Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Boosting Domain Incremental Learning: Selecting the Optimal Parameters is All You Need

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read SOYO claims that in parameter-isolation continual learning, the bottleneck is domain guessing, and a compact trainable selector with Gaussian-mixture memory consistently beats training-free selectors across six benchmarks.

desk verdict A useful, broadly evaluated domain-selection plug-in for PIDIL whose accuracy gains look real, but the memory numbers in Table 5 contradict the full-covariance GMM in the method section and need fixing. read the letter →

arxiv 2505.23744 v1 pith:AL4VNQDF submitted 2025-05-29 cs.CV cs.AI

Qiang Wang , Xiang Song , Yuhang He , Jizhou Han , Chenhao Ding , Xinyuan Gao , Yihong Gong This is my paper
classification cs.CVcs.AI
keywords domainincrementallearningparameter-isolationlabelpredictionGaussianmixturecompressionfeatureresamplingparameter-efficientfine-tuningcontinualmulti-levelfusion
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 is trying to establish that the main weakness of parameter-isolation domain incremental learning is not how each domain is fine-tuned but how the model guesses which domain a test sample comes from. It proposes SOYO, a trainable domain selector, and shows that plugging it into existing methods improves accuracy. SOYO compresses previous-domain features into Gaussian mixtures, resamples synthetic features to balance training, and trains a small network on fused shallow and deep features. Experiments across six benchmarks in image classification, object detection, and speech enhancement consistently improve over training-free selectors. If correct, the paper's message is that selecting the optimal parameters is all you need: better domain selection converts directly into better continual-learning performance.

What carries the argument

The load-bearing mechanism is the trio of GMC, DFR, and MDFN. The Gaussian Mixture Compressor fits a K-component mixture model $p(x|\theta)=\sum_{k=1}^K \lambda_k \mathcal{N}(x|\mu_k,\Sigma_k)$ to the features of each previous domain at both the $L/2$-th and $L$-th transformer layers, storing only the mixture parameters. The Domain Feature Resampler draws $N_t$ synthetic samples from those mixtures, matched to the current domain's sample count, to rebalance the selector's training set without storing raw features. The Multi-level Domain Feature Fusion Network computes the fused feature $x^D = x^L + g_1(x^{L/2}) + g_2(x^L)$ and maps it through a linear head to domain probabilities, trained with cross-entropy. Together they convert domain selection from a training-free geometric match into a balanced, trainable classifier over a fixed-size memory of each past domain.

What would settle it

Take one image classification benchmark, such as DomainNet, and give SOYO the same memory budget but replace GMC and DFR with a randomly sampled real exemplar set of matched size from each previous domain. If SOYO's selection accuracy is not at least as high as this real-exemplar replay baseline, then the Gaussian-mixture approximation is not what carries the claimed gain.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that parameter-isolation domain incremental learning can be boosted more by improving the inference-time domain selector than by further refining the per-domain fine-tuning. SOYO treats domain selection as a trainable classification problem: for each previous domain it fits a Gaussian mixture model to backbone features at two depths, stores only mixture parameters, resamples pseudo-features so the selector's training set is balanced, and trains a Multi-level Domain Feature Fusion Network that combines intermediate spatial features with final semantic features to predict the domain label. The experiments show consistent gains in parameter selection accuracy and in downstream metrics, bringing LDB within 0.3 mAP of the oracle on Pascal VOC and raising LNA's SI-SNR from 16.07 to 17.50 dB. The paper concludes that a generic, compact selector of this kind is compatible with multiple parameter-efficient fine-tuning methods and transfers across image, detection, and audio tasks.

Load-bearing premise

The Gaussian-mixture approximation of each previous domain's features must preserve enough of the domain-discriminative structure that pseudo-features resampled from it train a domain classifier that still works on real test features.

Editorial extensions

If this is right

  • Existing parameter-isolation methods, such as prompt- and adapter-based ones, can get an immediate accuracy boost by swapping their KNN or nearest-mean selector for SOYO, without retraining the domain-specific parameters.
  • The gap between a PIDIL method and its oracle, which knows the true domain label, narrows: on the Pascal VOC series, mAP rises from 56.8 to 59.6, within 0.3 of the oracle.
  • Parameter selection accuracy becomes a measurable bottleneck, since gains in selection accuracy track gains in downstream metrics on all three tasks.
  • The approach scales to longer domain sequences because the stored representation per domain is a fixed-size Gaussian mixture rather than a growing set of exemplar features.

Reading between the lines

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

  • Because GMC stores only mixture parameters, SOYO is a form of generative replay in feature space; the same compression could be plugged into rehearsal-based continual learning to reduce buffer size while keeping a balanced training signal.
  • The paper fixes the number of Gaussian components per task via the Bayesian Information Criterion; an adaptive or online choice of K could further improve fidelity when domains have heterogeneous internal structure.
  • The selector is trained after each new domain arrives; training it jointly with the domain-specific fine-tuning, or updating the mixture with streaming EM, could cut the reported 9% training overhead.
  • Since the selector is task-agnostic, it may transfer to other parameter-isolation settings, such as selecting among LoRA adapters or test-time adapters for unseen domains, though the paper only demonstrates seen-domain selection plus cross-domain classification on CORe50.
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 / 6 minor

Summary. The paper proposes SOYO, a lightweight trainable domain selector for parameter-isolation domain incremental learning (PIDIL). SOYO consists of a Gaussian Mixture Compressor (GMC) that compresses features from previous domains into Gaussian mixtures, a Domain Feature Resampler (DFR) that resamples pseudo-features to balance training, and a Multi-level Domain Feature Fusion Network (MDFN) that fuses features from intermediate and final transformer layers for domain classification. The method is evaluated on six benchmarks spanning image classification (DomainNet, CDDB, CORe50), object detection (Pascal VOC, BDD100K), and speech enhancement (WSJ0), covering multiple PEFT backbones. Results show consistent improvements in parameter-selection accuracy and downstream task metrics over baselines such as S-Prompts, PINA, LDB, and LNA.

Significance. If the reported gains are reproducible, SOYO offers a generic plug-in that can improve the accuracy of existing PIDIL methods across modalities with only a modest amount of extra computation, which would be practically useful. The paper's breadth of evaluation across three tasks and six benchmarks is a notable strength, as is its compatibility with multiple PEFT methods. The methodological contribution is simple but intuitive: train a lightweight domain classifier with resampled GMM features rather than relying on training-free nearest-centroid schemes. However, the significance is tempered by the absence of statistical error bars, an apparently inconsistent memory-usage claim, and an underspecified feature-extraction protocol at inference time.

major comments (4)
  1. [Sec. 3.4, Eq. (3)-(6); Table 5] The reported memory usage for GMC appears inconsistent with the full-covariance Gaussian model. Equation (3) explicitly defines a d-dimensional Gaussian with a full covariance matrix Σ, and Eq. (6) applies GMC to both the L/2-th and L-th layers for every previous domain. On DomainNet with ViT-B/16 (d=768, K=2), at the final session there are 5 previous domains, so storing the parameters requires 5 domains × 2 layers × 2 components × (768^2 covariance entries + 768 mean entries + 1 weight) ≈ 11.8M scalars, which is roughly 13.7% of the 86M ViT-B/16 parameters. Table 5 reports 1.293% for 'SOYO (GMC, K=2)', nearly a factor of ten lower. If the implementation actually uses diagonal covariance matrices, then the description of Eq. (3) is misleading and the comparison of GMC against Mean&std—which would be a close competitor—needs reinterpretation. If full covariances are truly stored, the paper's claims of 'lightweight' and 'without increasing memory usage' (Abstract, Sec. 3.4) are off by an order of magnitude. Either way, the authors must clarify the exact covariance structure, recompute the memory figures, and reconcile the text with the reported numbers.
  2. [Sec. 3.3, Fig. 2] The feature extraction procedure for domain prediction is underspecified, and this is load-bearing for the method's correctness. At inference, SOYO must predict the domain label of a test image before the corresponding parameters are selected. The features x^{L/2} and x^L used by the MDFN are outputs of transformer blocks that, during training, contain domain-specific parameters (e.g., adapters). If at inference these features are extracted from the pre-trained backbone without any domain-specific parameters, there is a train/test feature-distribution mismatch with the GMC-resampled features, which were extracted from the model with adapters. If, instead, the features depend on which adapter is applied, then the domain prediction becomes circular: selecting the optimal parameters requires knowing the domain, but the domain prediction is based on features that may already be adapted. The paper does not state which backbone state is used for feature extraction at test time. Please specify this explicitly and, if necessary, retrain/evaluate the MDFN under the same feature-extraction protocol used at inference.
  3. [Sec. 4.3, Fig. 3 and Table 5] The hyperparameter K (the number of Gaussian components) is selected using test-set metrics. Figure 3 reports the mAP on the Pascal VOC series dataset for different K, and Table 5 reports DomainNet accuracy for different K. Using these test metrics to choose K constitutes a form of test-set leakage and inflates the reported performance relative to the baselines, whose hyperparameters are presumably fixed. The paper should either select K on a held-out validation split or demonstrate that the results are not sensitive to K (e.g., a table of AT/ST for K = 1,...,5 on a validation subset). Without this, the comparison in Tables 1 and 2 is not entirely fair.
  4. [Tables 1-3] All reported results are single runs with no error bars or multiple seeds. The improvements over baselines are modest in several cases (e.g., +3.02 AT on DomainNet, +2.15 AT on CDDB in Table 1; +2.8 mAP on the Pascal VOC series in Table 2). These differences could plausibly fall within run-to-run variance. To support the claim of 'consistent superiority', the authors should report mean and standard deviation over at least three independent runs (or seeds) for the main benchmarks. This is particularly important because SOYO involves random resampling in DFR and stochastic training of the MDFN.
minor comments (6)
  1. [Abstract and Sec. 3.4] The phrase 'without increasing memory usage' is contradicted by the paper's own Table 5, which reports a non-zero memory overhead for every SOYO variant. Please rephrase to something like 'with small memory overhead' to avoid overclaiming.
  2. [Sec. 3.1] The notation 'theT -th domain' appears as 'theT -th' in the problem formulation; fix this typo.
  3. [Sec. 4.1, Implementation Details] The values of K (2, 3, and 1 for DIC, DIOD, and DISE) are given only in prose. A small table summarizing the dataset, backbone, K, and feature dimensions would improve readability.
  4. [Table 4] In the LNA+K&K confusion matrix, the row labels 'DD' and 'MG' are missing for Session 4*; consider presenting the session-wise selection accuracy more compactly with a single table of ST percentages.
  5. [Fig. 1] The radar chart in Fig. 1(b) has unreadably small labels and unclear axis scaling (e.g., 'lower bound of the radar chart is set to 90% of pretrain'). Please enlarge fonts and describe the normalization in the caption or text.
  6. [References] Some references are missing venue details (e.g., [24] and [25] are preprint/arXiv entries; the text uses 'Method [47]' in Related Work without a name). Ensure all entries are complete and consistent with the bibliography style.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: SOYO's domain-selection accuracy is measured on held-out test labels, and the GMM/DFR training pipeline does not encode those labels.

full rationale

SOYO's derivation chain is: GMC fits a K-component Gaussian mixture to the l-th-layer features of each previous domain; DFR resamples pseudo-features from those fitted Gaussians; MDFN is trained with cross-entropy on pseudo-features plus current-domain features using known domain indices as labels; at inference the MDFN predicts a domain label for test features and the corresponding PEFT parameters are selected. The reported metrics (AT, ST, mAP, SI-SNR, etc.) are computed on held-out test samples whose labels are not used in fitting the GMM, resampling pseudo-features, or training MDFN. The GMM is a lossy compressor of the training feature distribution, not a function of the test outcome, so the domain prediction is not forced by construction. The ablation in Table 5 compares GMC with Mean&std and PCA under the same MDFN and the same final accuracy criterion, so the GMC advantage is an empirical result rather than a fitted artifact. Same-group baselines (PINA, LDB, LNA) are used, but SOYO is also evaluated on the external S-Prompts and MoP-CLIP methods, and no load-bearing argument rests on a self-citation. The Fig. 3 choice of K=3 by Pascal VOC mAP is a test-set hyperparameter-selection leak and a reproducibility concern, not a by-construction equivalence between a fitted input and the reported prediction. Likewise, the Table 5 memory percentage appears inconsistent with the full-covariance GMM in Eqs. (3)-(6), but that is an internal-consistency/cost-accounting issue rather than circular reasoning. Under the hard rules, no specific circular step can be quoted and exhibited, so the appropriate finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

The main free parameter is K, the number of Gaussian components, chosen per task and partly via test-set metrics. The key domain assumption is that features from each domain are well approximated by a small K-component Gaussian mixture. No new physical entities are introduced.

free parameters (1)
  • GMC number of Gaussian components K = K=2 (DIC), 3 (DIOD), 1 (DISE)
    Chosen per task and partly validated using BIC and final test metrics (Fig. 3, Tab. 5). K controls the fidelity of the feature approximation and thus the quality of resampled pseudo-features.
assumptions (2)
  • standard math The EM algorithm for Gaussian mixture models converges to a useful approximation of the feature distribution.
    Invoked in Sec. 3.4 to fit GMC parameters; standard statistical tool.
  • domain assumption Features of each previous domain are well approximated by a small K-component Gaussian mixture.
    Sec. 3.4, Eq. (4)-(6): the whole resampling procedure relies on this approximation preserving domain-discriminative structure. There is no verification of how well the GMM matches the true feature distribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Boosting Domain Incremental Learning: Selecting the Optimal Parameters is All You Need." pith.science (2026). https://pith.science/paper/AL4VNQDF

@misc{pith2026250523744,
  author       = {Pith},
  title        = {Pith review of: Boosting Domain Incremental Learning: Selecting the Optimal Parameters is All You Need},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AL4VNQDF}},
  note         = {Machine review of arXiv:2505.23744}
}
read the original abstract

Deep neural networks (DNNs) often underperform in real-world, dynamic settings where data distributions change over time. Domain Incremental Learning (DIL) offers a solution by enabling continual model adaptation, with Parameter-Isolation DIL (PIDIL) emerging as a promising paradigm to reduce knowledge conflicts. However, existing PIDIL methods struggle with parameter selection accuracy, especially as the number of domains and corresponding classes grows. To address this, we propose SOYO, a lightweight framework that improves domain selection in PIDIL. SOYO introduces a Gaussian Mixture Compressor (GMC) and Domain Feature Resampler (DFR) to store and balance prior domain data efficiently, while a Multi-level Domain Feature Fusion Network (MDFN) enhances domain feature extraction. Our framework supports multiple Parameter-Efficient Fine-Tuning (PEFT) methods and is validated across tasks such as image classification, object detection, and speech enhancement. Experimental results on six benchmarks demonstrate SOYO's consistent superiority over existing baselines, showcasing its robustness and adaptability in complex, evolving environments. The codes will be released in https://github.com/qwangcv/SOYO.

Figures

Figures reproduced from arXiv: 2505.23744 by the authors.

Figure 1
Figure 1. Illustration of the proposed SOYO. (a) The green, blue, and red small squares represent learnable parameters for dif￾ferent domains. (b) The “pretrain” represents evaluation using the pre-trained model directly. The lower bound of the radar chart is set to 90% of the “pretrain”, while the upper bound represents the achievable upper limit of the baseline if the domain labels of test samples are known in advance. Best… view at source ↗
Figure 2
Figure 2. Illustration of the proposed framework. The numbers 1, 2, 3, and 4 in (b) indicate the sequence of steps. Best viewed in color. versity. The MDFN is a lightweight trainable network that fuses features from shallow and deep layers to produce more discriminative domain features and obtain the domain prediction result. In the inference phase, the SOYO predicts the domain label of the input image and integrates the corr… view at source ↗
Figure 3
Figure 3. BIC score and mAP on the Pascal VOC series dataset. In the right table, K = 0 represents the baseline without GMC. Method Memory #Params AT ∆AT ST ∆ST KMeans&KNN 0.004% 0 61.58 -0.65 80.35 -0.20 Nearest Mean Classifier 0.001% 61.16 -1.07 79.58 -0.97 Patch Shuffle Selector 0.001% 62.23 0 80.55 0 SOYO (Mean&std) 0.002% 0.06% 63.48 +1.25 82.83 +2.28 SOYO (PCA, N = 10) 0.009% 63.74 +1.51 83.15 +2.60 SOYO (PCA, N = 100) … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Model training convergence curves on DomainNet. The [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: t-SNE visualization of image features. Features are extracted by randomly selecting one image from each domain and class (345 classes across 6 domains) in the DomainNet dataset. (a) visualized the features extracted from the pre-trained ViT-B, and (b) shows the feature…

Discussion (0). Sign in 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. Continual Knowledge Consolidation LORA for Domain Incremental Learning

    cs.LG 2025-10 conditional novelty 5.0 of 10

    CONEC-LoRA reports state-of-the-art accuracy on four domain-incremental benchmarks by combining task-shared and task-specific LoRAs with a stochastic classifier and a learned domain-ID selector.

Reference graph

Works this paper leans on

61 extracted references · 44 canonical work pages · cited by 1 Pith paper

  1. [1]

    Subspace Regularizers for Few-Shot Class Incremental Learning

    Afra Feyza Aky ¨urek, Ekin Aky¨urek, Derry Tanti Wijaya, and Jacob Andreas. Subspace regularizers for few-shot class in- cremental learning. arXiv preprint arXiv:2110.07059, 2021. 2

    work page Pith review arXiv 2021

  2. [2]

    Memory aware synapses: Learning what (not) to forget

    Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision (ECCV), pages 139–154, 2018. 2

    work page 2018

  3. [3]

    Leaving none behind: Data-free domain incremental learn- ing for major depressive disorder detection

    Tao Chen, Yanrong Guo, Shijie Hao, and Richang Hong. Leaving none behind: Data-free domain incremental learn- ing for major depressive disorder detection. IEEE Transac- tions on Affective Computing, 2024. 1

    work page 2024

  4. [4]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 3213–3223, 2016. 5

    2016

  5. [5]

    Space rotation with basis transformation for training-free test-time adaptation

    Chenhao Ding, Xinyuan Gao, Songlin Dong, Yuhang He, Qiang Wang, Xiang Song, Alex Kot, and Yihong Gong. Space rotation with basis transformation for training-free test-time adaptation. arXiv preprint arXiv:2502.19946 ,

    arXiv

  6. [6]

    Domain incremental object detection based on feature space topology preserv- ing strategy

    Li Ding, Xiang Song, Yuhang He, Changxin Wang, Songlin Dong, Xing Wei, and Yihong Gong. Domain incremental object detection based on feature space topology preserv- ing strategy. IEEE Transactions on Circuits and Systems for Video Technology, 34(1):424–437, 2023. 2

    work page 2023

  7. [7]

    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 Vi- sion, pages 18711–18720, 2023. 2

    work page 2023

  8. [8]

    An image is worth 16x16 words: Trans- formers 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, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6

    arXiv 2010

Show all 61 references
  1. [9]

    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. 6

    2022

  2. [10]

    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, 88:303–338, 2010. 5

    2010

  3. [11]

    Overcoming catastrophic forgetting in incremental object detection via elastic response distillation

    Tao Feng, Mang Wang, and Hangjie Yuan. Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 9427–9436, 2022. 3, 6

    2022

  4. [12]

    Beyond prompt learning: Continual adapter for efficient rehearsal-free continual learning

    Xinyuan Gao, Songlin Dong, Yuhang He, Qiang Wang, and Yihong Gong. Beyond prompt learning: Continual adapter for efficient rehearsal-free continual learning. In European Conference on Computer Vision , pages 89–106. Springer,

  5. [13]

    Multi-domain incremental learning for semantic segmenta- tion

    Prachi Garg, Rohit Saluja, Vineeth N Balasubramanian, Chetan Arora, Anbumani Subramanian, and CV Jawahar. Multi-domain incremental learning for semantic segmenta- tion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 761–771, 2022. 2

    2022

  6. [14]

    Csr-i (wsj0) complete ldc93s6a

    John Garofolo, David Graff, Doug Paul, and David Pallett. Csr-i (wsj0) complete ldc93s6a. Web Download. Philadel- phia: Linguistic Data Consortium, 83, 1993. 5

    1993

  7. [15]

    Learn by reasoning: Analogical weight generation for few-shot class- incremental learning

    Jizhou Han, Chenhao Ding, Yuhang He, Songlin Dong, Qiang Wang, Xinyuan Gao, and Yihong Gong. Learn by reasoning: Analogical weight generation for few-shot class- incremental learning. arXiv preprint arXiv:2503.21258 ,

  8. [16]

    An end-to- end architecture for class-incremental object detection with knowledge distillation

    Yu Hao, Yanwei Fu, Yu-Gang Jiang, and Qi Tian. An end-to- end architecture for class-incremental object detection with knowledge distillation. In 2019 IEEE International Con- ference on Multimedia and Expo (ICME), pages 1–6. IEEE,

    2019

  9. [17]

    Distill- ing the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distill- ing the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2

    2015 arXiv

  10. [18]

    Learning a unified classifier incrementally via rebalancing

    Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831–839,

  11. [19]

    Depth-attentional features for single-image rain removal

    Xiaowei Hu, Chi-Wing Fu, Lei Zhu, and Pheng-Ann Heng. Depth-attentional features for single-image rain removal. In Proceedings of the IEEE/CVF Conference on computer vi- sion and pattern recognition, pages 8022–8031, 2019. 5

    2019

  12. [20]

    Minimum class confusion for versatile domain adaptation

    Ying Jin, Ximei Wang, Mingsheng Long, and Jianmin Wang. Minimum class confusion for versatile domain adaptation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI 16, pages 464–480. Springer, 2020. 6

    2020

  13. [21]

    Class- incremental learning by knowledge distillation with adaptive feature consolidation

    Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class- incremental learning by knowledge distillation with adaptive feature consolidation. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 16071–16080, 2022. 2

    2022

  14. [22]

    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- 9 Barwinska, et al. Overcoming catastrophic forgetting in neu- ral networks. Proceedings of the national academy of ...

    2017

  15. [23]

    Clustering- based domain-incremental learning

    Christiaan Lamers, Ren ´e Vidal, Nabil Belbachir, Niki van Stein, Thomas B ¨aeck, and Paris Giampouras. Clustering- based domain-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 3384–3392, 2023. 1

    2023

  16. [24]

    Seril: Noise adaptive speech enhance- ment using regularization-based incremental learning

    Chi-Chang Lee, Yu-Chen Lin, Hsuan-Tien Lin, Hsin-Min Wang, and Yu Tsao. Seril: Noise adaptive speech enhance- ment using regularization-based incremental learning. arXiv preprint arXiv:2005.11760, 2020. 7

    2005 arXiv

  17. [25]

    A continual deepfake detection benchmark: Dataset, meth- ods, and essentials

    Chuqiao Li, Zhiwu Huang, Danda Pani Paudel, Yabin Wang, Mohamad Shahbazi, Xiaopeng Hong, and Luc Van Gool. A continual deepfake detection benchmark: Dataset, meth- ods, and essentials. In Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision , page...

    2023

  18. [26]

    Learning from students: Online contrastive distillation net- work for general continual learning

    Jin Li, Zhong Ji, Gang Wang, Qiang Wang, and Feng Gao. Learning from students: Online contrastive distillation net- work for general continual learning. In Proc. 31st Int. Joint Conf. Artif. Intell., pages 3215–3221, 2022. 2

    2022

  19. [27]

    Learning without forgetting

    Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelli- gence, 40(12):2935–2947, 2017. 2, 6

    2017

  20. [28]

    Few-shot class-incremental learning via entropy-regularized data-free replay

    Huan Liu, Li Gu, Zhixiang Chi, Yang Wang, Yuanhao Yu, Jun Chen, and Jin Tang. Few-shot class-incremental learning via entropy-regularized data-free replay. In European Con- ference on Computer Vision, pages 146–162. Springer, 2022. 2

    2022

  21. [29]

    Multi-task incremental learn- ing for object detection

    Xialei Liu, Hao Yang, Avinash Ravichandran, Rahul Bhotika, and Stefano Soatto. Multi-task incremental learn- ing for object detection. arXiv preprint arXiv:2002.05347 ,

    2002 arXiv

  22. [30]

    Continual detection transformer for incremen- tal object detection

    Yaoyao Liu, Bernt Schiele, Andrea Vedaldi, and Christian Rupprecht. Continual detection transformer for incremen- tal object detection. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 23799–23808, 2023. 3

    2023

  23. [31]

    Compositional prompting for anti-forgetting in domain incremental learn- ing

    Zichen Liu, Yuxin Peng, and Jiahuan Zhou. Compositional prompting for anti-forgetting in domain incremental learn- ing. International Journal of Computer Vision, pages 1–18,

  24. [32]

    Core50: a new dataset and benchmark for continuous object recognition

    Vincenzo Lomonaco and Davide Maltoni. Core50: a new dataset and benchmark for continuous object recognition. In Conference on robot learning, pages 17–26. PMLR, 2017. 5

    2017

  25. [33]

    Mop-clip: A mixture of prompt-tuned clip models for domain incremental learning

    Julien Nicolas, Florent Chiaroni, Imtiaz Ziko, Ola Ahmad, Christian Desrosiers, and Jose Dolz. Mop-clip: A mixture of prompt-tuned clip models for domain incremental learning. In Proceedings of the IEEE/CVF Winter Conference on Ap- plications of Computer Vision, pages 1762–177...

    2024

  26. [34]

    Moment matching for multi-source domain adaptation

    Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 1406–1415,

  27. [35]

    Learning transferable visual models from natural language supervi- sion

    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 supervi- sion. In International conference on machine learning, ...

    2021

  28. [36]

    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 con- ference on Computer Vision and Pattern Recognition, pages 2001–2010, 2017. 2

    2001

  29. [37]

    A unified approach to do- main incremental learning with memory: Theory and algo- rithm

    Haizhou Shi and Hao Wang. A unified approach to do- main incremental learning with memory: Theory and algo- rithm. Advances in Neural Information Processing Systems, 36, 2024. 1

    2024

  30. [38]

    Multi-granularity knowl- edge distillation and prototype consistency regularization for class-incremental learning

    Yanyan Shi, Dianxi Shi, Ziteng Qiao, Zhen Wang, Yi Zhang, Shaowu Yang, and Chunping Qiu. Multi-granularity knowl- edge distillation and prototype consistency regularization for class-incremental learning. Neural Networks, 164:617–630,

  31. [39]

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

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

    2023

  32. [40]

    Non-exemplar domain incremental object detection via learning domain bias

    Xiang Song, Yuhang He, Songlin Dong, and Yihong Gong. Non-exemplar domain incremental object detection via learning domain bias. In Proceedings of the AAAI Con- ference on Artificial Intelligence, pages 15056–15065, 2024. 2, 3, 5

    2024

  33. [41]

    Overcoming catastrophic forgetting for multi-label class-incremental learning

    Xiang Song, Kuang Shu, Songlin Dong, Jie Cheng, Xing Wei, and Yihong Gong. Overcoming catastrophic forgetting for multi-label class-incremental learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision, pages 2389–2398, 2024. 2

    2024

  34. [42]

    Assessment for automatic speech recognition: Ii

    Andrew Varga and Herman JM Steeneken. Assessment for automatic speech recognition: Ii. noisex-92: A database and an experiment to study the effect of additive noise on speech recognition systems. Speech communication , 12(3):247– 251, 1993. 5

    1993

  35. [43]

    Instance rela- tion graph guided source-free domain adaptive object detec- tion

    Vibashan VS, Poojan Oza, and Vishal M Patel. Instance rela- tion graph guided source-free domain adaptive object detec- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 3520–3530,

  36. [44]

    Multi-domain incremental learning for face presentation attack detection

    Keyao Wang, Guosheng Zhang, Haixiao Yue, Ajian Liu, Gang Zhang, Haocheng Feng, Junyu Han, Errui Ding, and Jingdong Wang. Multi-domain incremental learning for face presentation attack detection. In Proceedings of the AAAI Conference on Artificial Intelligence , pages 5499– 550...

    2024

  37. [45]

    Non-exemplar domain incremental learning via cross-domain concept integration

    Qiang Wang, Yuhang He, Songlin Dong, Xinyuan Gao, Shaokun Wang, and Yihong Gong. Non-exemplar domain incremental learning via cross-domain concept integration. In European Conference on Computer Vision , pages 144–

  38. [46]

    Dualcp: Rehearsal-free domain-incremental learning via dual-level concept proto- type

    Qiang Wang, Yuhang He, Songlin Dong, Xiang Song, Jizhou Han, Haoyu Luo, and Yihong Gong. Dualcp: Rehearsal-free domain-incremental learning via dual-level concept proto- type. arXiv preprint arXiv:2503.18042, 2025. 2

    2025 arXiv

  39. [47]

    Importance-aware shared param- eter subspace learning for domain incremental learning

    Shiye Wang, Changsheng Li, Jialin Tang, Xing Gong, Ye Yuan, and Guoren Wang. Importance-aware shared param- eter subspace learning for domain incremental learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8874–8883, 2024. 3

    2024

  40. [48]

    S-prompts learning with pre-trained transformers: An occam’s razor for domain incremental learning

    Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam’s razor for domain incremental learning. Advances in Neural Informa- tion Processing Systems, 35:5682–5695, 2022. 2, 3, 5, 6

    2022

  41. [49]

    Isolation and impartial aggre- gation: A paradigm of incremental learning without interfer- ence

    Yabin Wang, Zhiheng Ma, Zhiwu Huang, Yaowei Wang, Zhou Su, and Xiaopeng Hong. Isolation and impartial aggre- gation: A paradigm of incremental learning without interfer- ence. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10209–10217, 2023. 1

    2023

  42. [50]

    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, Vin- cent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. InEuropean Conference on Computer Vision , pages 631–648. Springer,

  43. [51]

    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,

  44. [52]

    Choice of peft technique in continual learning: Prompt tuning is not all you need

    Martin Wistuba, Prabhu Teja Sivaprasad, Lukas Balles, and Giovanni Zappella. Choice of peft technique in continual learning: Prompt tuning is not all you need. arXiv preprint arXiv:2406.03216, 2024. 3

    2024 arXiv

  45. [53]

    Improving monaural speech enhancement by mapping to fixed simulation space with knowledge distilla- tion

    Xinmeng Xu. Improving monaural speech enhancement by mapping to fixed simulation space with knowledge distilla- tion. IEEE Signal Processing Letters, 2024. 3

    2024

  46. [54]

    An ex- perimental study on speech enhancement based on deep neu- ral networks

    Yong Xu, Jun Du, Li-Rong Dai, and Chin-Hui Lee. An ex- perimental study on speech enhancement based on deep neu- ral networks. IEEE Signal processing letters , 21(1):65–68,

  47. [55]

    One-shot replay: Boosting incremental object detection via retrospecting one object

    Dongbao Yang, Yu Zhou, Xiaopeng Hong, Aoting Zhang, and Weiping Wang. One-shot replay: Boosting incremental object detection via retrospecting one object. In Proceed- ings of the AAAI Conference on Artificial Intelligence, pages 3127–3135, 2023. 3

    2023

  48. [56]

    Learning noise adapters for incremental speech en- hancement

    Ziye Yang, Xiang Song, Jie Chen, C´edric Richard, and Israel Cohen. Learning noise adapters for incremental speech en- hancement. IEEE Signal Processing Letters, 2024. 2, 3, 5, 7

    2024

  49. [57]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Dar- rell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition ...

    2020

  50. [58]

    Speech enhance- ment using deep learning methods: A review

    Asri Rizki Yuliani, M Faizal Amri, Endang Suryawati, Ade Ramdan, and Hilman Ferdinandus Pardede. Speech enhance- ment using deep learning methods: A review. Jurnal Elek- tronika dan Telekomunikasi, 21(1):19–26, 2021. 3

    2021

  51. [59]

    Contin- ual learning through synaptic intelligence

    Friedemann Zenke, Ben Poole, and Surya Ganguli. Contin- ual learning through synaptic intelligence. In International conference on machine learning, pages 3987–3995. PMLR,

  52. [60]

    Prototype augmentation and self-supervision for incremental learning

    Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and Cheng- Lin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 5871–5880, 2021. 6 11

    2021

  53. [162]

    2, 3, 5, 6 10

    Springer, 2024. 2, 3, 5, 6 10

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.