Pith. sign in

REVIEW 4 major objections 4 minor 33 references

TypiCore: A Hybrid Active Query Strategy for Class-Incremental Learning on Time Series

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read TypiCore claims that alternating typical and diverse sample selection beats every baseline in label-scarce continual time-series learning.

desk verdict The empirical ACIL time-series comparison is useful and the hybrid idea is plausible, but the headline 'significant over all baselines' is contradicted by the paper's own Section 6.3 and by the flawed rank test. read the letter →

arxiv 2607.17632 v1 pith:OTEJLANY submitted 2026-07-20 cs.LG cs.AI

Gabor Szucs , Samuel Jacsev , Marcell Nemeth , Davide Dalle Pezze , Gian Antonio Susto This is my paper · ORCID
classification cs.LGcs.AI
keywords activeclass-incrementallearningtimeseriesclassificationcontinualrehearsal-basedmemoryreplayTypiCoretypicalitydiversity-basedquerystrategies
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

Active class-incremental learning on time series faces a double bind: a model must keep learning new classes without forgetting old ones, yet only a small portion of incoming labels can be afforded. This paper argues that neither uncertainty-based nor distribution-aware query strategies alone solve that problem. It proposes TypiCore, a hybrid strategy that alternates between typicality-based selection (TypiClust) and diversity-based selection (CoreSet) across active-learning cycles, building memory buffers that are both representative and diverse. The paper reports that TypiCore achieves statistically significant gains over baseline query strategies and, on several benchmark datasets, matches or exceeds fully supervised continual learning while using only a fraction of the labels. A sympathetic reader would care because this directly targets the real-world annotation bottleneck in continual learning.

What carries the argument

TypiCore is the central mechanism: a hybrid active query strategy that strictly alternates TypiClust and CoreSet across successive active-learning cycles within a task. TypiClust selects typical samples by clustering the unlabeled pool and scoring each candidate by the inverse average Euclidean distance to its K nearest neighbors; CoreSet applies a greedy k-center furthest-first heuristic to maximize geometric coverage of the labeled set. The alternation, initialized with TypiClust, is what allows the cumulative labeled set to capture both the distributional core of each class and the broader feature-space spread.

What would settle it

Re-run the benchmark and recompute the Friedman/Nemenyi test using the five individual runs as repeated measures rather than the 24 aggregate ranks; if any adjusted pairwise p-value for TypiCore versus a baseline reaches 0.02 or higher, the claimed significance over all baselines is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that memory buffers for rehearsal-based class-incremental learning need to be both representative and diverse, and that no single query strategy supplies both. TypiCore achieves this by alternating two distribution-aware selection steps within each task: a TypiClust step that anchors the buffer on typical, high-density samples, and a CoreSet step that extends coverage to underrepresented regions of the feature space. Starting each task with TypiClust handles the cold-start problem, when no labels from the new classes exist yet. On the TSCIL benchmark, the paper reports that TypiCore attains the best average accuracy on most datasets and, despite labeling only a f

Load-bearing premise

The claim that TypiCore beats every baseline statistically depends on a significance test that treats 24 summary numbers as independent, even though they come from only five experimental runs; the paper itself notes one competitor was not significantly beaten.

Editorial extensions

If this is right

  • On the TSCIL benchmark, TypiCore paired with Experience Replay achieves the highest average accuracy on three of the four datasets, with clear margins on UWave, GRABMyo, and WISDM.
  • TypiCore surpasses fully supervised regularization-based continual learning methods (LwF and MAS) on all datasets, suggesting that selective labeling can act as an implicit regularizer that filters noisy or redundant samples.
  • The simpler Experience Replay mechanism outperforms ASER across active-learning settings, especially under severe annotation budgets, indicating that importance-driven memory management designed for full supervision is less robust when the labeled pool is small.
  • Uncertainty-based query strategies consistently fail to improve over random selection and frequently degrade performance, while CoreSet alone causes high forgetting and TypiClust alone underrepresents regions of the feature space.
  • The successful combination of typicality and diversity provides evidence that memory-buffer quality, not just sample informativeness, is the key lever in active class-incremental learning.

Reading between the lines

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

  • Beyond the paper, the alternation pattern suggests a general principle: cold-start cycles should prioritize representativeness, while later cycles should prioritize coverage, and this could be tested with other typicality and diversity measures beyond TypiClust and CoreSet.
  • Also beyond the paper, the label-efficiency gains imply that active continual learning systems could trade annotation effort for model capacity or task complexity in deployment, but the paper does not quantify that trade-off.
  • A testable extension would be dynamic budget allocation: instead of uniform alternation, the ratio of typicality to coverage steps could adapt to the estimated distribution shift between tasks, potentially improving performance on datasets with irregular task difficulties.
  • The comparison to fully supervised methods is suggestive but not a controlled equivalence claim; joint training remains a much higher upper bound, and the paper leaves open whether the gap is due to label scarcity or to the incremental protocol itself.
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.

Referee Report

4 major / 4 minor

Summary. The paper studies Active Class-Incremental Learning (ACIL) for multivariate time series under a fixed annotation budget. It integrates active query strategies into rehearsal-based continual learning (ER and ASER) on the TSCIL benchmark, evaluates uncertainty-based, distribution-aware, and hybrid strategies, and proposes TypiCore, a strategy that alternates TypiClust (typicality) and CoreSet (diversity) selection across active-learning cycles. The authors report point estimates across four datasets and claim that TypiCore achieves statistically significant improvements over all baseline query strategies, while matching or exceeding fully supervised continual learning methods on several datasets.

Significance. If the central empirical claim were fully supported, the paper would make a useful contribution: it extends active continual learning to the time-series domain, provides a systematic comparison of seven query strategies under two replay mechanisms across four datasets, and proposes a plausible hybrid strategy with an intuitive motivation. The observation that simpler ER can outperform ASER in low-budget active regimes is also interesting. However, the headline statistical claim is not currently supported as stated: the paper contains an explicit internal contradiction between the abstract/conclusion and Section 6.3, and the Friedman/Nemenyi analysis used to support the claim is applied to non-independent aggregate ranks. The empirical contribution is thus best seen as a set of favorable point estimates whose significance and robustness require additional statistical and ablation evidence.

major comments (4)
  1. [Section 6.3 / Abstract / Conclusion] The abstract and conclusion state that TypiCore 'delivers statistically significant improvements over all baselines' and that this holds 'under both ER and ASER'. Section 6.3, however, reports that TypiCore 'achieved significantly better values than all competitors (except one)', and Table 4 confirms that TypiClust is the exception (no asterisk in the 'All' row). Since TypiClust is one of the baseline query strategies, the two statements are contradictory. This is a load-bearing issue: the central claim must be corrected either by withdrawing the 'all baselines' formulation or by providing a statistical test that actually demonstrates significance over TypiClust as well.
  2. [Section 6.3, Table 4] The Friedman/Nemenyi analysis is performed on 24 average rank values obtained by pooling three metrics (ACC, FT, A_curr) across four datasets and two replay methods. These metrics are computed from the same five runs and are therefore correlated, so they cannot be treated as independent blocks in a Friedman test. Moreover, using aggregate ranks discards run-level variance, so the pairwise p-values cannot establish significance over all baselines. The reported p-value (7.1e-06) and the pairwise criterion (p < 0.02) do not fix this design issue. I recommend replacing this analysis with per-run paired comparisons (e.g., Wilcoxon signed-rank or paired t-tests with appropriate multiple-comparison correction) on a pre-specified primary metric, and reporting the actual p-values and effect sizes rather than only asterisks.
  3. [Section 5.2 vs. Tables 2, 3, and 5] The text states that results are 'averaged over the five runs with 95% confidence intervals', but Tables 2, 3, and 5 report only point estimates, with no confidence intervals or per-run variability. Without these, the reader cannot assess whether the differences between TypiCore and the baselines are meaningful. This is especially important because the point estimates themselves do not consistently favor TypiCore: on UCI-HAR under ER, TypiCore's ACC is 70.70, lower than TypiClust's 75.26 (Table 2). The claim that TypiCore 'matches or surpasses fully supervised continual learning' also needs accompanying uncertainty quantification.
  4. [Section 4.4 / Section 5.4] TypiCore's core design choice is the strict alternation between TypiClust and CoreSet, initialized with TypiClust. The paper does not provide ablations testing this mechanism: there is no comparison against running TypiClust first versus CoreSet first, no variation in the alternation order or ratio, and no study of sensitivity to the number of active-learning cycles N_a or the budget ratio rho. Since TypiCore is defined directly as the alternation of two existing methods, these ablations are necessary to establish that the hybrid mechanism itself, rather than the strengths of one component, is responsible for the reported gains.
minor comments (4)
  1. [Section 6.3, Table 4] The paragraph describing how the ranking was constructed is confusing. It says 'We created a ranking based on the values ... of the Table 4', but Table 4 already shows rank averages. Please clarify the workflow: raw ranks per dataset/metric, aggregation, and then the Friedman/Nemenyi test.
  2. [Section 5.4] The dataset-specific values of q and rho are described only in prose ('UCI-HAR employs q=300 ... other datasets utilize q in {20,25,40} with ratios between 0.12 and 0.30'). A table or explicit per-dataset listing would improve reproducibility.
  3. [General] No code or repository is mentioned. Given the empirical nature of the work, providing code and per-run results would substantially strengthen reproducibility.
  4. [General] There are several typographical and formatting issues, e.g., 'UW aveis' in Section 5.1 and the spacing in 'T able 1'. A careful proofread is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: TypiCore is a direct composition of two external query strategies; the statistical-significance concerns are correctness issues, not circular reasoning.

full rationale

TypiCore is defined in Section 4.4 as an alternation of two previously published query strategies, TypiClust [26] and Core-Set [23]. There is no equation-level derivation in which an output quantity is defined in terms of the quantity it is claimed to predict, and no fitted parameter is renamed as a prediction. The paper's headline claim is an empirical comparison on the external TSCIL benchmark, with results reported in Tables 2, 3, and 5 against external baselines. The only self-citation, [28], appears in Section 4.2 as background support for ASER's state-of-the-art status ('achieving state-of-the-art performance on the TSCIL benchmark [4, 28]'); it is not used to justify TypiCore's design or to establish the central claim. The design motivation 'from the experimental analysis of individual active learning strategies' is standard empirical method development, not circularity. The concerns raised about the Friedman/Nemenyi aggregation and the internal 'except one' versus 'all baselines' discrepancy are statistical-validity and reporting issues, not self-referential reductions of the claimed result to its inputs. The derivation chain is therefore self-contained with respect to circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

TypiCore introduces no new physical or mathematical entities. Its load-bearing assumptions are benchmark validity, oracle reliability, the complementarity of typicality and coverage, and the stability of cycle-wise training. The main hand-chosen quantities are per-dataset budget ratios and unreported TypiClust parameters.

free parameters (3)
  • Per-dataset annotation budget ratio rho = UCI-HAR 0.01; UWave/GRABMyo/WISDM 0.12-0.30 (q=20/25/40)
    Chosen by hand per dataset to balance annotation cost and learning; the 'fraction of labels' claim depends directly on these values.
  • TypiClust neighbor count K and cluster count k = not reported
    Required to compute typicality and clustering; values are not given, affecting reproducibility and potentially the results.
  • Number of active learning cycles N_a = not reported per dataset (only q is given)
    Controls the alternation schedule and total budget; not specified, hampering replication and sensitivity analysis.
assumptions (4)
  • domain assumption TSCIL dataset splits and random class-order task partitions are a valid proxy for real-world class-incremental time-series streams.
    Section 5.1-5.2; all conclusions are benchmark-specific and may not transfer to non-stationary real-world streams.
  • domain assumption Queried samples' labels from the oracle are correct and the unlabeled pool is static within each task.
    Algorithm 2 assumes a perfect oracle and no intra-task drift, which may not hold in real deployments.
  • ad hoc to paper Typicality (inverse mean KNN distance) and geometric coverage (k-center) are complementary and strict alternation is beneficial.
    Section 4.4 postulates this trade-off and alternation schedule; no ablation or theoretical justification is provided.
  • domain assumption Training to convergence on the cumulative labeled set at each cycle (up to 100 epochs) is appropriate and does not introduce within-task catastrophic forgetting.
    Algorithm 2 and Section 5.4; no analysis of early-cycle overfitting or within-task forgetting is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TypiCore: A Hybrid Active Query Strategy for Class-Incremental Learning on Time Series." pith.science (2026). https://pith.science/paper/OTEJLANY

@misc{pith2026260717632,
  author       = {Pith},
  title        = {Pith review of: TypiCore: A Hybrid Active Query Strategy for Class-Incremental Learning on Time Series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OTEJLANY}},
  note         = {Machine review of arXiv:2607.17632}
}
read the original abstract

Time series data play a pivotal role across numerous domains, including healthcare and manufacturing. In real-world environments, models must cope with distribution shifts over time, a challenge commonly addressed through Continual Learning (CL) techniques. However, existing CL methods face a critical limitation: real-world data streams are rarely fully labeled, making annotation cost a major practical constraint. This paper investigates Active Class-Incremental Learning (ACIL) for multivariate time series, where a model must sequentially learn new classes while selectively querying labels under a fixed annotation budget. We present a systematic evaluation of a wide range of query strategies combined with multiple rehearsal-based approaches, assessing their impact on plasticity, stability, and label efficiency across four benchmark datasets. Our analysis reveals the limitations of uncertainty-based and distribution-aware methods in achieving strong performance under constrained labeling budgets. To address these shortcomings, we propose TypiCore, a novel hybrid query strategy that alternates between typicality-based and diversity-based sample selection across active learning cycles, enabling the construction of memory buffers that are both representative and diverse. Evaluated on the TSCIL benchmark, TypiCore delivers statistically significant improvements over all baselines and matches or surpasses fully supervised continual learning performance on multiple datasets while requiring a fraction of the available labels.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 1 canonical work pages

  1. [28]

    IF AC-PapersOnLine59(26), 229–234 (2025)

    Sz˝ ucs, G., N´ emeth, M., Jacsev, S.: Active learning for class-incremental time series with dual memory and shapley replay. IF AC-PapersOnLine59(26), 229–234 (2025)

    2025

  2. [1]

    Nature Machine Intelligence4(12), 1185–1197 (2022)

    Ven, G.M., Tuytelaars, T., Tolias, A.S.: Three types of incremental learning. Nature Machine Intelligence4(12), 1185–1197 (2022)

    2022

  3. [2]

    Evolving Systems16(4), 111 (2025)

    Pezze, D.D., Anello, E., Masiero, C., Susto, G.A.: Continual learning approaches for anomaly detection. Evolving Systems16(4), 111 (2025)

    2025

  4. [3]

    Evolving Systems17(2), 53 (2026)

    Carta, A., Marinelli, A.R., Passaro, L.C.: Adapting language models with continual learning for temporal drifts. Evolving Systems17(2), 53 (2026)

    2026

  5. [4]

    In: Pro- ceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp

    Qiao, Z., Pham, Q., Cao, Z., Le, H.H., Suganthan, P.N., Jiang, X., Ramasamy, S.: Class-incremental learning for time series: Benchmark and evaluation. In: Pro- ceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 5613–5624 (2024)

    2024

  6. [5]

    Settles, B.: Active learning literature survey (2009) 23

    2009

  7. [6]

    arXiv preprint arXiv:2602.04252 (2026)

    Bhattacharya, A.R., Goswami, D., Chakraborty, S.: Acil: Active class incremental learning for image classification. arXiv preprint arXiv:2602.04252 (2026)

    arXiv 2026

  8. [7]

    Evolving Systems14(3), 437– 460 (2023)

    Yamauchi, K.: Quick continual kernel learning on bounded memory space based on balancing between adaptation and forgetting. Evolving Systems14(3), 437– 460 (2023)

    2023

Show all 33 references
  1. [8]

    IEEE Transactions on Artificial Intelligence5(6), 2526–2546 (2023)

    Wickramasinghe, B., Saha, G., Roy, K.: Continual learning: A review of techniques, challenges, and future directions. IEEE Transactions on Artificial Intelligence5(6), 2526–2546 (2023)

    2023

  2. [9]

    Proceedings of the national academy of sciences114(13), 3521–3526 (2017)

    Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A.,et al.: Over- coming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences114(13), 3521–3526 (2017)

    2017

  3. [10]

    IEEE transactions on pattern analysis and machine intelligence40(12), 2935–2947 (2017)

    Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence40(12), 2935–2947 (2017)

    2017

  4. [11]

    arXiv preprint arXiv:1503.02531 (2015)

    Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)

    2015 arXiv

  5. [12]

    arXiv preprint arXiv:1701.08734 (2017)

    Fernando, C., Banarse, D., Blundell, C., Zwols, Y., Ha, D., Rusu, A.A., Pritzel, A., Wierstra, D.: Pathnet: Evolution channels gradient descent in super neural networks. arXiv preprint arXiv:1701.08734 (2017)

    2017 arXiv

  6. [13]

    arXiv preprint arXiv:1606.04671 (2016)

    Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)

    2016 arXiv

  7. [14]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7765–7773 (2018)

    2018

  8. [15]

    arXiv preprint arXiv:1902.10486 (2019)

    Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486 (2019)

    1902 arXiv

  9. [16]

    Advances in neural information processing systems33, 15920–15930 (2020)

    Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experi- ence for general continual learning: a strong, simple baseline. Advances in neural information processing systems33, 15920–15930 (2020)

    2020

  10. [17]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Rebuffi, S.-A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental clas- sifier and representation learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2001–2010 (2017)

    2001

  11. [18]

    24 Advances in neural information processing systems30(2017)

    Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. 24 Advances in neural information processing systems30(2017)

    2017

  12. [19]

    Advances in neural information processing systems32(2019)

    Aljundi, R., Belilovsky, E., Tuytelaars, T., Charlin, L., Caccia, M., Lin, M., Page- Caccia, L.: Online continual learning with maximal interfered retrieval. Advances in neural information processing systems32(2019)

    2019

  13. [20]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Shim, D., Mai, Z., Jeong, J., Sanner, S., Kim, H., Jang, J.: Online class- incremental continual learning with adversarial shapley value. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, pp. 9630–9638 (2021)

    2021

  14. [21]

    Evolving Systems10(1), 63–78 (2019)

    Cherman, E.A., Papanikolaou, Y., Tsoumakas, G., Monard, M.C.: Multi-label active learning: key issues and a novel query strategy. Evolving Systems10(1), 63–78 (2019)

    2019

  15. [22]

    Mathematics11(4), 820 (2023)

    Tharwat, A., Schenck, W.: A survey on active learning: State-of-the-art, practical challenges and research directions. Mathematics11(4), 820 (2023)

    2023

  16. [23]

    arXiv preprint arXiv:1708.00489 (2017)

    Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core- set approach. arXiv preprint arXiv:1708.00489 (2017)

    2017 arXiv

  17. [24]

    The Scientific World Journal2014(1), 827586 (2014)

    He, T., Zhang, S., Xin, J., Zhao, P., Wu, J., Xian, X., Li, C., Cui, Z.: An active learning approach with uncertainty, representativeness, and diversity. The Scientific World Journal2014(1), 827586 (2014)

    2014

  18. [25]

    arXiv preprint arXiv:2503.00867 (2025)

    Giouroukis, P.S., Gidiotis, A., Tsoumakas, G.: Dual: Diversity and uncertainty active learning for text summarization. arXiv preprint arXiv:2503.00867 (2025)

    2025 arXiv

  19. [26]

    arXiv preprint arXiv:2202.02794 (2022)

    Hacohen, G., Dekel, A., Weinshall, D.: Active learning on a budget: Opposite strategies suit high and low budgets. arXiv preprint arXiv:2202.02794 (2022)

    2022 arXiv

  20. [27]

    ACM Transactions on Mathe- matical Software (TOMS)11(1), 37–57 (1985)

    Vitter, J.S.: Random sampling with a reservoir. ACM Transactions on Mathe- matical Software (TOMS)11(1), 37–57 (1985)

    1985

  21. [29]

    UCI Machine Learning Repository (2013)

    Anguita, D., Reyes-Ortiz, J., Ghio, A., Oneto, L., Parra, X.: Human activity recognition using smartphones. UCI Machine Learning Repository (2013)

    2013

  22. [30]

    arXiv preprint arXiv:2603.20352 (2026)

    Middlehurst, M., Rushbrooke, A., Ismail-Fawaz, A., Devanne, M., Forestier, G., Dempster, A., Webb, G.I., Holder, C., Bagnall, A.: The multiverse of time series machine learning: an archive for multivariate time series classification. arXiv preprint arXiv:2603.20352 (2026)

    2026

  23. [31]

    PhysioNet

    Jiang, N., Pradhan, A., He, J.: Gesture recognition and biometrics electromyo- gram (grabmyo). PhysioNet. doi: https://doi. org/10.13026/dzhr-nj10 (2022) 25

    2022 doi

  24. [32]

    ACM SigKDD Explorations Newsletter12(2), 74–82 (2011)

    Kwapisz, J.R., Weiss, G.M., Moore, S.A.: Activity recognition using cell phone accelerometers. ACM SigKDD Explorations Newsletter12(2), 74–82 (2011)

    2011

  25. [33]

    In: Proceedings of the European Conference on Computer Vision (ECCV), pp

    Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T.: Memory aware synapses: Learning what (not) to forget. In: Proceedings of the European Conference on Computer Vision (ECCV), pp. 139–154 (2018) 26

    2018

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.