Continual Learning with Support Boundary Experience Blending

Chih-Fan Hsu; Ming-Ching Chang; Wei-Chao Chen

REVIEW 3 major objections 5 minor 26 references

Continual Learning with Support Boundary Experience Blending

T0 review · 3 major / 5 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 Latent-space noise injected into replay exemplars generates support boundary data, and blending it into training via dual-model aggregation cuts catastrophic forgetting and improves average accuracy by 6-14 points on four…

desk verdict Original SBD+DMA idea with strong short-horizon gains, but the ImageNet memory accounting and the 25-task results undercut the headline consistency claim. read the letter →

arxiv 2507.23534 v4 pith:FUOQQIXE submitted 2025-07-31 cs.LG cs.CV

Chih-Fan Hsu , Ming-Ching Chang , Wei-Chao Chen This is my paper · ORCID

classification cs.LGcs.CV

keywords continuallearningcatastrophicforgettingexperiencereplaysupportboundarydatadecisionregularizationdifferentialprivacydual-modelaggregationblurred

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 claims that replay-based continual learning can be substantially improved by adding Support Boundary Data (SBD): latent feature vectors perturbed with differential-privacy-style Laplace noise so that they land near class decision boundaries. Because these synthetic samples are stored alongside ordinary exemplars and used only during training, the model sees boundary-adjacent points that implicitly regularize the classifier and reduce catastrophic forgetting. The authors back this with a framework, Experience Blending (EB), that trains two models—one on replay plus SBD and one on SBD alone—and averages their weights, and report average-accuracy gains of 10.45% on CIFAR-10, 6.26% on CIFAR-100, 13.75% on Tiny ImageNet, and 9.15% on ImageNet-1k over existing continual-learning baselines. If these results hold, boundary-sample generation is a simple, inference-free add-on to experience replay that works across class-incremental, domain-incremental, and blurred-boundary settings.

What carries the argument

The load-bearing object is Support Boundary Data (SBD). For each task, an image encoder followed by a self-attention layer produces features, and the batch is perturbed as $E_b = f'_b + \mathrm{Lap}\left(0, \frac{\max(f'_b)-\min(f'_b)}{\epsilon |S^t_b|}\right)$, with $\epsilon=0.005$ by default. The noise is differential-privacy-inspired but used as regularization: it deliberately lowers the model's confidence on these points by placing them in ambiguous regions. The second piece is dual-model aggregation (DMA): two identical models are trained, one on replay memory plus SBD and one on SBD only, and their weights are averaged with $\alpha=0.5$ (FedAvg-style), which the authors find is necessary because training on SBD alone causes the model to memorize the synthetic samples rather than generalize. The self-attention layer emphasizes class-specific features before noise is added, and first-task fine-tuning aligns the encoder to the downstream domain; after the first task the encoder is frozen.

What would settle it

Train EB on a stream where the first task covers only a small, unrepresentative subset of classes, then check empirically whether the SBD generated for later tasks still falls near the decision boundary (by measuring margin distributions). If the margin gains vanish or accuracy reverts to the ER baseline, the frozen-encoder assumption fails. A sharper test: replace the Laplace noise with Gaussian or uniform noise of the same variance; if the gains disappear, the specific DP-inspired heavy-tailed mechanism, rather than mere perturbation, is what does the work.

Watch

Extended reading notes

Core claim

The central discovery is that injecting multivariate Laplace noise into the latent representations of stored data—a mechanism borrowed from differential privacy but used for regularization, not privacy—produces synthetic samples that populate the ambiguous region around class decision boundaries. Augmenting experience replay with these Support Boundary Data (SBD) samples, and training with a dual-model aggregation (DMA) step that averages the weights of a replay+SBD model and an SBD-only model, consistently raises average task accuracy. On the i-Blurry-50-10 setting the gains over the best prior baselines are 10.45 points on CIFAR-10, 6.26 points on CIFAR-100, 13.75 points on Tiny ImageNet, and 9.15 points on ImageNet-1k (with a CLIP encoder), and the method also raises average margins and reduces margin variance, which the authors interpret as more stable decision boundaries.

Load-bearing premise

After the first task, the image encoder and self-attention layer are frozen, and the paper assumes this single encoder keeps producing SBD that lies near decision boundaries for every later task, even as the data distribution shifts; the observed accuracy drop of 20.6% when moving from 5 to 25 tasks shows how much rests on that encoder.

Editorial extensions

If this is right

Experience replay methods can be upgraded by adding a latent-noise SBD memory with no change to the inference-time model.
The frozen-encoder design means SBD generation is cheap during training (about +0.13s per epoch for 100 SBD samples) and free at test time.
Under a fixed memory budget, allocating part of the budget to SBD rather than raw images yields roughly 4% higher accuracy than CLIB and FOSTER at the same memory footprint (Figure 6).
The method transfers across encoder choices (ImageNet-pretrained ResNet-18 and CLIP ViT-B/32) and across task counts, with CLIP initialization nearly eliminating the 5-to-25-task degradation.
Blurred-boundary and domain-incremental settings, which are harder for typical class-incremental methods, show the largest relative gains, suggesting SBD helps most when task boundaries are noisy.

Reading between the lines

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

The paper does not compare SBD against other boundary-regularization tricks such as mixup in latent space; a natural extension is to test whether the gains come from the Laplace distribution's heavy tails or simply from adding any controlled noise near the boundary.
Because after the first task the encoder is frozen, the method's success depends on the first task being representative; an adaptive version that periodically refreshes the encoder (or the SBD) with new-task data could reduce the 20.6% drop seen at 25 tasks with the ImageNet encoder.
The margin statistics used for the analysis (average margin up, variance down) could serve as a cheap online diagnostic to decide when to regenerate SBD, which the authors leave as an open direction.
The DP-noise mechanism suggests a possible link to certified robustness or to privacy-utility trade-offs, but the paper deliberately separates SBD from privacy, so any privacy guarantee would require a separate analysis.

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.

Desk Editor's Note

read the letter

This is a serious replay-based CL paper with a genuinely new mechanism and unusually large reported gains. The core idea—DP-style Laplace noise in latent space to create boundary-adjacent support data, blended via FedAvg-style aggregation of two internal models—is not in the cited ER or regularization literature. The ablations in Table V support each component, and the margin analysis in Section III-D is a nice diagnostic. On CIFAR-100 and Tiny ImageNet the gains are large and consistent, and the writing is clear.

Now the soft spots, in order of severity. First, the long-horizon degradation. Table III shows EB(IN) dropping 20.6% from 5 to 25 tasks, and at 25 tasks it is 38.46 vs CLIB's 43.64—yet the caption says "our method outperforms the compared methods," which is false for that row. The explanation about limited first-task coverage is plausible, but it means the benefit hinges on an encoder that doesn't drift. The CLIP variant does much better (57.44 at 25 tasks), so the problem is specifically with the ImageNet-initialized encoder, which is the default for most experiments. The blanket consistency claim overstates what the data show.

Second, memory fairness on ImageNet-1k: baselines get 20k replay images, EB gets 10k replay plus 10k SBD. That is not a controlled comparison unless memory is measured in bytes and SBD latents are shown to be cheaper. The paper asserts fairness but doesn't show the byte math. Third, the convergence claim: citing FedAvg's non-IID convergence for a two-model internal aggregation overstates what the theorem covers. Fourth, minor but real: the abstract says 2% for ImageNet-1k while the full text says 9.15%, and the code is not released.

Bottom line: the core idea is sound and the ablation supports it, but the paper needs revision before the headline claims are credible. A serious referee should focus on memory accounting and long-horizon behavior. I'd send it to review rather than desk-reject.

Referee Report

3 major / 5 minor

Summary. The paper proposes Experience Blending (EB), a replay-based continual learning method that augments stored exemplars with Support Boundary Data (SBD): latent representations generated by injecting Laplace noise, motivated by the differential-privacy mechanism, into features produced by a first-task-fine-tuned encoder and self-attention layer. Training uses two models, one trained on replay plus SBD and one on SBD alone, aggregated by FedAvg-style weight averaging (DMA). The authors report large average-accuracy gains over prior replay methods on CIFAR-10, CIFAR-100, Tiny ImageNet, and ImageNet-1k under the i-Blurry-50-10 setting, and provide ablations, margin analyses, and memory-budget experiments supporting the contribution of each component.

Significance. If the reported gains hold under controlled comparisons, EB would be a simple, training-only addition to replay-based continual learning that improves decision-boundary stability. The paper's strengths are the clear component-level ablation in Table V, the margin evidence in Table I, the memory-budget analysis in Figure 6, and the breadth of datasets, including a large-scale ImageNet-1k result. The framing is not circular: the PCA and margin analyses are post hoc interpretations, not quantities that reduce to the fitted SBD by construction. However, two load-bearing concerns must be resolved: the default frozen-encoder version degrades sharply on longer task horizons and even falls below CLIB at 25 tasks, contradicting the paper's consistency claim; and the ImageNet-1k comparison does not control the replay-memory budget between EB and its baselines. The theoretical convergence claim for DMA is also overstated. These issues are fixable with additional experiments and corrected claims, but they currently affect the central message of 'consistent improvements.'

major comments (3)

[Section III-C and Table III]
[Section IV and Table II]
[Section III-B]

minor comments (5)

[Abstract]
[Section III-A]
[Table I]
[Section III-A]
[Section III-C]

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the SBD/EB construction does not reduce to its inputs and the main claims rest on external benchmarks and independent components.

full rationale

Walking the derivation chain, the paper's construction is non-circular. SBD is defined by an explicit generative operation—Laplace noise added to SA-layer outputs of a fixed encoder (Section III-A)—not in terms of the accuracy or forgetting metrics it later reports; the notion of being 'boundary-adjacent' is a hypothesis validated post hoc via PCA and margin statistics (Section III-D), not a definition that presupposes the claimed gains. The dual-model aggregation uses FedAvg from [20] with convergence cited from [21], both external and independent; no load-bearing argument reduces to a self-citation. The replay-memory importance sampler is taken from [12], and the encoder fine-tuning is a first-task procedure, not a fit to the evaluation targets. The choice of epsilon=0.005 is selected after testing on the evaluation datasets, which is a test-set-selection caveat, but it is a fixed scalar and does not make any reported accuracy forced by construction—the experiments are ordinary benchmark comparisons. Table III's decline of EB(IN) at 25 tasks is a genuine robustness weakness and an internal inconsistency with its caption, but it is a correctness/experimental-design concern, not circularity: the numbers are independently measured. The Limitations paragraph admits scope restrictions (task boundaries, overhead, memory allocation), none of which assert or imply that a result was assumed into existence. Therefore no circular step meeting the quote-and-reduction bar is present.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The central mechanism rests on SBD as a new type of replay content, but its boundary-adjacent property is assumed rather than proven. Two free parameters (epsilon, alpha) are chosen by hand or light tuning. The method adds no new physical entities; it is an algorithmic construct.

free parameters (2)

epsilon (Laplace noise scale divisor) = 0.005 (tuned over 0.001, 0.005, 0.01)
Section III-A states the authors tested epsilon values of 0.001, 0.005, and 0.01 and adopted 0.005 as default. This scale controls how much SBD is pushed toward boundaries and is chosen based on downstream accuracy.
alpha (aggregation weight) = 0.5
Section III-C sets the final model as M = (1-alpha)MR∪E + alpha*ME with alpha=0.5. No sweep or analysis is provided, so this is a hand-picked constant that determines the balance between the two models.

assumptions (4)

ad hoc to paper Laplace noise injected into latent features, with scale (max(f')-min(f'))/(epsilon*|S_b|), produces representations that lie near class decision boundaries.
Section III-A claims this design intuition, but it is not proven; only empirical PCA visualization (Figure 4) and margin statistics (Table I) are offered. The paper does not derive a bound relating noise scale to boundary proximity.
domain assumption A single first-task fine-tuning of PR and SA yields a feature encoder that remains valid for all subsequent tasks.
Section III-C states PR and SA are fixed after the first task. The paper acknowledges this breaks down as the number of tasks increases (Table III), so the assumption is load-bearing.
domain assumption FedAvg convergence guarantees for non-IID clients transfer to the dual-model aggregation of MR∪E and ME within a single continual learner.
Section III-B claims 'guaranteed convergence under arbitrary data distribution' citing [20,21]. The cited FedAvg convergence proofs assume multiple clients with their own data, not two models with overlapping data trained sequentially in CL; the transfer is unproven.
ad hoc to paper The transposed CNN PE can map SBD latent features back to image space without losing boundary information, so that cross-entropy loss on e with the original label y is meaningful.
Section III-C defines PE as a shallow transposed CNN but provides no analysis of reconstruction fidelity. The method depends on this mapping for the SBD branch of MR∪E and for ME.

invented entities (1)

Support Boundary Data (SBD)
purpose: Latent feature vectors with Laplace noise injected, stored and replayed to regularize decision boundaries in continual learning.
The paper presents PCA visualizations and margin statistics (Section III-D) as evidence, but these are computed on the same datasets and models used to design the method; there is no external prediction or independent benchmark that confirms SBD's boundary-adjacent property. It is an internal construct.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Continual Learning with Support Boundary Experience Blending." pith.science (2026). https://pith.science/paper/FUOQQIXE

@misc{pith2026250723534,
  author       = {Pith},
  title        = {Pith review of: Continual Learning with Support Boundary Experience Blending},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FUOQQIXE}},
  note         = {Machine review of arXiv:2507.23534}
}

read the original abstract

Continual learning (CL) seeks to mitigate catastrophic forgetting when models are trained with sequential tasks. A common approach, experience replay (ER), stores past exemplars but only sparsely approximates the data distribution, yielding fragile and oversimplified decision boundaries. We address this limitation by introducing Support Boundary Data (SBD), generated via differential-privacy-inspired noise into latent features to create boundary-adjacent representations that implicitly regularize decision boundaries. Building on this idea, we propose Experience Blending (EB), a framework that jointly trains on exemplars and SBD through a dual-model aggregation strategy. EB has two components: (1) latent-space noise injection to generate support boundary data, and (2) end-to-end training that jointly leverages exemplars and SBD. Unlike standard experience replay, SBD enriches the feature space near decision boundaries, leading to more stable and robust continual learning. Extensive experiments on CIFAR-10, CIFAR-100, Tiny ImageNet, and ImageNet1K demonstrate consistent accuracy improvements of 10%, 6%, 14%, 2%, respectively.

Figures

Figures reproduced from arXiv: 2507.23534 by the authors.

**Figure 2.** Experience blending: Our method aggregates the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** (a) We generate SBD with an image encoder [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Latent visualization. The dispersion of SBD beyond [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Validation and test accuracy plots. TABLE II: Comparison of average accuracy Aavg (%) across the datasets using the i-Blurry-50-10 setting. The gray rows show the prior works. Numbers following ± represent the standard deviation across three experiments with different …

**Figure 6.** Figure 6: This figure shows how varying replay memory [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 25 canonical work pages

[1]

A recent survey of heterogeneous transfer learning,

R. Bao, Y. Sun, Y. Gao, J. Wang, Q. Yang, Z.-H. Mao, and Y. Ye, “A recent survey of heterogeneous transfer learning,” arXiv preprint arXiv:2310.08459, 2024

arXiv 2024
[2]

A continual learningsurvey:Defyingforgettinginclassificationtasks,

M. D. Lange, R. Aljundi, M. Masana, S. Parisot, X. Jia, A. Leonardis, G. Slabaugh, and T. Tuytelaars, “A continual learningsurvey:Defyingforgettinginclassificationtasks,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, pp. 3366–3385, 2022

work page 2022
[3]

A comprehensive survey of continual learning: Theory, method and application,

L. Wang, X. Zhang, H. Su, and J. Zhu, “A comprehensive survey of continual learning: Theory, method and application,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–20, 2024

work page 2024
[4]

iCaRL: Incremental classifier and representation learning,

S.-A. Rebuﬀi, A. Kolesnikov, G. Sperl, and C. H. Lampert, “iCaRL: Incremental classifier and representation learning,” 2017 IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), pp. 5533–5542, 2017

work page 2017
[5]

Continual prototype evo- lution: Learning online from non-stationary data streams,

M. De Lange and T. Tuytelaars, “Continual prototype evo- lution: Learning online from non-stationary data streams,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 8230–8239

work page 2021
[6]

Learning multiple layers of features from tiny images,

A. Krizhevsky, “Learning multiple layers of features from tiny images,” Technical Report in Univer- sity of Toronto, pp. 32–33, 2009. [Online]. Avail- able: https://www.cs.toronto.edu/ kriz/learning-features-2009- TR.pdf

work page 2009
[7]

Tiny ImageNet,

M. A. mnmoustafa, “Tiny ImageNet,” 2017. [Online]. Available: https://kaggle.com/competitions/tiny-imagenet

work page 2017
[8]

Im- agenet large scale visual recognition challenge,

O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “Im- agenet large scale visual recognition challenge,” International Journal of Computer Vision (IJCV), vol. 115, no. 3, pp. 211– 252, 2015

work page 2015

Show all 26 references

[9]

Online continual learning with maximally interfered retrieval,

R. Aljundi, E. Belilovsky, T. Tuytelaars, L. Charlin, M. Caccia, M. Lin, and L. Page-Caccia, “Online continual learning with maximally interfered retrieval,” in International Conference on Neural Information Processing Systems, 2019

2019
[10]

Gdumb: A simple approach that questions our progress in continual learning,

A. Prabhu, P. Torr, and P. Dokania, “Gdumb: A simple approach that questions our progress in continual learning,” in The European Conference on Computer Vision (ECCV), 2020

2020
[11]

Rainbow memory: Continual learning with a memory of diverse samples,

J. Bang, H. Kim, Y. Yoo, J.-W. Ha, and J. Choi, “Rainbow memory: Continual learning with a memory of diverse samples,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, p. 8218–8227

2021
[12]

Online continual learning on class incremental blurry task configuration with anytime inference,

H. Koh, D. Kim, J.-W. Ha, and J. Choi, “Online continual learning on class incremental blurry task configuration with anytime inference,” in ICLR, 2022

2022
[13]

Online continual learning on hierarchical label expansion,

B. H. Lee, O. Jung, J. Choi, and S. Y. Chun, “Online continual learning on hierarchical label expansion,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 11761–11770

2023
[14]

FOSTER: Feature boosting and compression for class-incremental learn- ing,

F.-Y. Wang, D.-W. Zhou, H.-J. Ye, and D.-C. Zhan, “FOSTER: Feature boosting and compression for class-incremental learn- ing,” in European Conference on Computer Vision (ECCV), 2022

2022
[15]

Large scale incremental learning,

Y. Wu, Y. Chen, L. Wang, Y. Ye, Z. Liu, Y. Guo, and Y. Fu, “Large scale incremental learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019

2019
[16]

Riemannian walk for incremental learning: Understanding forgetting and intransigence,

A. Chaudhry, P. K. Dokania, T. Ajanthan, and P. H. S. Torr, “Riemannian walk for incremental learning: Understanding forgetting and intransigence,” in Proceedings of the European Conference on Computer Vision (ECCV), 2018

2018
[17]

Doubly perturbed task free continual learning,

B. H. Lee, M. hwan Oh, and S. Y. Chun, “Doubly perturbed task free continual learning,” AAAI, vol. 38, no. 12, pp. 13346– 13354, Mar. 2024

2024
[18]

Differential privacy,

C. Dwork, “Differential privacy,” in Automata, Languages and Programming, 2006, pp. 1–12

2006
[19]

Self- attention generative adversarial networks,

H. Zhang, I. Goodfellow, D. Metaxas, and A. Odena, “Self- attention generative adversarial networks,” in Proceedings of the 36th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, K. Chaudhuri and R. Salakhutdinov, Eds., vol. 97. PMLR,...

2019
[20]

Communication-Eﬀicient Learning of Deep Networks from Decentralized Data,

B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y. Arcas, “Communication-Eﬀicient Learning of Deep Networks from Decentralized Data,” in International Conference on Ar- tificial Intelligence and Statistics, A. Singh and J. Zhu, Eds., vol. 54. PMLR, 2017, pp. 1273–1282

2017
[21]

On the convergence of fedavg on non-iid data,

X. Li, K. Huang, W. Yang, S. Wang, and Z. Zhang, “On the convergence of fedavg on non-iid data,” in 8th International Conference on Learning Representations ICLR 2020, 2020

2020
[22]

Revisiting class-incremental learning with pre-trained models: Generalizabilityandadaptivityareallyouneed,

D.-W. Zhou, Z.-W. Cai, H.-J. Ye, D.-C. Zhan, and Z. Liu, “Revisiting class-incremental learning with pre-trained models: Generalizabilityandadaptivityareallyouneed,”inIJCV,2024

2024
[23]

First session adaptation: A strong replay-free baseline for class-incremental learning,

A. Panos, Y. Kobe, D. O. Reino, R. Aljundi, and R. E. Turner, “First session adaptation: A strong replay-free baseline for class-incremental learning,” in 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 18774– 18784

2023
[24]

Learning transferable visual models from natural language supervision,

A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” in Proceedings of the 38th International Conference on Machine L...

2021
[25]

Predicting the generalization gap in deep networks with margin distribu- tions,

Y. Jiang, D. Krishnan, H. Mobahi, and S. Bengio, “Predicting the generalization gap in deep networks with margin distribu- tions,” in ICLR, 2019

2019
[26]

Gradient episodic memory for continual learning,

D. Lopez-Paz and M. Ranzato, “Gradient episodic memory for continual learning,” in Proceedings of the 31st International Conference on Neural Information Processing Systems, ser. NIPS’17. Red Hook, NY, USA: Curran Associates Inc., 2017, p. 6470–6479

2017

Pith tools

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

[1] [1]

A recent survey of heterogeneous transfer learning,

R. Bao, Y. Sun, Y. Gao, J. Wang, Q. Yang, Z.-H. Mao, and Y. Ye, “A recent survey of heterogeneous transfer learning,” arXiv preprint arXiv:2310.08459, 2024

arXiv 2024

[2] [2]

A continual learningsurvey:Defyingforgettinginclassificationtasks,

M. D. Lange, R. Aljundi, M. Masana, S. Parisot, X. Jia, A. Leonardis, G. Slabaugh, and T. Tuytelaars, “A continual learningsurvey:Defyingforgettinginclassificationtasks,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, pp. 3366–3385, 2022

work page 2022

[3] [3]

A comprehensive survey of continual learning: Theory, method and application,

L. Wang, X. Zhang, H. Su, and J. Zhu, “A comprehensive survey of continual learning: Theory, method and application,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–20, 2024

work page 2024

[4] [4]

iCaRL: Incremental classifier and representation learning,

S.-A. Rebuﬀi, A. Kolesnikov, G. Sperl, and C. H. Lampert, “iCaRL: Incremental classifier and representation learning,” 2017 IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), pp. 5533–5542, 2017

work page 2017

[5] [5]

Continual prototype evo- lution: Learning online from non-stationary data streams,

M. De Lange and T. Tuytelaars, “Continual prototype evo- lution: Learning online from non-stationary data streams,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 8230–8239

work page 2021

[6] [6]

Learning multiple layers of features from tiny images,

A. Krizhevsky, “Learning multiple layers of features from tiny images,” Technical Report in Univer- sity of Toronto, pp. 32–33, 2009. [Online]. Avail- able: https://www.cs.toronto.edu/ kriz/learning-features-2009- TR.pdf

work page 2009

[7] [7]

Tiny ImageNet,

M. A. mnmoustafa, “Tiny ImageNet,” 2017. [Online]. Available: https://kaggle.com/competitions/tiny-imagenet

work page 2017

[8] [8]

Im- agenet large scale visual recognition challenge,

O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “Im- agenet large scale visual recognition challenge,” International Journal of Computer Vision (IJCV), vol. 115, no. 3, pp. 211– 252, 2015

work page 2015