A Comprehensive Study of Class Incremental Learning Algorithms for Visual Tasks
Open accessThe DOI matches its DataCite record; 102 of the 102 references it cites were found in a registry or an index; 43 of the 89 sources its claims cite were checked for support.
Abstract
The ability of artificial agents to increment their capabilities when\nconfronted with new data is an open challenge in artificial intelligence. The\nmain challenge faced in such cases is catastrophic forgetting, i.e., the\ntendency of neural networks to underfit past data when new ones are ingested. A\nfirst group of approaches tackles forgetting by increasing deep model capacity\nto accommodate new knowledge. A second type of approaches fix the deep model\nsize and introduce a mechanism whose objective is to ensure a good compromise\nbetween stability and plasticity of the model. While the first type of\nalgorithms were compared thoroughly, this is not the case for methods which\nexploit a fixed size model. Here, we focus on the latter, place them in a\ncommon conceptual and experimental framework and propose the following\ncontributions: (1) define six desirable properties of incremental learning\nalgorithms and analyze them according to these properties, (2) introduce a\nunified formalization of the class-incremental learning problem, (3) propose a\ncommon evaluation framework which is more thorough than existing ones in terms\nof number of datasets, size of datasets, size of bounded memory and number of\nincremental states, (4) investigate the usefulness of herding for past\nexemplars selection, (5) provide experimental evidence that it is possible to\nobtain competitive performance without the use of knowledge distillation to\ntackle catastrophic forgetting and (6) facilitate reproducibility by\nintegrating all tested methods in a common open-source repository. The main\nexperimental finding is that none of the existing algorithms achieves the best\nresults in all evaluated settings. Important differences arise notably if a\nbounded memory of past classes is allowed or not.\n
Overview
CoChat Summary
TL;DR
This paper provides a comparative study of class incremental learning (IL) methods for visual tasks under a unified formalization and evaluation protocol. It finds that no single algorithm is best in all settings, and that the best choice depends strongly on whether a bounded memory of past classes is available.
Key findings
- With memory allowed, the strongest overall results came from methods that treat class IL as an imbalanced learning problem: the paper reports FT th as best overall with GIL = -3.62, followed by ScaIL (-3.70), BiC (-4.03), and LUCIR CNN (-4.13).
- In the memory setting, the authors report that vanilla FT can outperform iCaRL globally: FT has GIL = -5.19 versus iCaRL’s -7.36; they also state that FT is globally better than iCaRL by over 2 GIL points.
- Without memory, fixed-representation methods are reported to be much more competitive than fine-tuning methods; the best global score is said to come from DeeSIL, followed by Deep-SLDA, FR, and REMIND.
- Herding-based exemplar selection outperformed random selection for all methods tested in the reported experiments, though the gain varied by method.
Methods
- The paper compares class incremental learning algorithms in a common conceptual and experimental framework, focusing on methods for constant model complexity with bounded or no memory of past classes.
- Experiments use four datasets: ILSVRC (1,231,167 train / 50,000 test), VGGFACE2 (491,746 / 50,000), LANDMARKS (374,367 / 20,000), and CIFAR100 (50,000 / 10,000).
- Evaluation uses a ResNet-18 backbone with SGD, top-5 accuracy averaged over incremental states, and the Incremental Learning Gap (GIL) across configurations with varying numbers of states (T = 10, 20, 50) and memory sizes (including K = 2%, 1%, 0.5%, and K = 0 in the no-memory case).
Limitations
- The paper states that the evaluation scenario is still simplified and recommends future work on more realistic settings: no assumption of readily annotated incoming data, variable numbers of classes per incremental state, and imbalanced datasets.
- The authors note that their conclusions are limited to the tested setting of fixed-complexity visual class incremental learning; methods that grow model capacity, generative replay methods, and some online approaches are discussed but not evaluated under this protocol.
- Evident from the text: the study is constrained to four datasets and a specific set of algorithm families, so the findings may not generalize beyond the tested visual tasks and configurations.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
28 Valid · 3 Not valid · 29 Not checked · 127 Own results and interpretations
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“The method does not need a bounded memory for the update.”
What the sources say
- Disputedcites [56] [56] Rebuffi, S., Kolesnikov, A., Sperl, G., Lampert, C.H., 2017b. icarl: Incremental classifier and representation learning, in: Conference on Computer Vision and Pattern Recognition. - sourceExperimental — accuracy not yet measured
If the cited method is iCaRL, the paper says the opposite. Section 2.4 (p. 4) states that its stored exemplars never exceed a memory budget of K, and the representation-learning discussion (p. 4) says those exemplars are used during updates. The copy does not identify a method called FR. (LLM, conf=0.84)
Checked by a reasoning model — may be wrong.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“However, fixed-representations work well only if the task does not change over time, as it is the case in the evaluated scenarios presented in [2].”
What the sources say
- Disputedcites [2] [2] Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T., 2018. Memory aware synapses: Learning what (not) to forget, in: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (Eds.), Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part III, Springer. pp. 144-161. - sourceExperimental — accuracy not yet measured
The evaluated scenarios do change tasks over time. Section 5.1 (p. 10) describes sequences such as Scenes → Birds, and the Longer Sequence results (p. 11) evaluate performance after successive tasks. The paper therefore contradicts the claim that its evaluated scenarios involve an unchanging task; it does not establish the broader rule about when fixed representations work well. (LLM, conf=0.89)
Checked by a reasoning model — may be wrong.
plainstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“FT BAL is also tributary to the fixed memory K because past exemplars are needed for the balancing step.”
What the sources say
- Supportedcites [7] [7] Belouadah, E., Popescu, A., 2019. Il2m: Class incremental learning with dual memory, in: Proceedings of the IEEE International Conference on Computer Vision, pp. 583-592. - source
Section 5.3.1 says FT BAL uses fine-tuning with memory and that, as memory decreases, more data must be dropped during its balancing step. Section 5.1 identifies that step as balanced fine-tuning after the initial imbalanced training. Together, these passages support the claim that FT BAL's balancing depends on stored past exemplars. (LLM, conf=0.88)
- Not matchedcites [8] [8] Belouadah, E., Popescu, A., 2020. Scail: Classifier weights scaling for class incremental learning, in: The IEEE Winter Conference on Applications of Computer Vision (WACV). -
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Partly supportedcites [14] [14] Castro, F.M., Marín-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K., 2018. End-to-end incremental learning, in: Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 814, 2018, Proceedings, Part XII, pp. 241-257. - sourceExperimental — accuracy not yet measured
Section 4 (Incremental Learning; page not provided) says balanced fine-tuning uses equal numbers of samples from old and new classes, supporting the need for stored past exemplars. But Section 3.1 (Representative memory; page not provided) describes both a fixed-capacity memory K and a memory that grows with the number of classes. The balancing step therefore depends on past exemplars, not specifically on fixed memory K. (LLM, conf=0.87)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“one early example of such work [56] is shown to outperform continual learning approaches when tested in a common experimental setting [35].”
What the sources say
- Supportedcites [56] [56] Rebuffi, S., Kolesnikov, A., Sperl, G., Lampert, C.H., 2017b. icarl: Incremental classifier and representation learning, in: Conference on Computer Vision and Pattern Recognition. - source
The Abstract (p. 1) identifies iCaRL as a class-incremental learning strategy. The Results (p. 7) report that iCaRL outperformed the other methods tested. The separate citation [35] is attached to the common experimental setting, so that setting is not attributed to this source. (LLM, conf=0.91)
- Supportedcites [35] [35] Lange, M.D., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G.G., Tuytelaars, T., 2019. Continual learning: A comparative study on how to defy forgetting in classification tasks. CoRR abs/1909.08383. - source
Section 5.1 (p. 4) identifies iCaRL as an early class-incremental method and says the authors adapted it to their task-incremental, multi-head setting for a fair comparison. Section 6.3.1 (p. 10) reports that iCaRL outperformed all compared continual-learning methods for the DEEP model on Tiny Imagenet. The reported result is specific to that configuration. (LLM, conf=0.92)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“FT NEM can only be deployed if a bounded memory is allowed to alleviate catastrophic forgetting.”
What the sources say
- Supportedcites [7] [7] Belouadah, E., Popescu, A., 2019. Il2m: Class incremental learning with dual memory, in: Proceedings of the IEEE International Conference on Computer Vision, pp. 583-592. - source
Section 5.1 defines FT NEM by its nearest-exemplars-mean classifier. Section 5.3 reports that, with no memory (K = 0), all methods derived from fine tuning perform identically because their score-rectification methods rely on exemplars. Thus FT NEM’s distinctive classifier requires exemplar memory, even though the paper still reports a K = 0 result under the FT NEM label. (LLM, conf=0.86)
- Not matchedcites [8] [8] Belouadah, E., Popescu, A., 2020. Scail: Classifier weights scaling for class incremental learning, in: The IEEE Winter Conference on Applications of Computer Vision (WACV). -
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
106 more claims · 6 with their cited sources checked
Create a free account to see all 187 claims and to verify the 12 references still open.
- It may state two separate claims as one. 52
- Probably not a substantial claim. 27
- The quote only partly states this claim. 2
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The check wasn’t confident either way, so it gives no verdict — never disputed.
In the citing paper
“Although some surveys briefly mention FSCIL [32, 35], they only provide a short introduction to the concept and a few studies without offering a thorough or systematic analysis.”
From Few-Shot Class-Incremental Learning for Classification and Object Detection: A Survey (2025) · cites this paper as [32]
Experimental — accuracy not yet measured
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 5×
“In an end-to-end DNN [17], a softmax function is applied to transform raw scores 𝒐 𝑡 = ( 1 𝑡 , 𝑜 2 𝑡 , ..., ) to probabilities 𝒑 𝑡 = ( 1 𝑡 , 𝑝 2 𝑡 , ..., ) .”
BackgroundMethodInfluentialA Mathematical Theory of Communication
Shannon, C. E. (1948) · Bell System Technical Journal · doi:10.1002/j.1538-7305.1948.tb01338.x · Cited by 88,486
Scikit-learn: Machine Learning in Python
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Müller, A., Nothman, J., Louppe, G., et al. (2012) · arXiv · doi:10.48550/arxiv.1201.0490 · Cited by 71,016
A Mathematical Theory of Communication
Shannon, C. E. (1948) · Bell System Technical Journal · doi:10.1002/j.1538-7305.1948.tb00917.x · Cited by 48,949
ImageNet Large Scale Visual Recognition Challenge
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. (2015) · International Journal of Computer Vision · doi:10.1007/s11263-015-0816-y · Cited by 39,380 · cited 2×
“• ILSVRC [47]”
MethodA mathematical theory of communication
Hockett, C. F., Shannon, C., & Weaver, W. (1948) · MOCO · doi:10.1145/584091.584093 · Cited by 30,543 · cited 1×
“We also tried exemplar selection techniques inspired from active learning approaches such as: entropy [52], min margin [51], core-set [50], k-means [62].”
MethodLearning Multiple Layers of Features from Tiny Images
Krizhevsky, A. (2024) · doi:10.57702/zp44cu3g · Cited by 23,395
Distilling the Knowledge in a Neural Network
Hinton, G., Vinyals, O., & Dean, J. (2015) · arXiv · doi:10.48550/arxiv.1503.02531 · Cited by 14,204
A training algorithm for optimal margin classifiers
Boser, B. E., Guyon, I. M., & Vapnik, V. N. (1992) · Proceedings of the fifth annual workshop on Computational learning theory · doi:10.1145/130385.130401 · Cited by 11,992 · cited 1×
“batch of classes and a battery of Support Vector Machines (SVMs) [7] to incrementally learn new classes.”
Background(2017) · Cited by 11,625
Research works printed but not linked: 4
- [19][19] Fritzke, B., 1994b. A growing neural gas network learns topologies, in: Tesauro, G., Touretzky, D.S., Leen, T.K. (Eds.), Advances in Neural Information Processing Systems 7, [NIPS Conference, Denver, Colorado, USA, 1994], MIT Press. pp. 625-632. -
- [70][70] Tao, X., Chang, X., Hong, X., Wei, X., Gong, Y., 2020a. Topologypreserving class-incremental learning, in: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (Eds.), Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX, Springer. pp. 254-270. -
- [73][73] Wang, Y., Ramanan, D., Hebert, M., 2017. Growing a brain: Finetuning by increasing model capacity, in: Conference on Computer Vision and Pattern Recognition. -
- [80][80] Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., van de Weijer, J., 2020. Semantic drift compensation for classincremental learning, in: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, IEEE. pp. 6980-6989. -
Cited by
Class-Incremental Learning: Survey and Performance Evaluation on Image Classification
Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A. D., & van de Weijer, J. (2023) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2022.3213473 · Cited by 740
“[63] perform a comparison of several classIL methods.”
MethodThree types of incremental learning
van de Ven, G. M., Tuytelaars, T., & Tolias, A. S. (2022) · Nature Machine Intelligence · doi:10.1038/s42256-022-00568-3 · Cited by 665
“An important challenge in this scenario is learning to discriminate between classes that are not observed together, which has turned out to be very challenging for deep neural networks, especially when storing examples of previously seen classes is not allowed 29,30 .”
Class-Incremental Learning: A Survey
Zhou, D.-W., Wang, Q.-W., Qi, Z.-H., Ye, H.-J., Zhan, D.-C., & Liu, Z. (2024) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2024.3429383 · Cited by 376
“Other surveys either focus on the specific field [20], [21] or lack the performance evolution among state-of-the-arts [8], [22], [23].”
BackgroundDing, A., Qin, Y., Wang, B., Guo, L., Jia, L., & Cheng, X. (2024) · Mechanical Systems and Signal Processing · doi:10.1016/j.ymssp.2024.111175 · Cited by 254
SS-IL: Separated Softmax for Incremental Learning
Ahn, H., Kwak, J., Lim, S., Bang, H., Kim, H., & Moon, T. (2021) · 2021 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.00088 · Cited by 177
“Also, results in [4] and [6] show that CIL models are sensitive to incremental conditions such as the number of incremental tasks (T ) and exemplar-memory size (|M|).”
BackgroundMethodResult
Metadata
- Authors
- Belouadah, E., Popescu, A., & Kanellos, I.
- Affiliations
- Commissariat à l'Énergie Atomique et aux Énergies Alternatives (FR)
- IMT Atlantique (FR)
- Corresponding author
- Popescu, Adrian
- Venue
- arXiv
- Publisher
- arXiv
- Published
- 2020
- Volume
- 135
- Pages
- 38-54
- ISSN
- 0893-6080 (ISSN-L)
- arXiv
- 2011.01844
- OpenAlex
- W3096398238
- Cited by
- 235
- References
- 86
- Citations by year
- 2026: 28 · 2025: 40 · 2024: 61 · 2023: 58 · 2022: 32 · 2021: 16
- FWCI
- 0.91
- Type
- journal article
- Language
- en
- Open access
- bronze
- doi.org
- Locations
- Neural Networks (acceptedVersion)
- PubMed (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- HAL (submittedVersion, cc-by-nc)
- Zenodo (CERN European Organization for Nuclear Research) (submittedVersion, cc-by)
- arXiv (Cornell University)
- Funders
- European Commission (951911, HORIZON2020)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Data Stream Mining Techniques
- Keywords
- class-incremental learning, catastrophic forgetting, stability-plasticity balance, herding behavior, knowledge distillation, evaluation framework, reproducibility, open-source repositories, visual task
- MeSH terms
- Algorithms, Artificial Intelligence, Humans, Memory, Psychomotor Performance, Visual Perception, Reproducibility of Results, Neural Networks, Computer
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: A Comprehensive Study of Class Incremental Learning Algorithms for Visual Tasks. https://app.cochat.ai/papers/10.48550/arxiv.2011.01844
- The paper
- Belouadah, E., Popescu, A., & Kanellos, I. (2020). A Comprehensive Study of Class Incremental Learning Algorithms for Visual Tasks. arXiv. https://doi.org/10.48550/arxiv.2011.01844
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.