXiv preprint arXiv:1810.12488, 2018. 4

    2018 arXiv

  3. [11]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR,

  4. [12]

    Kvasir-seg: A segmented polyp dataset

    Debesh Jha, Pia H Smedsrud, Michael A Riegler, P ˚al Halvorsen, Thomas De Lange, Dag Johansen, and H˚avard D Johansen. Kvasir-seg: A segmented polyp dataset. In In- ternational conference on multimedia modeling, pages 451–

  5. [13]

    Dataless knowledge fusion by merging weights of language models

    Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. arXiv preprint arXiv:2212.09849 , 2022. 2, 3

    2022 arXiv

  6. [14]

    Lifelong vision models with memory-constrained rehearsal

    Kenneth Joseph and Alexa Smith. Lifelong vision models with memory-constrained rehearsal. CVPR, 2022. 2

    2022

  7. [15]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. In ICCV, 2023. 1, 5

    2023

  8. [16]

    Overcoming catastrophic forgetting in neu- ral 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 neu- ral networks. Proceedings of the national academy of sc...

    2017

  9. [17]

    Anabranch network for camouflaged object segmentation

    Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, Minh- Triet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer vision and im- age understanding, 2019. 5

    2019

  10. [18]

    Federated learning: Challenges, methods, and future directions

    Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 2020. 3

    2020

  11. [19]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV,

  12. [20]

    Continual semantic segmentation via structure preserving and projected feature alignment

    Zihan Lin, Zilei Wang, and Yixin Zhang. Continual semantic segmentation via structure preserving and projected feature alignment. In ECCV, 2022. 6

    2022

  13. [21]

    Gradient episodic memory for continual learning

    David Lopez-Paz and Marc’Aurelio Ranzato. Gradient episodic memory for continual learning. NeurIPS, 2017. 6

    2017

  14. [22]

    Merging models with fisher-weighted averaging

    Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. NeurIPS, 2022. 3, 5

    2022

  15. [23]

    Incremental learn- ing techniques for semantic segmentation

    Umberto Michieli and Pietro Zanuttigh. Incremental learn- ing techniques for semantic segmentation. In ICCV, 2019. 2

    2019

  16. [24]

    V-net: Fully convolutional neural networks for volumetric medical image segmentation

    Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 Fourth International Conference on 3D Vision (3DV), 2016. 5

    2016

  17. [25]

    icarl: Incremental classifier and representation learning

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 2

    2017

  18. [26]

    Generative continual concept learning

    Mohammad Rostami, Soheil Kolouri, Praveen Pilly, and James McClelland. Generative continual concept learning. In AAAI, 2020. 2

    2020

  19. [27]

    Progressive neural networks

    Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Raz- van Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2

    2016 arXiv

  20. [28]

    Three scenar- ios for continual learning

    Gido M Van de Ven and Andreas S Tolias. Three scenar- ios for continual learning. arXiv preprint arXiv:1904.07734,

    1904 arXiv

  21. [29]

    Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal

    Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In CVPR, 2018. 5

    2018

  22. [30]

    Or- thogonal subspace learning for language model continual learning

    Xiao Wang, Tianze Chen, Qiming Ge, Han Xia, Rong Bao, Rui Zheng, Qi Zhang, Tao Gui, and Xuanjing Huang. Or- thogonal subspace learning for language model continual learning. arXiv preprint arXiv:2310.14152, 2023. 6

    2023 arXiv

  23. [31]

    Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing in- ference time

    Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Re- becca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Ko- rnblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing ...

    2022

  24. [32]

    Medical sam adapter: Adapting segment anything model for medical im- age segmentation

    Junde Wu, Rao Fu, Huihui Fang, Yuanpei Liu, Zhao-Yang Wang, Yanwu Xu, Yueming Jin, and Tal Arbel. Medical sam adapter: Adapting segment anything model for medical im- age segmentation. Medical image analysis, 2023. 1

    2023

  25. [33]

    Ties-merging: Resolving interference when merging models

    Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raf- fel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. NeurIPS, 2023. 3

    2023

  26. [34]

    Language models are super mario: Absorbing abilities from homologous models as a free lunch

    Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In Forty-first Interna- tional Conference on Machine Learning, 2024. 3

    2024

  27. [35]

    Learning at a glance: Towards interpretable data-limited continual seman- tic segmentation via semantic-invariance modelling

    Bo Yuan, Danpei Zhao, and Zhenwei Shi. Learning at a glance: Towards interpretable data-limited continual seman- tic segmentation via semantic-invariance modelling. IEEE TPAMI, 2024. 6

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.