Pith. sign in

REVIEW 5 major objections 5 minor 67 references

Noise-Tolerant Coreset-Based Class Incremental Continual Learning

T0 review · 5 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Coreset-based replay buffers let class-incremental learners keep accuracy and cut forgetting under label and instance noise, with a new training-time guarantee for perturbed inputs.

desk verdict The empirical part is worth a look, but the central instance-noise theorem is unproven—Eq. (7) is false for the nonlinear networks they actually train. read the letter →

arxiv 2504.16763 v1 pith:PLIO37BT submitted 2025-04-23 cs.LG cs.AIcs.CVcs.NE

Edison Mucllari , Aswin Raghavan , Zachary Alan Daniels This is my paper · ORCID
classification cs.LGcs.AIcs.CVcs.NE
keywords class-incrementallearningcontinualcoresetsCRUSTlabelnoiseinstancereplaybuffercatastrophicforgetting
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

This paper claims that memory-based class-incremental learners, which replay old examples from a stored buffer, can be made noise-tolerant by building the buffer with gradient-selected Coresets. It derives a new convergence bound showing that a Coreset-trained network still reaches correct classification when a fraction of the training inputs are additively perturbed, and that degradation is gradual in the noise fraction rather than abrupt. On that basis it introduces Continual CRUST and Continual CosineCRUST, two replay-buffer methods for class-incremental learning, and reports that they hold accuracy and reduce forgetting compared with existing memory-based learners under label flipping and salt-and-pepper instance noise across five datasets. If correct, continual learning systems could tolerate failing sensors and imperfect labeling without a separate denoising stage.

What carries the argument

The load-bearing object is the Jacobian of the network's loss with respect to its parameters, evaluated on the selected Coreset. CRUST is a greedy submodular Coreset-selection procedure that picks replay samples whose gradients best approximate the full gradient matrix, exploiting the fact that clean samples cluster in the Jacobian spectrum. The new theory extends the standard gradient-descent residual recursion to perturbed inputs by decomposing the Jacobian at $X+EX$ additively into its value at $X$ and its value at $EX$ (equations 7-10), producing error terms $E_{J_1}$ and $E_{J_2}$ that enter the final iteration bound. The proposed algorithms operationalize this by keeping one Coreset per class, refining each Coreset after every experience with CRUST, and optionally adding a spectral-clustering step on cosine distance before submodular selection in Continual CosineCRUST.

What would settle it

Take a ResNet18 on CIFAR10, add salt-and-pepper noise to a fraction of inputs, and numerically check equation (7): compute $J(W, X+EX)^T$ and compare it with $J(W, X)^T + J(W, EX)^T$ at the same weights. If the norm of the difference is on the same order as the $E_{J_1}$ and $E_{J_2}$ terms driving the bound, then the residual decay predicted by Lemma 14 does not hold for that network, and the theorem's iteration guarantee would not be observed in practice.

Watch

Extended reading notes

Core claim

The paper's central claim is that Coreset replay, originally designed for label-noise robustness in static training, extends to class-incremental learning under both label noise and additive input noise. The theoretical engine is a new bound (Theorem 4): for a fraction $\delta$ of training inputs perturbed as $\tilde{X} = X + E X$, gradient descent with MSE loss classifies all Coreset samples correctly after $T \geq O\big((1/\eta)\big(\alpha/2 + E_{J_2}\alpha + E_{J_1}\alpha - \eta E_{J_2}\beta^{3/2} - \eta E_{J_1}\beta/3\big)^{-1}\log(\|r_0\|^2/\nu)\big)$ iterations, assuming the Coreset approximates the Jacobian within $\epsilon \leq O(\delta\alpha^2/(k\beta\log\delta))$. The qualitative message is that instance noise degrades accuracy gradually with $\delta$, at a slower rate than label noise, and requires more training iterations as clean data become scarcer. Building on this, the paper maintains per-class Coresets across experiences, refining each by CRUST's greedy submodular selection, and adds a variant that first clusters gradients by cosine distance and filters out small clusters (Continual CosineCRUST). Across MNIST, FashionMNIST, CIFAR10, MSTAR, and PathMNIST+, the proposed methods report higher final accuracy and lower forgetting than iCaRL, Dark ER, and replay-based baselines under label flipping up to 0.5 and salt-and-pepper instance noise up to 0.8.

Load-bearing premise

The proof assumes that the sensitivity of the network's outputs to its parameters at a noisy input is exactly the sensitivity at the clean input plus the sensitivity at the noise alone, and that the same split holds for the path-averaged sensitivity; this is true only for a network that is a straight-line function of its input, which the experimental networks are not.

Editorial extensions

If this is right

  • A continual learner can store only a small per-class Coreset and still maintain accuracy when label flipping probability reaches 0.5, whereas random replay, iCaRL, Dark ER, and Replay+EWC degrade sharply in the reported comparisons.
  • Under salt-and-pepper instance noise with up to 80 percent of samples perturbed, the proposed methods keep the forgetting metric below 0.1 on FashionMNIST, about half the forgetting of the strongest replay baseline.
  • Coreset purity stays above 90 percent up to 30 percent label noise and above 75 percent at 50 percent label noise, which the authors note satisfies the purity condition required by the prior label-noise theorem.
  • The bound predicts that the number of iterations needed to fit the Coreset decreases as the noise fraction grows, so noisy class-incremental settings may train faster while still fitting the Coreset, though with a less representative buffer.

Reading between the lines

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

  • A direct test the authors do not run: since the additive Jacobian decomposition is exact only for affine networks, the theorem's quantitative iteration count should first be checked on a linear model, where the assumption holds, before being expected to transfer to deep networks.
  • The CosineCRUST clustering step is essentially outlier detection in gradient space; the same mechanism could transfer to other settings where per-class or per-client gradient clusters need cleaning, such as noisy federated learning, though the paper offers no evidence for that extension.
  • The authors' own caveat that clustering can split data into two dense clusters, one clean and one noisy, suggests a concrete mitigation: restrict the submodular Coreset selection to the clean cluster or weight samples by cluster purity; this variant is not tested in the paper.
  • Because the bound expresses robustness in terms of the Coreset's Jacobian approximation error, it implies that any future replay method that better approximates the Jacobian under perturbation should inherit a stronger instance-noise guarantee, a connection the paper leaves implicit.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper addresses class-incremental continual learning (CIL) under label noise and uncorrelated instance noise. It proposes two replay-buffer construction methods, Continual CRUST and Continual CosineCRUST, built on the CRUST coreset selection algorithm. The central theoretical claim is Theorem 4, which asserts an iteration bound and correct-classification guarantee for gradient descent with MSE loss on a dataset with a delta fraction of perturbed instances, under a stated condition on the Jacobian approximation error epsilon. The empirical section compares the proposed methods against memory-based baselines (Random Replay, Random Replay with EWC, Dark ER, iCaRL) on MNIST, FashionMNIST, CIFAR10, MSTAR, and PathMNIST+ under label-flipping and salt-and-pepper instance noise, reporting accuracy, forgetting, and coreset purity. The supplementary material contains the proof of Theorem 4 as Lemmas 7-15 and Theorem 16.

Significance. If Theorem 4 were correct, it would provide a quantitative robustness guarantee for coreset-based replay under instance noise, including a comparison of the hardness of label versus instance noise and a bound on the required number of training iterations. The empirical study is broad: five datasets, two noise types, multiple noise levels, five seeds, and a statistical significance test, and the authors honestly record limitations of their method (e.g., the worst-case two-cluster scenario for CosineCRUST and the lack of adversarial-noise experiments). However, the theoretical contribution is the stated core of the paper, and it is not supported by the supplied proof: the key additive Jacobian identity is false for the nonlinear networks used, the perturbation terms EJ1 and EJ2 are never bounded, and the epsilon hypothesis of Theorem 4 is vacuous as written owing to a sign error. The empirical results, while suggestive, do not validate Theorem 4 because the experiments use cross-entropy loss and Adam rather than the MSE/gradient-descent setting of the theorem. Given that the load-bearing theoretical claims collapse, the paper in its current form cannot be accepted.

major comments (5)
  1. [Section 3, Eq. (7)] Equation (7) asserts J(W,X+EX)^T = J(W,X)^T + J(W,EX)^T and defines EJ1 by this decomposition; Eq. (9) makes the same additive assumption for the average Jacobian and defines EJ2. This identity is false for any nonlinear network: the Jacobian of the network output with respect to the weights is not an affine function of the input, so the Jacobian evaluated at X+EX is not the sum of the Jacobians at X and EX. The experiments use ResNet18 and EfficientNetV2s (Table 4), which are nonlinear. Because Lemmas 10, 11, and 14 of the supplement all inherit this decomposition, the residual recursion leading to Eq. (47) and the final iteration bound in Theorem 4 have no valid derivation.
  2. [Supplement, Lemma 10] Lemma 10 states that the update takes the form rhat approx (I - eta(C(W) + EJ2 J(W)^T + EJ1 Jbar(What,W)))r and then says 'Numerical analysis suggests we can ignore the second order error term EJ2 EJ1.' Dropping a term without bounding it is not a proof step. The product of the two perturbation terms is of the same formal order as the kept cross terms, and no bound in terms of delta or the noise amplitude is given for any of EJ1 or EJ2. The subsequent Lemma 14 and Eq. (47) therefore rest on an unproved approximation rather than an inequality.
  3. [Theorem 4, epsilon hypothesis] The theorem states the hypothesis epsilon <= O(delta alpha^2 / (k beta log(delta))). For a fraction delta of perturbed samples with delta < 1, log(delta) is negative, so the right-hand side is negative, making the hypothesis unsatisfiable for any positive epsilon. This appears to be a sign error (log(1/delta) was likely intended), but as written the condition is vacuous and the theorem cannot hold for any instance-noise level delta in (0,1).
  4. [Theorem 4 and Supplement Theorem 16, contraction factor] The contraction factor in Eq. (47) and in Lemma 15 is written as 1 - eta(alpha/2 + EJ2 alpha + EJ1 alpha - eta EJ2 beta^3/2 - eta EJ1 beta/3), with EJ1 and EJ2 appearing with mixed signs. No bounds for EJ1 and EJ2 in terms of delta, the noise amplitude, or the problem dimensions are ever established; consequently the positivity of the contraction factor, and hence the stated iteration bound, is not guaranteed. Additionally, the proof of Theorem 16 asserts the epsilon bound by saying one can 'follow the steps outlined in [36]' and reinterpret delta as the image-perturbation fraction; the label-noise proof in [36] relies on label margin and coreset purity, for which no instance-noise analogues are provided, so the derivation of the epsilon condition is absent.
  5. [Section 2.1 and Section 5; Supplement Sections 10.2, 10.6] Theorem 4 assumes MSE loss and vanilla gradient descent, but the experiments use cross-entropy loss with the Adam optimizer (e.g., Supplement Sections 10.2 and 10.6), and the main text states only that labels are treated as ordinal values in the theory. No argument is given to transfer the theorem to the cross-entropy/Adam setting, and no experiment uses MSE loss with gradient descent. The empirical validation therefore does not test Theorem 4's assumptions, and the claim that the algorithms are 'theoretically-sound' in the experimental setting is unsupported.
minor comments (5)
  1. [Section 5.7, Table 3] The caption says '(top)' and '(bottom)' but the table shown contains only one set of rows; the reference to a bottom half for uniform random noise is confusing and should be corrected to match the actual table content.
  2. [Supplement, Section 9] The sentence 'Not that Coreset size refers to...' should read 'Note that Coreset size refers to...'.
  3. [Section 2.2, Eq. (6)] The marginal utility Fhat(e|S_t) = F(S_t union {e}) - F(S_t) is written with S_t in the definition but the sup over e in V is over the full dataset; the notation should specify that the union is with the current set S_{t-1} to avoid confusion with the set S_t being defined.
  4. [Section 5.6] The phrase 'prior work on CL shows only a small degradations in performance' should be 'small degradation'; also 'the roughly average 0.2 achieved by DER' would be clearer as 'the average forgetting of roughly 0.2 achieved by DER'.
  5. [Theorem 4, notation] The theorem statement uses alpha = sqrt(rmin) sigma_min(J(W,X_S)) - Emin and beta = ||J(W,X)||_2 + epsilon + Emax, but Emin and Emax are defined only later in the supplement proof; a forward reference to Definition 2 or to the supplement would improve readability.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 4's robustness bound is written in terms of the perturbation Jacobians EJ1/EJ2, which are never bounded in terms of δ; the advertised δ-dependent guarantee reduces to the definitions in Eqs. (7)-(10).

  1. self definitional [Section 3, Eqs. (7)-(10) and Theorem 4; supplementary Lemma 14 / Eq. (47)]
    "The following terms EJ1 and EJ2 appear in the bound. They are associated with the Jacobian of the NN loss evaluated at the perturbation. For the Jacobian, we have J (W,X +EX)T =J (W,X )T +J (W,EX)T (7) =J (W,X )T +EJ1 (8) ... ¯J ( ˆW,W,X +EX) = ¯J ( ˆW,W,X ) + ¯J ( ˆW,W,E X) (9) = ¯J ( ˆW,W,X ) +EJ2 (10)"

    Theorem 4's iteration bound is T≥O(1/η (α/2 + EJ2α + EJ1α − ηEJ2β^3/2 − ηEJ1β/3)^−1 log(||r0||^2/ν)). The only perturbation-dependent quantities in this bound are EJ1 and EJ2, which are defined in Eqs. (8) and (10) as the Jacobian / average-Jacobian of the perturbation EX. No lemma in the paper bounds EJ1 or EJ2 by δ or by the noise amplitude; the proof of Theorem 16 simply solves the residual recurrence of Lemma 14 after substituting these definitions into the clean-data recurrence. The convergence conclusion is therefore equivalent to assuming that the perturbed-Jacobian contraction factor is favorable, which is exactly the object the theorem was supposed to control.

full rationale

The only substantial circularity is in the theoretical 'Robustness to Data Perturbations' result. Theorem 4 promises a δ-dependent iteration guarantee for CRUST under additive instance noise, but its bound is expressed through EJ1 and EJ2, defined in Eqs. (7)-(10) as the Jacobian and average-Jacobian of the perturbation. No step connects these terms to δ or to the noise amplitude; the proof reduces to substituting the definitions into the old clean-data recurrence and solving for T. Thus the central theoretical contribution is close to tautological: it restates the residual contraction in terms of the very perturbed-Jacobian quantities it needed to control. The empirical part is different: Continual CRUST and Continual CosineCRUST are evaluated against external baselines on standard benchmarks, with no fitted parameter relabeled as a prediction and no load-bearing self-citation chain. The experimental results stand independently, so the circularity is partial rather than total. Score 6 reflects that one central claimed prediction reduces by construction, while the empirical core remains self-contained.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central theoretical contribution depends on the additive Jacobian decomposition and unquantified perturbation matrices EJ1 and EJ2. The empirical contribution depends on several unreported hyperparameters and the untested transfer from MSE theory to cross-entropy training. No new physical entities are postulated.

free parameters (3)
  • nA (minimum Coreset size / cluster size threshold) = not reported
    Algorithm 1 selects clusters only if size > nA; no value or ablation is listed in Table 4 or the text, so the behavior of CosineCRUST depends on an unreported hand choice.
  • spectral clustering settings (number of clusters, affinity, etc.) = not reported
    CosineCRUST's clustering step in Algorithm 1 lines 20-24 is not parameterized in the paper; different settings would change which samples are kept.
  • per-strategy learning rates = varies by dataset and strategy, e.g., MNIST 0.0007-0.005, FashionMNIST 0.001-0.03
    Learning rates are tuned separately for each method on one seed, which is a fitting choice that affects the baseline comparisons.
assumptions (6)
  • ad hoc to paper Additive decomposition of the Jacobian: J(W,X+EX)^T = J(W,X)^T + J(W,EX)^T
    Eq (7)-(8) introduce EJ1 through this equality, but it is false for nonlinear network maps; the proof of Theorem 4 depends on it.
  • ad hoc to paper Average Jacobian decomposes additively in the input perturbation: Jbar(What,W,X+EX) = Jbar(What,W,X) + Jbar(What,W,EX)
    Eq (9)-(10) and Lemma 10 use this to define EJ2; it has the same nonlinearity problem as Eq (7).
  • standard math The Jacobian is L-smooth and has a spectral gap with alpha and beta bounds
    Adopted from CRUST [36] and [32]; reasonable for the analysis but not verified for the networks used in the experiments.
  • domain assumption MSE convergence bounds transfer to cross-entropy-trained classifiers
    Section 2.1 limits the theory to MSE and ordinal labels, while experiments use cross-entropy and categorical labels; no transfer argument is given.
  • domain assumption Noisy examples form small outlier clusters in cosine-Jacobian space
    Section 4 states this as a hypothesis for CosineCRUST; the paper's own limitations concede the clustering can instead find two dense clean and noisy clusters and pollute the coreset.
  • domain assumption Label margin delta >= 1 so that 8*rho <= delta holds at 90% coreset purity
    Section 5.7 claims Theorem 1's requirement is met because delta >= 1; the label margin is a dataset-specific constant from [32] and is not measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Noise-Tolerant Coreset-Based Class Incremental Continual Learning." pith.science (2026). https://pith.science/paper/PLIO37BT

@misc{pith2026250416763,
  author       = {Pith},
  title        = {Pith review of: Noise-Tolerant Coreset-Based Class Incremental Continual Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PLIO37BT}},
  note         = {Machine review of arXiv:2504.16763}
}
read the original abstract

Many applications of computer vision require the ability to adapt to novel data distributions after deployment. Adaptation requires algorithms capable of continual learning (CL). Continual learners must be plastic to adapt to novel tasks while minimizing forgetting of previous tasks.However, CL opens up avenues for noise to enter the training pipeline and disrupt the CL. This work focuses on label noise and instance noise in the context of class-incremental learning (CIL), where new classes are added to a classifier over time, and there is no access to external data from past classes. We aim to understand the sensitivity of CL methods that work by replaying items from a memory constructed using the idea of Coresets. We derive a new bound for the robustness of such a method to uncorrelated instance noise under a general additive noise threat model, revealing several insights. Putting the theory into practice, we create two continual learning algorithms to construct noise-tolerant replay buffers. We empirically compare the effectiveness of prior memory-based continual learners and the proposed algorithms under label and uncorrelated instance noise on five diverse datasets. We show that existing memory-based CL are not robust whereas the proposed methods exhibit significant improvements in maximizing classification accuracy and minimizing forgetting in the noisy CIL setting.

Figures

Figures reproduced from arXiv: 2504.16763 by the authors.

Figure 1
Figure 1. Representative samples from the datasets used for eval [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Looking at the final accuracy at different label flipping noise levels (0.0-0.5) for five benchmark datasets for different strategies; [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Left: Average Coreset purity over all experiences on PathMNIST+ dataset for different levels of label noise. Right: Examining [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Looking at the final accuracy (left) and forgetting metric (right) at different Coreset sizes (100, 200, and 300 samples per [PITH_FULL_IMAGE:figures/full_fig_p024_4.png]
Figure 5
Figure 5. Figure 5: Looking at the final accuracy (left) and forgetting metric (right) at different label flipping noise levels (0.0-0.5) for MNIST for [PITH_FULL_IMAGE:figures/full_fig_p028_5.png]
Figure 6
Figure 6. Figure 6: Looking at the final accuracy (left) and forgetting metric (right) at different label flipping noise levels (0.0-0.5) for FashionMNIST [PITH_FULL_IMAGE:figures/full_fig_p030_6.png]
Figure 7
Figure 7. Figure 7: Looking at the final accuracy (left) and forgetting metric (right) at different label flipping noise levels (0.0-0.5) for CIFAR10 for [PITH_FULL_IMAGE:figures/full_fig_p031_7.png]
Figure 8
Figure 8. Figure 8: Looking at the final accuracy (left) and forgetting metric (right) at different label flipping noise levels (0.0-0.5) for MSTAR for [PITH_FULL_IMAGE:figures/full_fig_p034_8.png]
Figure 9
Figure 9. Figure 9: Looking at the final accuracy (left) and forgetting metric (right) at different label flipping noise levels (0.0-0.5) for PathMNIST+ [PITH_FULL_IMAGE:figures/full_fig_p036_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

67 extracted references · 50 canonical work pages

  1. [36]

    Coresets for robust training of deep neural networks against noisy labels

    Baharan Mirzasoleiman, Kaidi Cao, and Jure Leskovec. Coresets for robust training of deep neural networks against noisy labels. Advances in Neural Information Processing Systems, 33:11465–11477, 2020. 1, 2, 3, 4, 8, 10, 11

    work page 2020

  2. [1]

    Brainwash: A poisoning attack to forget in continual learning

    Ali Abbasi, Parsa Nooralinejad, Hamed Pirsiavash, and So- heil Kolouri. Brainwash: A poisoning attack to forget in continual learning. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 24057–24067, 2024. 1, 8

    work page 2024

  3. [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. 7

    work page 2018

  4. [3]

    Gradient based sample selection for online continual learning

    Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Ben- gio. Gradient based sample selection for online continual learning. Advances in neural information processing sys- tems, 32, 2019. 7, 8

    work page 2019

  5. [4]

    Tackling Online One-Class Incremental Learning by Removing Negative Contrasts

    Nader Asadi, Sudhir Mudur, and Eugene Belilovsky. Tack- ling online one-class incremental learning by removing neg- ative contrasts. arXiv preprint arXiv:2203.13307, 2022. 4

    work page Pith review arXiv 2022

  6. [5]

    Towards adversarially robust continual learning

    Tao Bai, Chen Chen, Lingjuan Lyu, Jun Zhao, and Bi- han Wen. Towards adversarially robust continual learning. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 1–5. IEEE, 2023. 8

    work page 2023

  7. [6]

    Coresets via bilevel optimization for continual learning and stream- ing

    Zal ´an Borsos, Mojmir Mutny, and Andreas Krause. Coresets via bilevel optimization for continual learning and stream- ing. Advances in neural information processing systems, 33: 14879–14890, 2020. 8

    work page 2020

  8. [7]

    Dark experience for gen- eral continual learning: a strong, simple baseline

    Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for gen- eral continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920–15930,

Show all 67 references
  1. [8]

    Avalanche: A pytorch li- brary for deep continual learning.Journal of Machine Learn- ing Research, 24(363):1–6, 2023

    Antonio Carta, Lorenzo Pellegrini, Andrea Cossu, Hamed Hemati, and Vincenzo Lomonaco. Avalanche: A pytorch li- brary for deep continual learning.Journal of Machine Learn- ing Research, 24(363):1–6, 2023. 5

    2023

  2. [9]

    Lifelong machine learning

    Zhiyuan Chen and Bing Liu. Lifelong machine learning. Synthesis Lectures on AI and ML, 2018. 7

    2018

  3. [10]

    Lifelong machine learning

    Zhiyuan Chen and Bing Liu. Lifelong machine learning . Springer Nature, 2022. 1

    2022

  4. [11]

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

    Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ale ˇs Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying for- getting in classification tasks. IEEE TPAMI, 2021. 7

    2021

  5. [12]

    Don’t forget, there is more than for- getting: new metrics for continual learning

    Natalia D ´ıaz-Rodr´ıguez, Vincenzo Lomonaco, David Filliat, and Davide Maltoni. Don’t forget, there is more than for- getting: new metrics for continual learning. arXiv preprint arXiv:1810.13166, 2018. 5

    2018 arXiv

  6. [13]

    Progressive learning: A deep learning framework for contin- ual learning

    Haytham M Fayek, Lawrence Cavedon, and Hong Ren Wu. Progressive learning: A deep learning framework for contin- ual learning. Neural Networks, 128, 2020. 7

    2020

  7. [14]

    Persis- tent backdoor attacks in continual learning

    Zhen Guo, Abhinav Kumar, and Reza Tourani. Persis- tent backdoor attacks in continual learning. arXiv preprint arXiv:2409.13864, 2024. 8

    2024 arXiv

  8. [15]

    Remind your neural net- work to prevent catastrophic forgetting

    Tyler L Hayes, Kushal Kafle, Robik Shrestha, Manoj Acharya, and Christopher Kanan. Remind your neural net- work to prevent catastrophic forgetting. In European con- ference on computer vision, pages 466–483. Springer, 2020. 8

    2020

  9. [16]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 19, 21

    2016

  10. [17]

    Compacting, picking and growing for unforgetting continual learning

    Ching-Yi Hung, Cheng-Hao Tu, Cheng-En Wu, Chien-Hung Chen, Yi-Ming Chan, and Chu-Song Chen. Compacting, picking and growing for unforgetting continual learning. NeurIPS, 2019. 7

    2019

  11. [18]

    Selective experience re- play for lifelong learning

    David Isele and Akansel Cosgun. Selective experience re- play for lifelong learning. In Proceedings of the AAAI Con- ference on Artificial Intelligence, 2018. 7, 8

    2018

  12. [19]

    Robustness-preserving lifelong learning via dataset condensation

    Jinghan Jia, Yihua Zhang, Dogyoon Song, Sijia Liu, and Alfred Hero. Robustness-preserving lifelong learning via dataset condensation. In ICASSP 2023-2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Process- ing (ICASSP), pages 1–5. IEEE, 2023. 8

    2023

  13. [20]

    Continual poi- soning of generative models to promote catastrophic forget- ting

    Siteng Kang, Zhan Shi, and Xinhua Zhang. Continual poi- soning of generative models to promote catastrophic forget- ting. In NeurIPS ML Safety Workshop, 2022. 8

    2022

  14. [21]

    Poisoning gen- erative replay in continual learning to promote forgetting

    Siteng Kang, Zhan Shi, and Xinhua Zhang. Poisoning gen- erative replay in continual learning to promote forgetting. In International Conference on Machine Learning , pages 15769–15785. PMLR, 2023. 8

    2023

  15. [22]

    Not all sam- ples are created equal: Deep learning with importance sam- pling

    Angelos Katharopoulos and Franc ¸ois Fleuret. Not all sam- ples are created equal: Deep learning with importance sam- pling. In International conference on machine learning , pages 2525–2534. PMLR, 2018. 4

    2018

  16. [23]

    Mstar extended operating conditions: A tutorial

    Eric R Keydel, Shung Wu Lee, and John T Moore. Mstar extended operating conditions: A tutorial. Algorithms for Synthetic Aperture Radar Imagery III, 2757:228–242, 1996. 5

    1996

  17. [24]

    Adversarially robust continual learning

    Hikmat Khan, Nidhal Carla Bouaynaya, and Ghulam Rasool. Adversarially robust continual learning. In 2022 Interna- tional Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2022. 8

    2022

  18. [25]

    Knowledge transfer in lifelong machine learning: a system- atic literature review

    Pouya Khodaee, Herna L Viktor, and Wojtek Michalowski. Knowledge transfer in lifelong machine learning: a system- atic literature review. Artificial Intelligence Review , 57(8): 217, 2024. 1

    2024

  19. [26]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980, 2014. 15, 19, 21, 22

    2014 arXiv

  20. [27]

    Overcoming catastrophic forgetting in neu- ral networks

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska- Barwinska, et al. Overcoming catastrophic forgetting in neu- ral networks. Proceedings of the national academy of sc...

    2017

  21. [28]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5

    2009

  22. [29]

    Y . LECUN. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/. 5

  23. [30]

    Targeted data poisoning at- tacks against continual learning neural networks

    Huayu Li and Gregory Ditzler. Targeted data poisoning at- tacks against continual learning neural networks. In2022 In- ternational Joint Conference on Neural Networks (IJCNN) , pages 1–8. IEEE, 2022. 8 9

    2022

  24. [31]

    Pacol: Poisoning attacks against continual learners

    Huayu Li and Gregory Ditzler. Pacol: Poisoning attacks against continual learners. arXiv preprint arXiv:2311.10919,

    arXiv

  25. [32]

    Gradient descent with early stopping is provably robust to label noise for overparameterized neural networks

    Mingchen Li, Mahdi Soltanolkotabi, and Samet Oymak. Gradient descent with early stopping is provably robust to label noise for overparameterized neural networks. In Inter- national conference on artificial intelligence and statistics , pages 4313–4324. PMLR, 2020. 2, 3, 5, 6

    2020

  26. [33]

    Learn to grow: A continual structure learn- ing framework for overcoming catastrophic forgetting

    Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to grow: A continual structure learn- ing framework for overcoming catastrophic forgetting. In ICML. PMLR, 2019. 7

    2019

  27. [34]

    Avalanche: an end-to-end library for contin- ual learning

    Vincenzo Lomonaco, Lorenzo Pellegrini, Andrea Cossu, An- tonio Carta, Gabriele Graffieti, Tyler L Hayes, Matthias De Lange, Marc Masana, Jary Pomponi, Gido M Van de Ven, et al. Avalanche: an end-to-end library for contin- ual learning. In Proceedings of the IEEE/CVF Conference...

    2021

  28. [35]

    Gradient episodic memory for continual learning

    David Lopez-Paz and Marc’Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neu- ral information processing systems, 30, 2017. 5, 7, 8

    2017

  29. [37]

    Coresets- methods and history: A theoreticians design pattern for ap- proximation and streaming algorithms

    Alexander Munteanu and Chris Schwiegelshohn. Coresets- methods and history: A theoreticians design pattern for ap- proximation and streaming algorithms. KI-K¨unstliche Intel- ligenz, 32:37–53, 2018. 2

    2018

  30. [38]

    Lifelong learning metrics

    Alexander New, Megan Baker, Eric Nguyen, and Gau- tam Vallabha. Lifelong learning metrics. arXiv preprint arXiv:2201.08278, 2022. 5

    2022 arXiv

  31. [39]

    Latent replay for real-time continual learning

    Lorenzo Pellegrini, Gabriele Graffieti, Vincenzo Lomonaco, and Davide Maltoni. Latent replay for real-time continual learning. In 2020 IEEE/RSJ International Conference on In- telligent Robots and Systems (IROS) , pages 10203–10209. IEEE, 2020. 8

    2020

  32. [40]

    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. 5, 7, 8

    2001

  33. [41]

    Experience replay for continual learning

    David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lil- licrap, and Gregory Wayne. Experience replay for continual learning. Advances in neural information processing sys- tems, 32, 2019. 7

    2019

  34. [42]

    Main- taining adversarial robustness in continuous learning

    Xiaolei Ru, Xiaowei Cao, Zijia Liu, Jack Murdoch Moore, Xin-Ya Zhang, Xia Zhu, Wenjia Wei, and Gang Yan. Main- taining adversarial robustness in continuous learning. arXiv preprint arXiv:2402.11196, 2024. 8

    2024 arXiv

  35. [43]

    Progressive neural networks

    Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Raz- van Pascanu, and Raia Hadsell. Progressive neural networks. arXiv:1606.04671, 2016. 7

    2016 arXiv

  36. [44]

    Hidden trigger backdoor attacks

    Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pir- siavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI conference on artificial intelligence, pages 11957– 11965, 2020. 1

    2020

  37. [45]

    Prioritized experience replay

    Tom Schaul. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015. 8

    2015 arXiv

  38. [46]

    Continual learning with deep generative replay

    Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 8

    2017

  39. [47]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML. PMLR,

  40. [48]

    Lifelong learning algorithms

    Sebastian Thrun. Lifelong learning algorithms. In Learning to learn, pages 181–209. Springer, 1998. 1

    1998

  41. [49]

    Gcr: Gradient coreset based replay buffer selection for continual learning

    Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 99–108, 2022. 8

    2022

  42. [50]

    Adversarial targeted forgetting in regularization and generative based continual learning models

    Muhammad Umer and Robi Polikar. Adversarial targeted forgetting in regularization and generative based continual learning models. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2021. 8

    2021

  43. [51]

    Adversary aware con- tinual learning

    Muhammad Umer and Robi Polikar. Adversary aware con- tinual learning. IEEE Access, 2024. 8

    2024

  44. [52]

    Tar- geted forgetting and false memory formation in continual learners through adversarial backdoor attacks

    Muhammad Umer, Glenn Dawson, and Robi Polikar. Tar- geted forgetting and false memory formation in continual learners through adversarial backdoor attacks. In 2020 In- ternational Joint Conference on Neural Networks (IJCNN) , pages 1–8. IEEE, 2020. 1

    2020

  45. [53]

    Three scenar- ios for continual learning

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

    1904 arXiv

  46. [54]

    Three types of incremental learning

    Gido M van de Ven, Tinne Tuytelaars, and Andreas S To- lias. Three types of incremental learning. Nature Machine Intelligence, 2022. 7

    2022

  47. [55]

    Prioritized generative replay

    Renhao Wang, Kevin Frans, Pieter Abbeel, Sergey Levine, and Alexei A Efros. Prioritized generative replay. arXiv preprint arXiv:2410.18082, 2024. 8

    2024 arXiv

  48. [56]

    Metamix: Towards corruption-robust continual learning with tempo- rally self-adaptive data transformation

    Zhenyi Wang, Li Shen, Donglin Zhan, Qiuling Suo, Yan- jun Zhu, Tiehang Duan, and Mingchen Gao. Metamix: Towards corruption-robust continual learning with tempo- rally self-adaptive data transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Re...

    2023

  49. [57]

    Fashion- mnist: a novel image dataset for benchmarking machine learning algorithms

    Han Xiao, Kashif Rasul, and Roland V ollgraf. Fashion- mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 5

    2017 arXiv

  50. [58]

    Medm- nist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification

    Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. Medm- nist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data, 10(1):41,

  51. [59]

    Lifelong learning with dynamically expandable net- works

    Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable net- works. arXiv:1708.01547, 2017. 7

    2017 arXiv

  52. [60]

    Online coreset selection for rehearsal-based contin- 10 ual learning

    Jaehong Yoon, Divyam Madaan, Eunho Yang, and Sung Ju Hwang. Online coreset selection for rehearsal-based contin- 10 ual learning. In International Conference on Learning Rep- resentations, 2021. 8

    2021

  53. [61]

    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,

  54. [63]

    We first need to define subspacesS+ andS−

    Proof of Theorem 4 As it is mentioned in [36], the Jacobian spectrum can be split into information spaceI and nuisance spaceN , associated with the large and small singular values. We first need to define subspacesS+ andS−. First we need: Definition 5. In our following analysi...

  55. [64]

    At (c) we combine what we have at (b) and use the fact thatrT TJ (WT )J (WT )TrT = ∥J (WT )TrT∥2

  56. [65]

    At (e) we use the fact that∥J (WT )∥2≥α and so−∥J (WT )∥2≤−α

    At (d) we use the fact that η≤ 1 2β2 and plug inη at (c). At (e) we use the fact that∥J (WT )∥2≥α and so−∥J (WT )∥2≤−α. In the next Lemma, we are going to introduce the perturbed dataset: Lemma 14. By following the results from Lemma 10 prove that, ∥rt+1∥2 2 ⪅ (1−η(α 2 + 2EJ2α...

  57. [66]

    splitMNIST

    Experimental parameters In the following table, we define the hyperparameters used for the different datasets in our experiments. Not that Coreset size refers to the number of samples stored in the individual per-class Coresets. Note that in all experiments, we use five random...

  58. [67]

    Additional Experiments In this section, we present additional ablations and more fine-grained experimental data that support the results reported in the main text. Specifically, we look at the ef- fect of Coreset size, report full results over all data sets for label flipping ...

  59. [2017]

    7 11 Noise-Tolerant Coreset-Based Class Incremental Continual Learning Supplementary Material

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.