Pith. sign in

REVIEW 2 major objections 5 minor 49 references

Channel-wise gating of a frozen backbone reduces catastrophic forgetting in wearable activity recognition while training under 2% of parameters.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-08-02 18:34 UTC pith:MMAKIBQ3

load-bearing objection Solid empirical study of channel-wise gates on a frozen HAR backbone, but Theorem 4's proof has a variable mix-up and the abstract oversells what the gates alone contribute. the 2 major comments →

arxiv 2603.10046 v2 pith:MMAKIBQ3 submitted 2026-03-08 cs.LG cs.AI

Gated Adaptation for Continual Learning in Human Activity Recognition

Reza Rahimi Azghan , Gautham Krishna Gudur , Mohit Malu , Edison Thomaz , Giulia Pedrielli , Pavan Turaga , Hassan Ghasemzadeh This is my paper
classification cs.LG cs.AI
keywords continual learningcatastrophic forgettinghuman activity recognitionparameter-efficient fine-tuningchannel-wise gatingfrozen backbonedomain-incremental learningstability-plasticity tradeoff
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that continual learning in human activity recognition can be made stable and plastic at once by freezing a pretrained feature extractor and learning only lightweight channel-wise gates that rescale its outputs. The core claim is that adaptation should select and scale existing features rather than generate new ones, because a diagonal scaling operator bounds how much representations can drift when a new subject arrives. On PAMAP2, this reduces catastrophic forgetting from 39.7% to 16.2% and lifts final accuracy from 56.7% to 77.7%, while training under 2% of the model's parameters. The authors also give a theoretical account: gating bounds feature and logit drift, and under a channel-wise domain-shift assumption, diagonal gates can express the needed subject-specific transformation. The practical payoff is a replay-free, privacy-preserving on-device learner that does not store prior users' sensor data.

Core claim

The paper's central claim is that structured diagonal operators — channel-wise multiplicative gates applied to a frozen pretrained backbone — achieve a strong stability-plasticity tradeoff in domain-incremental human activity recognition. The authors prove three things: gating bounds feature drift by δ(x)∥U(x)∥_F with δ(x)<1 (Theorem 1); logit drift decomposes into a gate-induced term and a classifier-induced term (Theorem 2); and a margin condition guarantees zero forgetting (Corollary 1). They also prove that if cross-subject shift is approximately channel-wise (Assumption 1), a suitable gate can reproduce the subject-specific feature map up to global scaling (Theorem 4). Empirically, free

What carries the argument

The central object is the channel-wise gate, a squeeze-and-excitation style module inserted after each block of a frozen pretrained CNN. It computes a per-channel descriptor by global average pooling, passes it through a two-layer bottleneck MLP with sigmoid, and applies the resulting vector g∈(0,1)^C as a diagonal scaling D(g) on the feature map. This implements a bounded diagonal operator: it rescales channel magnitudes but never mixes channels, so the geometry of the pretrained representation is preserved. The proofs use the fact that the spectral norm of a diagonal matrix equals the infinity norm of its diagonal, giving a multiplicative bound on drift. The expressiveness proof relies on

Load-bearing premise

Assumption 1 — that a new subject's backbone features are approximately a channel-wise scaling of a canonical feature map — is the load-bearing premise. If real subject shifts involve cross-channel rotation or mixing, diagonal gates cannot express the needed transformation, and the expressiveness theorem and the stability-plasticity benefit would not follow. The paper's only direct empirical support for this assumption is a correlation analysis on PAMAP2 activity-class centro

What would settle it

Compute the cross-channel correlation matrix of backbone features on UCI-HAR and DSA (as done for PAMAP2 in Figure 4). If off-diagonal correlations are systematically nonzero, or if artificially rotating the sensor coordinate frame between subjects causes this method to forget more than a full adapter, the channel-wise assumption fails. Alternatively, measure the residual ε_t(x) in U_t(x) vs. D(s_t)U_bar(x) for each subject; if the residual is not small relative to the feature norm, Theorem 4 does not apply.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Wearable models can adapt to new users on-device without storing raw sensor data from previous users, addressing privacy and memory constraints in IoT deployments.
  • Freezing the backbone plus gating reduces task-order variance (std of final accuracy drops from 8.9% to 2.5% on PAMAP2), making results more predictable across subject orders.
  • Because the method is replay-free and trains under 2% of parameters, it is compatible with low-power edge devices; combining it with replay (when allowed) further cuts forgetting to 6.1% on PAMAP2.
  • The margin-based forgetting bound gives a per-sample sufficient condition for zero forgetting, so safety-critical deployments could in principle flag low-margin samples as at-risk.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The channel-wise assumption suggests a testable general principle: in any domain-incremental setting where distribution shift is mostly per-channel gain/offset (e.g., EEG, EMG, or sensor calibration drift), diagonal gating of frozen features may yield similar stability; the paper's correlation analysis could be replicated on those modalities to check.
  • If cross-channel mixing dominates a shift — say, because a sensor is rotated between subjects so axes covary — the theory predicts this method will struggle; a direct experiment would artificially rotate the coordinate frame between subjects and compare forgetting against a full adapter.
  • The proof of Theorem 4 as written applies to the canonical feature map, not the actual subject-dependent backbone output; a revised proof or a corrected algorithm that gates on the canonical map would be needed to make the expressiveness guarantee formally airtight.
  • One could push further: instead of scalar gates, use per-channel affine transforms with a small Lipschitz constraint; this might recover residual shifts while keeping the bounded-drift argument.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. This paper proposes a parameter-efficient continual learning framework for subject-incremental human activity recognition. A CNN backbone pretrained on WISDM is frozen, and channel-wise gating modules (squeeze-and-excitation style bottleneck MLPs producing sigmoid multipliers in (0,1)) are inserted after each residual stack to reweight the frozen features; a shared linear classifier is trained across tasks. Only gates and classifier are updated (claimed <2% of parameters), with no replay buffer or task-specific regularization. The paper proves stability bounds (Theorems 1–3: bounded feature drift, bounded logit drift, margin-based prediction-preservation condition) and claims an expressiveness result (Theorem 4) under a channel-wise domain-shift assumption (Assumption 1). Experiments on PAMAP2 (8 subjects, 12 activities), DSA (8, 18), and UCI-HAR (30, 6) compare against EWC, LwF, HAT, and DER/DER++, with ablations isolating backbone freezing, gating, stacked adapters, knowledge distillation, task-aware gates, and replay hybrids. Headline results on PAMAP2: forgetting drops from 39.7% (trainable backbone) to 16.2%, final accuracy rises from 56.7% to 77.7%.

Significance. If the results hold, the paper is a solid, practical contribution to on-device continual learning for wearable HAR. Strengths include a clean architecture and careful ablations: the comparison with stacked trainable layers (Table V) supports the 'selection vs. generation' thesis, the freezing × gating ablation (Table IV) isolates both design choices, and the gates-with-replay hybrids (Table VI) demonstrate orthogonality to replay. The stability analysis (Theorems 1–3) is correct, though elementary, and gives a qualitative explanation of drift control. Reproducibility is good: architecture tables, hyperparameters, 10 random task orders, and mean±std statistics are reported. The main weakness is Theorem 4 — the theoretical justification for 'provably sufficient' diagonal gating — whose proof is invalid as written (see major comments). Because the empirical contribution is substantial and largely independent of that theorem, the paper is repairable, but the current version overstates its theoretical support.

major comments (2)
  1. [§IV-B, Theorem 4] The proof of Theorem 4 does not follow from Assumption 1. Under Assumption 1 the observed backbone output is U(x) ≈ D(s_t)\bar U(x)+ε_t(x). The theorem sets g_t=s_t/α_t and asserts D(g_t)U(x)=α_t^{-1}D(s_t)U(x)=α_t^{-1}U_t(x) (exact case). That equality holds only if the gate acts on the canonical map \bar U(x): it requires D(s_t)U(x)=U_t(x)=D(s_t)\bar U(x). Under the stated assumption, D(s_t)U(x)=D(s_t)^2\bar U(x), so the gated output is α_t^{-1}D(s_t)^2\bar U(x), not α_t^{-1}U_t(x). The only exact choice is g_t=1 (trivial, no adaptation); the approximate-case proof likewise writes U_t(x)=D(s_t)U(x)+ε_t(x), inconsistent with Assumption 1. Hence the stated bound and the 'provably sufficient expressiveness' claim fail. Separately, the theorem shows existence of a gate vector but not that the SE gate network (which maps z(x) through a bottleneck MLP) can realize it. Please correct the targ
  2. [§IV-B, Remark 2 / Fig. 4] The empirical support for Assumption 1 is insufficient. (i) It is computed only for PAMAP2, yet the assumption motivates the interpretation of all three benchmarks. (ii) Each entry of the 512×512 correlation matrix is estimated from K≈12 paired class-centroid observations per subject pair; the reported 'off-diagonal ≈ 0' needs error bars or a permutation baseline. (iii) Pearson correlation is invariant to per-channel scaling, so high diagonal correlation does not establish the assumed fixed per-subject scaling vector; per-class scaling or affine channel transforms would produce the same pattern. A direct test—estimate s_t by least squares and report ∥U(x)−D(s_t)\bar U(x)∥_F/∥U(x)∥_F per subject, and check stability of s_t across activity classes—would substantiate or delimit the assumption.
minor comments (5)
  1. [§V-B2, Table III, Abstract] Parameter-count inconsistency: the abstract and Table III report <2% (1.7%) trainable parameters, but the architecture description gives 65,536 parameters per 512-channel gate and gates after all four stacks (256/384/512/512 channels), totaling ≈184K gate parameters plus a 512×K classifier, i.e., ≈2.6% of the 7.4M-parameter model. Please clarify the computation or the number of inserted gates.
  2. [§VI-A] Please report the hyperparameters used for EWC, LwF, and HAT (e.g., EWC λ, distillation temperature/α, HAT mask sparsity), and state whether HAT receives task identity at test time. The main method is task-free; HAT's mask mechanism typically requires task labels, which affects the comparison's interpretation.
  3. [§V-C, Tables II and IV] The gate-only gain on a frozen backbone (PAMAP2: Pretrained 76.5±4.0 vs. Pretrained+Gates 77.7±2.5) is within one standard deviation. With 10 permutations, report paired differences or significance tests to support the claim that gates restore adaptation capacity.
  4. [§IV, Assumption 1 / Theorem 4] The notation U(x), U_t(x), and \bar U(x) is used inconsistently between Assumption 1 and Theorem 4. Define U_t(x) explicitly and maintain the distinction between the observed backbone output, the canonical map, and the target subject-specific map throughout the section.
  5. [§I and §VI] The Introduction states that the experiments 'validate our theoretical predictions,' but Corollary 1 involves unmeasured quantities (margins m(x), feature norms, gate drift δ(x), classifier drift), and the reported metrics (FA, FM, LA) do not directly test the bounds. Rephrase as 'consistent with' or measure the relevant quantities.

Circularity Check

1 steps flagged

Theorem 4's expressiveness proof defines the target U_t as D(s_t)U(x), making diagonal gating match its own output by construction; the paper's empirical comparisons remain independent.

specific steps
  1. self definitional [Section IV-B, Theorem 4 (Expressiveness of Diagonal Gating), proof]
    "Proof. Choose α_t > ∥s_t∥∞ (e.g., α_t = (1+ϵ)∥s_t∥∞ for small ϵ > 0) and define g_t = s_t/α_t. Then g_t ∈ (0,1)^C and D(g_t)U(x) = 1/α_t D(s_t)U(x). In the exact case (ε_t = 0), we have U_t(x) = D(s_t)U(x), so H_t(x) = α_t^{-1}U_t(x) exactly."

    Assumption 1 states the gate operates on U(x), the subject-dependent backbone output, with U(x) ≈ D(s_t)U̅(x). The proof's chain D(g_t)U(x) = α^{-1}D(s_t)U(x) = α^{-1}U_t(x) holds only if U_t(x) is defined as D(s_t)U(x), i.e., the target is defined to be the scaling the gate is about to apply. Then H_t(x)=α^{-1}U_t(x) is an identity by construction, not a demonstration that diagonal gates express an independently specified subject adaptation. If instead U_t(x) is the actual subject feature D(s_t)U̅(x), the equality fails because D(g_t)U(x)=α^{-1}D(s_t)^2U̅(x). Either reading makes the expressiveness claim reduce to its own definition or conflates U with U̅.

full rationale

The paper's main empirical results (PAMAP2, DSA, UCI-HAR) are evaluated on held-out test splits after sequential training, with no fitted constants renamed as predictions; the stability bounds in Theorems 1-3 are algebraic consequences of the frozen backbone and diagonal gate structure. Self-citations (CLAD-Net, AttengluCo) and the SE-Net inspiration appear only in related work and do not bear the derivation's weight. The one load-bearing circular/by-construction step is Theorem 4, where the proof sets U_t(x)=D(s_t)U(x), making the gate's success an immediate definitional identity. This does not invalidate the empirical comparisons, but it does make the claimed theoretical guarantee of diagonal-gating expressiveness partially circular and not a derivation from Assumption 1 as given. Hence a moderate partial-circularity score rather than a higher one.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central premise is Assumption 1 (channel-wise scaling). The remaining axioms are standard math and domain assumptions. No new particles, forces, or physical entities are introduced. The free parameters are common training/architecture hyperparameters; none are fitted to data in a way that manufactures the reported accuracy.

free parameters (4)
  • Gate reduction ratio r = 8
    Bottleneck dimension in the gate MLP (hidden = max(C/8,16)). Chosen by hand; no sensitivity analysis is given.
  • L2 regularization coefficient λ = 0.0001
    Weight decay for gates and classifier; reported as a fixed hyperparameter without tuning or justification.
  • Learning rate for continual training = 0.001
    Adam learning rate for gates and classifier; standard default, not motivated by analysis.
  • Placement and number of gates = 4 (after each residual stack)
    Architecture choice; alternative placements are not explored, so the optimality of this choice is unknown.
axioms (6)
  • domain assumption Assumption 1: Cross-subject domain shift is approximately channel-wise: U(x) ≈ D(s_t) \bar U(x) + ε_t(x) with small residual.
    Load-bearing premise for Theorem 4. Empirically supported only on PAMAP2 via activity-class centroid correlations; not validated on DSA or UCI-HAR.
  • domain assumption Domain-incremental setting has fixed label distribution: P_t(y)=P_{t+1}(y).
    Standard and plausible for subject-incremental HAR; stated in Section III-B.
  • domain assumption Frozen backbone pretrained on WISDM transfers to PAMAP2, DSA, and UCI-HAR.
    Needed for the empirical results; the paper acknowledges this may fail when source and target domains differ substantially.
  • standard math Standard matrix norm inequalities (submultiplicativity, Cauchy-Schwarz, spectral norm of diagonal = sup norm of diagonal).
    Used in Theorems 1-3; standard and correct.
  • standard math Sigmoid output lies in (0,1), giving δ(x)<1.
    Elementary; used to claim bounded drift.
  • ad hoc to paper In Theorem 4, the gate is effectively assumed to operate on the canonical feature map \bar U rather than the actual subject-dependent U(x).
    This is the notational conflation that invalidates the exact-match claim. The architecture applies D(g_t) to the subject's actual backbone output U_t(x), not to a canonical feature.

pith-pipeline@v1.3.0-alltime-deepseek · 20345 in / 15443 out tokens · 146552 ms · 2026-08-02T18:34:50.682711+00:00 · methodology

0 comments
read the original abstract

Wearable sensors in Internet of Things (IoT) ecosystems increasingly support applications such as remote health monitoring, elderly care, and smart home automation, all of which rely on robust human activity recognition (HAR). Continual learning systems must balance plasticity (learning new tasks) with stability (retaining prior knowledge), yet AI models often exhibit catastrophic forgetting, where learning new tasks degrades performance on earlier ones. This challenge is especially acute in domain-incremental HAR, where on-device models must adapt to new subjects with distinct movement patterns while maintaining accuracy on prior subjects without transmitting sensitive data to the cloud. We propose a parameter-efficient continual learning framework based on channel-wise gated modulation of frozen pretrained representations. Our key insight is that adaptation should operate through feature selection rather than feature generation: by restricting learned transformations to diagonal scaling of existing features, we preserve the geometry of pretrained representations while enabling subject-specific modulation. We provide a theoretical analysis showing that gating implements a bounded diagonal operator that limits representational drift compared to unconstrained linear transformations. Empirically, freezing the backbone substantially reduces forgetting, and lightweight gates restore lost adaptation capacity, achieving stability and plasticity simultaneously. On PAMAP2 with 8 sequential subjects, our approach reduces forgetting from 39.7% to 16.2% and improves final accuracy from 56.7% to 77.7%, while training less than 2% of parameters. Our method matches or exceeds standard continual learning baselines without replay buffers or task-specific regularization, confirming that structured diagonal operators are effective and efficient under distribution shift.

Figures

Figures reproduced from arXiv: 2603.10046 by Edison Thomaz, Gautham Krishna Gudur, Giulia Pedrielli, Hassan Ghasemzadeh, Mohit Malu, Pavan Turaga, Reza Rahimi Azghan.

Figure 1
Figure 1. Figure 1: Catastrophic forgetting in subject-incremental HAR: accuracy on [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed continual learning system for HAR, comprising data collection, preprocessing, and a classification model with a frozen [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Channel-wise gating mechanism. Given an intermediate feature map [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cross-subject feature channel correlation matrix averaged over all [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Per-subject accuracy evolution with our gating approach on PAMAP2. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 6 linked inside Pith

  1. [1]

    Continual lifelong learning with neural networks: A review,

    G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, and S. Wermter, “Continual lifelong learning with neural networks: A review,” Neural Networks, vol. 113, pp. 54–71, 2019. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0893608019300231

    2019

  2. [2]

    Catastrophic interference in connec- tionist networks: The sequential learning problem,

    M. McCloskey and N. J. Cohen, “Catastrophic interference in connec- tionist networks: The sequential learning problem,” inPsychology of Learning and Motivation. Elsevier, 1989, vol. 24, pp. 109–165. Subject A Subject B Subject C Subject D Trained Subjects 0 20 40 60 80 100 Accuracy (%) Subject A Subject B Subject C Subject D Fig. 5. Per-subject accuracy e...

    1989

  3. [3]

    Catastrophic forgetting in connectionist networks,

    R. M. French, “Catastrophic forgetting in connectionist networks,” Trends in Cognitive Sciences, vol. 3, no. 4, pp. 128–135, 1999. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S1364661399012942

    1999

  4. [4]

    Advanced internet of things for personalised healthcare systems: A survey,

    J. Qi, P. Yang, G. Min, O. Amft, F. Dong, and L. Xu, “Advanced internet of things for personalised healthcare systems: A survey,”Pervasive and Mobile Computing, vol. 41, pp. 132–149, 2017

    2017

  5. [5]

    Ultra-sensitive wireless capacitive nanocomposite- based pressure sensors for health monitoring,

    S. Hashemi, S. Ebrahimibasabi, M. Sajjadi, N. Shahraki, D. Tamjid Shabestari, M. Golshahi, S. Zeinolabedinzadeh, H. Arami, and L. Khalifehzadeh, “Ultra-sensitive wireless capacitive nanocomposite- based pressure sensors for health monitoring,”Advanced Materials Technologies, vol. 10, no. 19, p. e01316, 2025

    2025

  6. [6]

    A survey on wearable sensor modality centred human activity recognition in health care,

    Y . Wang, S. Cang, and H. Yu, “A survey on wearable sensor modality centred human activity recognition in health care,”Expert Systems with Applications, vol. 137, pp. 167–190, 2019

    2019

  7. [7]

    Deepsense: A unified deep learning framework for time-series mobile sensing data processing,

    S. Yao, S. Hu, Y . Zhao, A. Zhang, and T. Abdelzaher, “Deepsense: A unified deep learning framework for time-series mobile sensing data processing,” inProceedings of the 26th International Conference on World Wide Web, 2017, pp. 351–360

    2017

  8. [8]

    Deepear: robust smartphone audio sensing in unconstrained acoustic environments using deep learn- ing,

    N. D. Lane, P. Georgiev, and L. Qendro, “Deepear: robust smartphone audio sensing in unconstrained acoustic environments using deep learn- ing,” inProceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing, 2015, pp. 283–294

    2015

  9. [9]

    Deep learning for sensor-based human activity recognition: Overview, challenges, and opportunities,

    K. Chen, D. Zhang, L. Yao, B. Guo, Z. Yu, and Y . Liu, “Deep learning for sensor-based human activity recognition: Overview, challenges, and opportunities,”ACM Computing Surveys, vol. 54, no. 4, pp. 1–40, 2021

    2021

  10. [10]

    A survey of mobile phone sensing,

    N. D. Lane, E. Miluzzo, H. Lu, D. Peebles, T. Choudhury, and A. T. Campbell, “A survey of mobile phone sensing,”IEEE Communications Magazine, vol. 48, no. 9, pp. 140–150, 2010

    2010

  11. [11]

    PAMAP2 Physical Activity Monitoring,

    A. Reiss, “PAMAP2 Physical Activity Monitoring,” UCI Machine Learning Repository, 2012, DOI: https://doi.org/10.24432/C5NW2H

    doi:10.24432/c5nw2h 2012

  12. [12]

    Overcoming catastrophic forgetting in neural networks,

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska et al., “Overcoming catastrophic forgetting in neural networks,”Pro- ceedings of the National Academy of Sciences, vol. 114, no. 13, pp. 3521–3526, 2017

    2017

  13. [13]

    Continual learning through synap- tic intelligence,

    F. Zenke, B. Poole, and S. Ganguli, “Continual learning through synap- tic intelligence,” inInternational Conference on Machine Learning. PMLR, 2017, pp. 3987–3995

    2017

  14. [14]

    iCaRL: Incremental classifier and representation learning,

    S.-A. Rebuffi, A. Kolesnikov, G. Sperl, and C. H. Lampert, “iCaRL: Incremental classifier and representation learning,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 2001–2010

    2017

  15. [15]

    Dark experience for general continual learning: a strong, simple baseline,

    P. Buzzega, M. Boschini, A. Porrello, D. Abati, and S. Calderara, “Dark experience for general continual learning: a strong, simple baseline,”

  16. [16]

    Progressive neural networks,

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

    Pith/arXiv arXiv 2016

  17. [17]

    Overcoming 14 catastrophic forgetting with hard attention to the task,

    J. Serr `a, D. Sur ´ıs, M. Miron, and A. Karatzoglou, “Overcoming 14 catastrophic forgetting with hard attention to the task,” 2018. [Online]. Available: https://arxiv.org/abs/1801.01423

    Pith/arXiv arXiv 2018

  18. [18]

    Learning to prompt for continual learning,

    Z. Wang, Z. Zhang, C.-Y . Lee, H. Zhang, R. Sun, X. Ren, G. Su, V . Perot, J. Dy, and T. Pfister, “Learning to prompt for continual learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 139–149

    2022

  19. [19]

    DualPrompt: Complementary prompt- ing for rehearsal-free continual learning,

    Z. Wang, Z. Zhang, S. Ebrahimi, R. Sun, H. Zhang, C.-Y . Lee, X. Ren, G. Su, V . Perot, J. Dyet al., “DualPrompt: Complementary prompt- ing for rehearsal-free continual learning,” inEuropean Conference on Computer Vision (ECCV). Springer, 2022, pp. 631–648

    2022

  20. [20]

    Parameter-efficient transfer learning for NLP,

    N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly, “Parameter-efficient transfer learning for NLP,” inProceedings of the International Conference on Machine Learning (ICML), 2019, pp. 2790–2799

    2019

  21. [21]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 7132–7141

    2018

  22. [22]

    Learning without forgetting,

    Z. Li and D. Hoiem, “Learning without forgetting,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 40, no. 12, pp. 2935– 2947, 2017

    2017

  23. [23]

    Continual learning with deep generative replay,

    H. Shin, J. K. Lee, J. Kim, and J. Kim, “Continual learning with deep generative replay,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017

    2017

  24. [24]

    PackNet: Adding multiple tasks to a single network by iterative pruning,

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

    2018

  25. [25]

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

    L. Wang, X. Zhang, H. Su, and J. Zhu, “A comprehensive survey of continual learning: Theory, method and application,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 5, pp. 5362– 5383, 2024

    2024

  26. [26]

    Class-incremental learning: A survey,

    D.-W. Zhou, Q.-W. Wang, Z.-H. Qi, H.-J. Ye, D.-C. Zhan, and Z. Liu, “Class-incremental learning: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 8, pp. 5513–5533, 2024

    2024

  27. [27]

    Continual learning and catastrophic forgetting,

    G. M. van de Ven, N. Soures, and D. Kudithipudi, “Continual learning and catastrophic forgetting,”arXiv preprint arXiv:2403.05175, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inInternational Conference on Learning Representations (ICLR), 2022

    2022

  29. [29]

    CODA-Prompt: Contin- ual decomposed attention-based prompting for rehearsal-free continual learning,

    J. S. Smith, L. Karlinsky, V . Gutta, P. Cascante-Bonilla, D. Kim, A. Arbelle, R. Panda, R. Feris, and Z. Kira, “CODA-Prompt: Contin- ual decomposed attention-based prompting for rehearsal-free continual learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 11 909–11 919

    2023

  30. [30]

    Continual learning with pre-trained models: A survey,

    D.-W. Zhou, H.-L. Sun, J. Ning, H.-J. Ye, and D.-C. Zhan, “Continual learning with pre-trained models: A survey,” inProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI), 2024, pp. 6593–6601

    2024

  31. [31]

    Reflecting on the state of rehearsal-free continual learning with pretrained models,

    L. Thede, M. Mundt, B. Sick, and V . Ramesh, “Reflecting on the state of rehearsal-free continual learning with pretrained models,”arXiv preprint arXiv:2406.09384, 2024

    arXiv 2024

  32. [32]

    CBAM: Convolutional block attention module,

    S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon, “CBAM: Convolutional block attention module,” inEuropean Conference on Computer Vision (ECCV). Springer, 2018, pp. 3–19

    2018

  33. [33]

    FiLM: Visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. De Vries, V . Dumoulin, and A. Courville, “FiLM: Visual reasoning with a general conditioning layer,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 32, no. 1, 2018

    2018

  34. [34]

    Alleviating catastrophic forgetting using context-dependent gating and synaptic stabilization,

    N. Y . Masse, G. D. Grant, and D. J. Freedman, “Alleviating catastrophic forgetting using context-dependent gating and synaptic stabilization,” Proceedings of the National Academy of Sciences, vol. 115, no. 44, pp. E10 467–E10 475, 2018

    2018

  35. [35]

    ResNet-SE: Channel attention-based deep residual network for complex activity recognition using wrist-worn wearable sensors,

    S. Mekruksavanich, A. Jitpattanakul, K. Sitthithakerngkiet, P. Youplao, and P. Yupapin, “ResNet-SE: Channel attention-based deep residual network for complex activity recognition using wrist-worn wearable sensors,”IEEE Access, vol. 10, pp. 51 142–51 154, 2022

    2022

  36. [36]

    Continual learning for activity recognition,

    K. Chen, L. Yao, D. Zhang, B. Wang, Z. Lu, M. Hong, and C. Lofi, “Continual learning for activity recognition,” inIEEE International Conference on Pervasive Computing and Communications (PerCom). IEEE, 2022, pp. 1–10

    2022

  37. [37]

    Continual learning in sensor-based human activity recognition: An empirical benchmark analysis,

    S. Jha, M. Schiemer, and J. Ye, “Continual learning in sensor-based human activity recognition: An empirical benchmark analysis,”Infor- mation Sciences, vol. 575, pp. 1–21, 2021

    2021

  38. [38]

    Learning causal dynamics and reward machines: A framework for faster reinforce- ment learning with extended temporal tasks,

    H. Partovi Aria, H. Kim, S. Meshkat Alsadat, and Z. Xu, “Learning causal dynamics and reward machines: A framework for faster reinforce- ment learning with extended temporal tasks,” in2025 5th International Conference on Computer, Control and Robotics (ICCCR), 2025, pp. 519–525

    2025

  39. [39]

    Online continual learning for human activity recognition,

    M. Schiemer, L. Fang, S. Dobson, and J. Ye, “Online continual learning for human activity recognition,”Pervasive and Mobile Computing, vol. 93, p. 101817, 2023

    2023

  40. [40]

    Lifelong adaptive machine learning for sensor-based human activity recognition using prototypical networks,

    R. Adaimi and E. Thomaz, “Lifelong adaptive machine learning for sensor-based human activity recognition using prototypical networks,” Sensors, vol. 22, no. 18, p. 6881, 2022

    2022

  41. [41]

    Attengluco: Multimodal transformer-based blood glucose forecasting on ai-readi dataset,

    E. Farahmand, R. R. Azghan, N. T. Chatrudi, E. Kim, G. K. Gudur, E. Thomaz, G. Pedrielli, P. Turaga, and H. Ghasemzadeh, “Attengluco: Multimodal transformer-based blood glucose forecasting on ai-readi dataset,” 2025. [Online]. Available: https://arxiv.org/abs/2502.09919

    Pith/arXiv arXiv 2025

  42. [42]

    Balancing continual learning and fine-tuning for human activity recognition,

    M. B ¨ock, M. Moser, S. Yan, S. Suh, and T. Pl ¨otz, “Balancing continual learning and fine-tuning for human activity recognition,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, 2024, pp. 268–276

    2024

  43. [43]

    Clad-net: Continual activity recognition in multi-sensor wearable systems,

    R. R. Azghan, G. K. Gudur, M. Malu, E. Thomaz, G. Pedrielli, P. Turaga, and H. Ghasemzadeh, “Clad-net: Continual activity recognition in multi-sensor wearable systems,” 2025. [Online]. Available: https://arxiv.org/abs/2509.23077

    arXiv 2025

  44. [44]

    Harnet: Towards on-device incremental learning using deep ensembles on constrained devices,

    P. Sundaramoorthy, G. K. Gudur, M. R. Moorthy, R. N. Bhandari, and V . Vijayaraghavan, “Harnet: Towards on-device incremental learning using deep ensembles on constrained devices,” inProceedings of the 2nd International Workshop on Embedded and Mobile Deep Learning (EMDL ’18), Munich, Germany, 2018, pp. 1–6

    2018

  45. [45]

    Daily and Sports Activities,

    B. Barshan and K. Altun, “Daily and Sports Activities,” UCI Machine Learning Repository, 2010, DOI: https://doi.org/10.24432/C5C59F

    doi:10.24432/c5c59f 2010

  46. [46]

    Human activity recognition using smartphones,

    J. Reyes-Ortiz, D. Anguita, A. Ghio, L. Oneto, and X. Parra, “Human activity recognition using smartphones,” UCI Machine Learning Repos- itory, 2013, dOI: https://doi.org/10.24432/C54S4K

    doi:10.24432/c54s4k 2013

  47. [47]

    WISDM Smartphone and Smartwatch Activity and Bio- metrics Dataset ,

    G. Weiss, “WISDM Smartphone and Smartwatch Activity and Bio- metrics Dataset ,” UCI Machine Learning Repository, 2019, DOI: https://doi.org/10.24432/C5HK59

    doi:10.24432/c5hk59 2019

  48. [48]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” 2015. [Online]. Available: https://arxiv.org/abs/1503.02531

    Pith/arXiv arXiv 2015

  49. [2020]

    Available: https://arxiv.org/abs/2004.07211

    [Online]. Available: https://arxiv.org/abs/2004.07211

    Pith/arXiv arXiv 2004

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.