Constrained Few-shot Class-incremental Learning
Open accessAbstract
Continually learning new classes from fresh data without forgetting previous knowledge of old classes is a very challenging research problem. Moreover, it is imperative that such learning must respect certain memory and computational constraints such as (i) training samples are limited to only a few per class, (ii) the computational cost of learning a novel class remains constant, and (iii) the memory footprint of the model grows at most linearly with the number of classes observed. To meet the above constraints, we propose C-FSCIL, which is architecturally composed of a frozen meta-learned feature extractor, a trainable fixed-size fully connected layer, and a rewritable dynamically growing memory that stores as many vectors as the number of encountered classes. C-FSCIL provides three update modes that offer a trade-off between accuracy and compute-memory cost of learning novel classes. C-FSCIL exploits hyperdimensional embedding that allows to continually express many more classes than the fixed dimensions in the vector space, with minimal interference. The quality of class vector representations is further improved by aligning them quasi-orthogonally to each other by means of novel loss functions. Experiments on the CIFAR100, mini-ImageNet, and Omniglot datasets show that C-FSCIL outperforms the baselines with remarkable accuracy and compression. It also scales up to the largest problem size ever tried in this few-shot setting by learning 423 novel classes on top of 1200 base classes with less than 1.6% accuracy drop. Our code is available at https://github.com/IBM/constrained-FSCIL.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
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?
C-FSCIL framework consists of a feature extractor trained by meta-learning, a trainable fully connected layer, and a rewritable explicit memory.
“Hersche et al. [24] proposed a C-FSCIL framework, which consisted of a feature extractor trained by meta-learning, a trainable fully connected layer, and a rewritable explicit memory.”
From Few-Shot Class-Incremental Learning for Classification and Object Detection: A Survey (2025) · cites this paper as [24]
The indexed copy provides a repository listing for the paper (section “4.01 Contributo in Atti di convegno,” no page number), but no passage describing the framework’s components. Searches for the rewritable explicit memory and the meta-learned feature extractor and fully connected layer returned only repository text, so the claimed architecture cannot be checked against this copy. (LLM, conf=0.72)
Experimental — accuracy not yet measuredHyperdimensional embedding has three advantages: high probability of quasi-orthogonality between random vectors, rich expressive space, and good semantic representation capability.
“The core idea was introducing hyperdimensional embedding, which has three advantages: (1) the high probability of quasi-orthogonality between random vectors, (2) rich expressive space, and (3) good semantic representation capability.”
From Few-Shot Class-Incremental Learning for Classification and Object Detection: A Survey (2025) · cites this paper as [24]
The only passage in our indexed copy is a repository file listing (section ‘4.01 Contributo in Atti di convegno’, no page), not the paper’s substantive text. Searches for quasi-orthogonality, expressive space, and semantic representation found no passage addressing the claimed advantages, so the claim cannot be verified from our copy. (LLM, conf=0.72)
Experimental — accuracy not yet measuredUtilizing the average of these features to create class prototypes surpassed the state-of-the-art methods at that time.
“Utilizing the average of these features to create class prototypes surpassed the state-of-the-art methods at that time.”
From Few-Shot Class-Incremental Learning for Classification and Object Detection: A Survey (2025) · cites this paper as [24]
The available passage (repository listing, no page number) identifies a PDF of the paper but does not contain its methods or results. Searches for averaged features, class prototypes, and state-of-the-art performance found no relevant passage in our indexed copy, so the claim cannot be verified from it. (LLM, conf=0.78)
Experimental — accuracy not yet measuredTraining the backbone using the meta-learning strategy effectively can extract robust features.
“empirical evidence demonstrated that training the backbone using the meta-learning strategy effectively can extract robust features.”
From Few-Shot Class-Incremental Learning for Classification and Object Detection: A Survey (2025) · cites this paper as [24]
The indexed passages are repository-listing text, including a PDF filename under “4.01 Contributo in Atti di convegno” (no page number), rather than passages describing the study’s methods or results. Searches for meta-learning, backbone training, and robust features found no passage addressing the claim; the decisive paper text may be missing from the indexed copy. (LLM, conf=0.72)
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
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 · see citing works
ImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155 · see citing works
Overcoming catastrophic forgetting in neural networks
Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al. (2017) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.1611835114 · Cited by 7,899 · see citing works
iCaRL: Incremental Classifier and Representation Learning
Rebuffi, S.-A., Kolesnikov, A., Sperl, G., & Lampert, C. H. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.587 · Cited by 3,988 · see citing works
Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem
McCloskey, M., & Cohen, N. J. (1989) · Psychology of Learning and Motivation · doi:10.1016/s0079-7421(08)60536-8 · Cited by 3,770 · see citing works
Unsupervised Feature Learning via Non-parametric Instance Discrimination
Wu, Z., Xiong, Y., Yu, S. X., & Lin, D. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2018.00393 · Cited by 3,664 · see citing works
Human-level concept learning through probabilistic program induction
Lake, B. M., Salakhutdinov, R., & Tenenbaum, J. B. (2015) · Science · doi:10.1126/science.aab3050 · Cited by 2,817 · see citing works
Flagged first, then by citations
Cited by
Sorted by citations
A Comprehensive Survey of Continual Learning: Theory, Method and Application
Wang, L., Zhang, X., Su, H., & Zhu, J. (2024) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2024.3367329 · Cited by 1,113 · see citing works
“…extraconstraintsfrommargin-basedrepresentations[258],etc. Thesecondistoobtainanadaptiveclassifierfromasequence of novel classes, such as by evolving the classifier weights with a graph attention network [259], performing hyperdimen-sional computing [260], sampling stochastic classifiers [261], etc.”
Few-Shot Class-Incremental Learning for Classification and Object Detection: A Survey
Zhang, J., Liu, L., Silvén, O., Pietikäinen, M., & Hu, D. (2025) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2025.3529038 · Cited by 84 · see citing works
“For instance, in the C-FSCIL framework proposed by Hersche et al. [24], empirical evidence demonstrated that training the backbone using the meta-learning strategy effectively can extract robust features.”
Influential
- Authors
- Hersche, M., Karunaratne, G., Cherubini, G., Benini, L., Sebastian, A., & Rahimi, A.
- Affiliations
- ETH Zurich (CH)
- IBM Research - Zurich (CH)
- ETH Zurich
- Venue
- 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
- Publisher
- IEEE
- Published
- 2022-06
- Pages
- 9047-9057
- ISSN
- 1063-6919 (ISSN-L)
- OpenAlex
- W4312327301
- Cited by
- 167
- References
- 7
- Citations by year
- 2026: 17 · 2025: 52 · 2024: 59 · 2023: 30 · 2022: 9
- FWCI
- 20.74
- Type
- conference paper
- Language
- en
- Open access
- green
- hdl.handle.net
- Locations
- IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings (publishedVersion)
- Archivio istituzionale della ricerca (Alma Mater Studiorum Università di Bologna) (submittedVersion, other-oa)
- 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, few-shot learning, continual learning, catastrophic forgetting, memory-efficient models, model compression
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.
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
