Relative Parameter Importance in Task-Agnostic Replay-Free Continual Learning

Ikechukwu Nkisi-Orji; Malavika Suresh; Nirmalie Wiratunga

REVIEW 4 major objections 4 minor 46 references

Relative Parameter Importance in Task-Agnostic Replay-Free Continual Learning

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A look-ahead relative-parameter-importance regularizer gives small overall-F1 gains in replay-free, task-agnostic continual text classification, but the gains are noisy and threshold-dependent.

desk verdict Genuinely new relative-importance ratio for replay-free CL, but the reported gains are not statistically established — worth a serious referee. read the letter →

arxiv 2608.00630 v1 pith:CCQTAR2K submitted 2026-08-01 cs.LG

Malavika Suresh , Ikechukwu Nkisi-Orji , Nirmalie Wiratunga This is my paper

classification cs.LG

keywords importancelearningrelativeparametersaccesscontinualenablinghigh

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

Neural networks forget old abilities when trained on new data. This paper tackles that 'continual learning' problem in a strict setting: no replay of old data, and no task labels at test time. The proposed Look-Ahead method works in two steps for every new task. First, it briefly trains a temporary copy of the model on the new task alone. Using the gradient of the model's output with respect to each weight, it computes an importance of every parameter for the new task, and compares it with the importance stored for old tasks. A parameter whose old-task importance is large relative to the new-task importance is heavily regularized, so it barely moves; a parameter the new task needs is left free to change. The main model is then trained on the new task with a quadratic penalty scaled per parameter by this relative importance. The idea is to allow backward transfer: parameters important for old tasks but even more important for the new task may still update, which can help both.

On two text-classification benchmarks (class-incremental intents and domain-incremental counseling behavior), the method reports the best average overall F1 among the compared replay-free, task-agnostic methods. But the margins are small, about 3 F1 points on CIL and 0.5 on DIL, and the standard deviations are comparable to the margins. A threshold-dependent table shows the method is sometimes worse than the ANCL-MAS baseline, and the backward-transfer numbers have very high variance. A third, preliminary experiment with Llama-3.2 on math reasoning shows that simple importance regularization collapses both old and new abilities, so this part only provides insights, not a working generative method.

Overall the paper describes a sensible, clearly specified modification of parameter-regularization c

Extended reading notes

Core claim

The central claim, stated in the abstract and Section 3.2, is: 'Unlike existing methods, our approach allows the update of parameters with high past-task importance when they have low relative-importance, thus enabling backward knowledge transfer in addition to tackling the stability-plasticity trade-off'; plus Contribution 2: 'our method leads to improved performance on both domain-incremental and class-incremental learning text classification scenarios.' If correct, LA-MAS is a useful per-parameter regularizer for replay-free, task-agnostic CL, with expected gains of about 3 F1 points on CIL and 0.5 F1 on DIL over the tested baselines.

Load-bearing premise

The load-bearing empirical assumption, introduced in Eq 4-5 and Algorithm 1, is that the ratio alpha_rel = alpha_past/(alpha_past+alpha_la+epsilon), computed from MAS gradients on a look-ahead model trained only on the new task, correctly identifies parameters that can be updated without causing forgetting and may enable backward transfer. The paper assumes low alpha_rel means 'more important to new task knowledge accumulation' and therefore safe to update. If MAS importance on an unregularized look-ahead model is not a reliable proxy for cross-task functional importance, the per-parameter decisions are ungrounded. A second assumption is that the current-task validation signal used in the CHSF search is sufficient to set lambda, lambda_up, lambda_down, and tau_alpha_rel without access to past data, because no past-task performance is observed during selection.

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

Quick take: LA-MAS is a real idea — defining per-parameter importance as the ratio of past to past+current importance (alpha_rel) and using it to selectively relax the regularizer on parameters that look new-task-relevant. That's a clean modification of MAS+ANCL, and the two-phase look-ahead is spelled out well. It deserves referee time.

What the paper does well: the method is precisely specified (Eq 4–5, Alg 1), the ablations in Appendix C are unusually thorough for this literature, the cost analysis in Appendix E is useful, and the authors are honest that the text-generation part is preliminary and that naive MAS regularization fails there. Code is promised/linked. The relative-importance ratio is new in the cited literature, so novelty is fine — this is an incremental extension of existing ideas but a reasonably sharp one.

Where I'd be careful: the central empirical claim — that LA-MAS beats ANCL-MAS by ~3 F1 on CIL and ~0.5 on DIL — is built on three task orderings, with standard deviations that overlap (7.4 vs 6.4 on CIL). No significance tests are reported. Table 7 shows that on one CIL sequence at the main threshold (thr=90), LA-MAS is actually worse than ANCL-MAS (32.91 vs 36.98), so the headline is threshold-dependent. The BWT+ claim on CIL (7.88 ± 11.1) is driven by a single ordering; on DIL, LA-MAS's BWT+ (5.64) is below SEQ (7.01) and ANCL-LWF (7.61), which undercuts the backward-transfer narrative. More worrying: the key LA hyperparameters (tau_alpha_rel = 0.8*mean, lambda_down/lambda_up schedules) were selected from ablations on the same task sequences (Appendix C, D.2) that then produce the reported performance. The CHSF search uses only current-task validation, so the choices are not genuinely held out. That's a post-hoc tuning risk, not a proof of fraud — the paper is transparent about it — but it means the apparent gains could partly be artifacts.

Is the mechanism plausible? Yes. The ratio makes sense as a heuristic for which parameters are likely to be safe to update, and the per-parameter up/down scaling is reasonable. The math is simple, almost trivially so, but that's not a flaw given the empirical aim.

Bottom line: the paper deserves peer review, not a desk reject. A good referee should ask for (1) more task orderings, (2) significance tests or effect-size reporting, (3) a fixed hyperparameter protocol with a genuinely held-out sequence, and (4) explicit reporting of per-ordering results. With that, the method could be a useful addition for replay-free, task-agnostic text classification. I'd bring it to a reading group as a case study in why small-delta, few-seed CL results need better statistical care.

Referee Report

4 major / 4 minor

Summary. The paper proposes LA-MAS, a replay-free, task-agnostic continual learning method. For each new task, it first trains a temporary look-ahead model on the new task data, computes MAS-based parameter importances for the past and current tasks, and defines a relative importance ratio α^rel = α_past/(α_past + α_la + ε). Parameters with low relative importance are regularized less even if they had high past-task importance, supposedly permitting plasticity and backward knowledge transfer. The method is evaluated on class-incremental (CIL) and domain-incremental (DIL) text classification, compared with replay-free baselines, and includes a preliminary analysis for text generation. The central empirical claims are that LA-MAS improves overall F1 over state-of-the-art baselines on both CIL and DIL and that it enables backward transfer.

Significance. If the empirical claims hold, the relative parameter importance measure is a useful per-parameter regularizer for replay-free, task-agnostic CL, and the look-ahead two-phase procedure is a plausible way to estimate current-task importance without past data. The paper provides code, a broad baseline set, computation-cost comparisons (Appendix E), and detailed ablations (Appendices C, D), which are strengths. However, the central empirical support is currently weak: only three task orderings, no significance tests, clear threshold dependence that reverses the ranking, and hyperparameter selection performed on the same benchmark sequences that are later used for evaluation. The text-generation section is explicitly preliminary and shows that the regularizer can collapse training. Thus the contribution is better described as a plausible mechanism than as an established improvement.

major comments (4)

[§4.2, Table 1] The headline improvement over the strongest baseline is not statistically supported. On CIL, LA-MAS vs ANCL-MAS is 34.20±7.4 vs 31.19±6.4 (gap 3.01); on DIL, LA-MAS vs ANCL-LWF is 50.48±4.8 vs 50.02±4.3 (gap 0.46). Both gaps are within one standard deviation, and the comparison uses only three random task orderings (Table 6). No significance test is reported. Since the abstract and Contribution 2 claim improved performance, this evidence is not sufficient. Please report per-ordering results and provide significance tests (e.g., paired bootstrap) or additional orderings.
[§D.1, Table 7] The ranking reverses under the main configuration. At thr=90, the threshold used for CIL in §4.1, LA-MAS achieves Ov=32.91 vs ANCL-MAS 36.98 and BWT+=0.00 vs 0.33 on the second CIL sequence. At thr=80 ANCL-MAS is again better (52.62 vs 48.92). The advertised backward-transfer gain appears only at lower thresholds. Because thr=90 is selected a priori for the CIL experiments, Table 7 directly contradicts the claim that LA-MAS outperforms ANCL-MAS at that threshold on one of the three evaluated sequences. Please explain this discrepancy or justify the threshold independently of the outcomes.
[Appendix C, Appendix D.2, Algorithm 1] The hyperparameters τ_alpha_rel=0.8×mean, the CF-KT Balance weighting, and the λ_down/λ_up search schedule are selected from ablations in Appendix C and D.2 using the second class-incremental sequence (Table 6) — the same sequence is among the three used in Table 1. The CHSF search in Algorithm 1 uses only current-task validation accuracy and cannot observe past-task forgetting; the stopping condition (line 24) is a slope threshold on current-task accuracy. This creates a selection-on-test risk: the reported gains may be artifacts of tuning to the benchmark. Please show that the chosen configuration is stable on held-out sequences or justify the search procedure without using the evaluation sequences.
[Abstract, §3.2, Table 1, Table 7] The backward-transfer claim is not consistently supported. On DIL in Table 1, LA-MAS BWT+ (5.64±1.0) is below SEQ (7.01±1.7) and ANCL-LWF (7.61±2.2). On CIL, the BWT+ value 7.88±11.1 has enormous variance and appears driven by a single ordering; Table 7 reports BWT+=0.00 at thr=90. Since backward transfer is a central differentiator in the abstract and Section 3.2, the claim should be qualified or supported with evidence that the effect is reproducible across orderings and thresholds.

minor comments (4)

[Eq. (1)] Equation (1) defines α as the gradient of the L2 norm of the output logits, but the norm is missing in the displayed equation; also the notation α_k for aggregated importance overloads the task index. Please clarify.
[Algorithm 1] Algorithm 1 uses variables x and budget without defining them in the main text; the stopping criterion 'Slope(Acc(θ_k, v_k))' is also not precisely defined. Please add definitions.
[Table 7] Table 7 states 'Results on the second class-incremental sequence' but does not specify which of the three orders in Table 6 is used. Add an explicit reference to the order.
[§5, Table 3] The text-generation analysis shows that importance-based regularization causes both pretrained and target accuracy to collapse for all λ values. This is an honest negative result, but it should be more explicitly framed as a limitation, and the abstract should not imply that the method extends to generation without this caveat.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The method rests on several domain assumptions and hand-set hyperparameters; no new physical entities. The relative-importance measure is a derived quantity, not independently confirmed.

free parameters (6)

lambda (overall regularization strength) = varies per task; e.g., 4.50 in Appendix D
Selected by CHSF: start high, decrease by 10% until current-task validation reaches thr% of best unregularized accuracy; controls the global weight of the L2 regularizer.
lambda_up = lambda_max/lambda with lower bound 1/tau_alpha_rel
Scales the penalty for high-alpha_rel parameters; searched via the validation budget loop in Algorithm 1.
lambda_down = in [0,1]; e.g., 1.00, 0.80, 0.01 in ablations
Scales the penalty for low-alpha_rel parameters; initialized to 1 and decremented until current-task validation performance plateaus.
tau_alpha_rel (relative-importance cutoff) = 80% of per-layer alpha_rel distribution mean
Chosen from ablation Table 8 on the same CL sequences to balance Ov/CF/FWT; a post-hoc tuning choice.
thr = 90 (CIL), 95 (DIL)
Validation-accuracy threshold for lambda search; selected by authors and affects stability/plasticity trade-off.
drop = 10%
Decrement step for lambda, lambda_up, lambda_down in hyperparameter search.

assumptions (4)

domain assumption MAS gradient importance alpha = gradient of L2 norm of output logits w.r.t. theta is a valid per-parameter importance measure for continual learning.
Eq 1; the relative-importance ratio inherits all limitations of MAS; if this proxy fails, the LA regularizer's decisions fail.
domain assumption Mean-pooling importance across tasks preserves enough information to protect all past tasks.
Section 3.1, footnote 2; authors acknowledge granularity loss and leave alternatives to future work.
domain assumption The current-task validation set is sufficient to select lambda, lambda_up, lambda_down, tau_alpha_rel in a replay-free setting without access to past data.
Algorithm 1; no past data is used for validation, yet selected hyperparameters are assumed to generalize to past-task preservation.
ad hoc to paper Low relative importance (alpha_rel <= tau) implies the parameter contributes to new-task learning and shared knowledge, so updating it is safe and enables backward transfer.
Eq 4/5 and Section 3.2; this is the paper's core hypothesis, supported only by the reported experiments.

invented entities (1)

Relative parameter importance alpha^rel_k = alpha_{k-1}/(alpha_{k-1}+alpha^la_k+epsilon)
purpose: Per-parameter weight used to decide which parameters may be updated when learning task k (low alpha_rel) and which must be frozen (high alpha_rel).
A constructed ratio of two MAS importance vectors; the only evidence is the paper's own benchmarks, with no independent falsifiable prediction outside these datasets.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Relative Parameter Importance in Task-Agnostic Replay-Free Continual Learning." pith.science (2026). https://pith.science/paper/CCQTAR2K

@misc{pith2026260800630,
  author       = {Pith},
  title        = {Pith review of: Relative Parameter Importance in Task-Agnostic Replay-Free Continual Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CCQTAR2K}},
  note         = {Machine review of arXiv:2608.00630}
}

read the original abstract

Achieving continual learning (CL) with deep neural networks requires balancing stability and plasticity while enabling knowledge transfer. In this work, we focus on offline learning algorithms under the constraints: (I) no access to training data from prior tasks (II) no access to task-id at inference time. We introduce a novel measure, the relative parameter-importance, which measures the relative importance of each parameter with respect to both the current and past tasks. Parameters with high relative importance are interpreted as more important for maintaining past-task stability and thus heavily regularised, whereas parameters with low relative-importance are allowed to be more freely updated. Unlike existing methods, our approach allows the update of parameters with high past-task importance when they have low relative-importance, thus enabling backward knowledge transfer in addition to tackling the stability-plasticity trade-off. We demonstrate improvements against state-of-the-art CL methods on both class-incremental and domain-incremental learning text classification problems and provide insights for extending our method to text generation problems. Code available at: https://github.com/itsmemala/LACL

Figures

Figures reproduced from arXiv: 2608.00630 by the authors.

**Figure 2.** Impact of different weighting strategies (CF-Focus, KT-Focus) and look [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗

**Figure 3.** Impact of the CF-KT Balance strategy and look-ahead hyper-parameters [PITH_FULL_IMAGE:figures/full_fig_p024_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 2 linked inside Pith

[1]

In: Women in Machine Learning Workshop at NeurIPS (2026)

Abbes, I., et al.: Revisiting replay and gradient alignment for continual pretraining of large language models. In: Women in Machine Learning Workshop at NeurIPS (2026)

2026
[2]

In: Proceedings of the 33rd NeurIPS Conference (2019)

Ahn, H., et al.: Uncertainty-based continual learning with adaptive regularization. In: Proceedings of the 33rd NeurIPS Conference (2019)

2019
[3]

In: Proceedings of the European conference on computer vision (ECCV) (2018)

Aljundi, R., et al.: Memory aware synapses: Learning what (not) to forget. In: Proceedings of the European conference on computer vision (ECCV) (2018)

2018
[4]

CoRR (2021)

Austin, J., et al.: Program synthesis with large language models. CoRR (2021)

2021
[5]

In: Advances in Neural Information Processing Systems (2019)

de Masson d'Autume, C., et al.: Episodic memory in lifelong language learning. In: Advances in Neural Information Processing Systems (2019)

2019
[6]

In: Proceedings of the 34th NeurIPS Conference (2020)

Buzzega, P., et al.: Dark experience for general continual learning: A strong, simple baseline. In: Proceedings of the 34th NeurIPS Conference (2020)

2020
[7]

In: Proceedings of ECCV (2018)

Chaudhry, A., et al.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: Proceedings of ECCV (2018)

2018
[8]

IEEE Transactions on Circuits and Sys- tems for Video Technology (2025)

Cheng, D., et al.: Achieving plasticity-stability trade-off in continual learning through adaptive orthogonal projection. IEEE Transactions on Circuits and Sys- tems for Video Technology (2025)

2025

Show all 46 references

[9]

ArXiv (2021)

Cobbe, K., et al.: Training verifiers to solve math word problems. ArXiv (2021)

2021
[10]

Applied Intelligence (2023)

Dekhovich, A., et al.: Continual prune-and-select: class-incremental learning with specialized subnetworks. Applied Intelligence (2023)

2023
[11]

In: Continual Learning Workshop at NeurIPS (2018)

Diaz-Rodriguez, N., et al.: Don’t forget, there is more than forgetting: new metrics for continual learning. In: Continual Learning Workshop at NeurIPS (2018)

2018
[12]

Ebrahimi,S.,etal.:Adversarialcontinuallearning.In:ProceedingsofECCV(2020)

2020
[13]

In: ICLR (2024)

Elsayed, M., Mahmood, A.R.: Addressing loss of plasticity and catastrophic for- getting in continual learning. In: ICLR (2024)

2024
[14]

In: ICML (2019)

Houlsby, N., et al.: Parameter-efficient transfer learning for nlp. In: ICML (2019)

2019
[15]

In: ICML (2024)

Hu, Y., et al.: Task-aware orthogonal sparse network for exploring shared knowl- edge in continual learning. In: ICML (2024)

2024
[16]

In: Proceedings of NAACL (2021)

Huang, Y., et al.: Continual learning for text classification with information disen- tanglement based regularization. In: Proceedings of NAACL (2021)

2021
[17]

Computer Vision and Image Understanding (2022)

Jodelet, Q., et al.: Balanced softmax cross-entropy for incremental learning with and without memory. Computer Vision and Image Understanding (2022)

2022
[18]

In: Proceedings of the NeurIPS Conference (2020)

Jung, S., et al.: Continual learning with node-importance based adaptive group sparse regularization. In: Proceedings of the NeurIPS Conference (2020)

2020
[19]

Suresh et al

Kang,H.,etal.:Forget-freecontinuallearningwithwinningsubnetworks.In:ICML (2022) 12 M. Suresh et al

2022
[20]

Advances in Neural Information Processing Systems34(2021)

Ke, Z., et al.: Achieving forgetting prevention and knowledge transfer in continual learning. Advances in Neural Information Processing Systems34(2021)

2021
[21]

In: Proceedings of ECML-PKDD (2021)

Ke, Z., et al.: Continual learning with knowledge transfer for sentiment classifica- tion. In: Proceedings of ECML-PKDD (2021)

2021
[22]

In: Proceedings of the CVPR Conference (2023)

Kim, S., et al.: Achieving a better stability-plasticity trade-off via auxiliary net- works in continual learning. In: Proceedings of the CVPR Conference (2023)

2023
[23]

Pro- ceedings of the national academy of sciences114(13) (2017)

Kirkpatrick, J., et al.: Overcoming catastrophic forgetting in neural networks. Pro- ceedings of the national academy of sciences114(13) (2017)

2017
[24]

CoRR (2019)

Lange, M.D., et al.: Continual learning: A comparative study on how to defy for- getting in classification tasks. CoRR (2019)

2019
[25]

IEEE TPAMI (2016)

Li, Z., Hoiem, D.: Learning without forgetting. IEEE TPAMI (2016)

2016
[26]

In: International Conference on Learning Representations (2022)

Lin, S., et al.: TRGP: Trust region gradient projection for continual learning. In: International Conference on Learning Representations (2022)

2022
[27]

In: Proceedings of the 41st ICML (2024)

Lin, W., et al.: An effective dynamic gradient calibration method for continual learning. In: Proceedings of the 41st ICML (2024)

2024
[28]

Liu, X., et al.: Benchmarking natural language understanding services for building conversational agents. In: Int. Workshop on Spoken Dialogue Systems (2021)

2021
[29]

Advances in neural information processing systems30(2017)

Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. Advances in neural information processing systems30(2017)

2017
[30]

Lu, K., et al.: Online merging optimizers for boosting rewards and mitigating tax in alignment (2024),https://arxiv.org/abs/2405.17931

2024 arXiv
[31]

In: CoLLAs (2021)

Lubana, E.S., et al.: How do quadratic regularizers prevent catastrophic forgetting: The role of interpolation. In: CoLLAs (2021)

2021
[32]

IEEE Transactions on Multimedia (2024)

Luo, Y., et al.: Modeling inner- and cross-task contrastive relations for continual image classification. IEEE Transactions on Multimedia (2024)

2024
[33]

In: Proceedings of the CVPR Conference (2020)

Rajasegaran, J., et al.: Itaml: An incremental task-agnostic meta-learning ap- proach. In: Proceedings of the CVPR Conference (2020)

2020
[34]

In: Interna- tional Conference on Learning Representations (2021)

Saha, G., et al.: Gradient projection memory for continual learning. In: Interna- tional Conference on Learning Representations (2021)

2021
[35]

In: International Conference on Machine Learning (2025)

Sanyal, S., et al.: Upweighting easy samples in fine-tuning mitigates forgetting. In: International Conference on Machine Learning (2025)

2025
[36]

In: International Conference on Learning Representations (2025)

Sarfraz, F., et al.: Semantic aware representation learning for lifelong learning. In: International Conference on Learning Representations (2025)

2025
[37]

In: ICML (2018)

Serra, J., et al.: Overcoming catastrophic forgetting with hard attention to the task. In: ICML (2018)

2018
[38]

In: International Conference on Learning Representations (2020)

Sun, F., et al.: LAMOL: language modeling for lifelong language learning. In: International Conference on Learning Representations (2020)

2020
[39]

In: Proceedings of the ACM International Conference on Multimedia (2024)

Sun, W., et al.: Incremental learning via robust parameter posterior fusion. In: Proceedings of the ACM International Conference on Multimedia (2024)

2024
[40]

In: Proceedings of the WACV Conference (2024)

Szatkowski, F., et al.: Adapt your teacher: Improving knowledge distillation for exemplar-free continual learning. In: Proceedings of the WACV Conference (2024)

2024
[41]

In: 7th Workshop on Computational Linguistics and Clinical Psychology (2021)

Tavabi, L., et al.: Analysis of behavior classification in motivational interviewing. In: 7th Workshop on Computational Linguistics and Clinical Psychology (2021)

2021
[42]

In: NeurIPS (2020)

Wortsman, M., et al.: Supermasks in superposition. In: NeurIPS (2020)

2020
[43]

In: IEEE International Conference on Acoustics, Speech and Signal Processing (2022)

Wu, Z., et al.: Anno-mi: A dataset of expert-annotated counselling dialogues. In: IEEE International Conference on Acoustics, Speech and Signal Processing (2022)

2022
[44]

IEEE TPAMI (2025)

Yang, E., et al.: Revisiting flatness-aware optimization in continual learning with orthogonal gradient projection. IEEE TPAMI (2025)

2025
[45]

In: International Conference on Learning Representations (2024)

Yu, L., et al.: Metamath: Bootstrap your own mathematical questions for large language models. In: International Conference on Learning Representations (2024)

2024
[46]

Zeng, G., et al.: Continual learning of context-dependent processing in neural net- works. Nature Machine Intelligence1(8) (2019) Relative Parameter Importance in Continual Learning 13 A Look Ahead Continual Learning with Continual Hyper-Parameter Search Selecting the optimal ...

2019 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

In: Women in Machine Learning Workshop at NeurIPS (2026)

Abbes, I., et al.: Revisiting replay and gradient alignment for continual pretraining of large language models. In: Women in Machine Learning Workshop at NeurIPS (2026)

2026

[2] [2]

In: Proceedings of the 33rd NeurIPS Conference (2019)

Ahn, H., et al.: Uncertainty-based continual learning with adaptive regularization. In: Proceedings of the 33rd NeurIPS Conference (2019)

2019

[3] [3]

In: Proceedings of the European conference on computer vision (ECCV) (2018)

Aljundi, R., et al.: Memory aware synapses: Learning what (not) to forget. In: Proceedings of the European conference on computer vision (ECCV) (2018)

2018

[4] [4]

CoRR (2021)

Austin, J., et al.: Program synthesis with large language models. CoRR (2021)

2021

[5] [5]

In: Advances in Neural Information Processing Systems (2019)

de Masson d'Autume, C., et al.: Episodic memory in lifelong language learning. In: Advances in Neural Information Processing Systems (2019)

2019

[6] [6]

In: Proceedings of the 34th NeurIPS Conference (2020)

Buzzega, P., et al.: Dark experience for general continual learning: A strong, simple baseline. In: Proceedings of the 34th NeurIPS Conference (2020)

2020

[7] [7]

In: Proceedings of ECCV (2018)

Chaudhry, A., et al.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: Proceedings of ECCV (2018)

2018

[8] [8]

IEEE Transactions on Circuits and Sys- tems for Video Technology (2025)

Cheng, D., et al.: Achieving plasticity-stability trade-off in continual learning through adaptive orthogonal projection. IEEE Transactions on Circuits and Sys- tems for Video Technology (2025)

2025