pith. sign in

arxiv: 2605.04617 · v2 · submitted 2026-05-06 · 💻 cs.CV · cs.HC· cs.LG

Temporal Structure Matters for Efficient Test-Time Adaptation in Wearable Human Activity Recognition

Zishu Zhou , Zaipeng Xie , Xuanyao Jie This is my paper

Pith reviewed 2026-05-12 02:49 UTC · model grok-4.3

classification 💻 cs.CV cs.HCcs.LG
keywords test-time adaptationwearable human activity recognitiontemporal structureprototype-based adaptationedge deploymentonline model updatesensor stream adaptation
0
0 comments X

The pith

SIGHT adapts wearable activity models at test time by comparing features to prototypes to decide when to update and when to hold steady.

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

Wearable human activity recognition models lose accuracy when the user changes because training data does not match new movement patterns. Existing test-time adaptation methods treat each sensor window in isolation and therefore miss the fact that activities usually continue across consecutive windows. The paper shows that measuring how far the current feature vector deviates from a stored prototype of expected behavior supplies two signals at once: how surprised the model should be and whether the current activity is likely to be ending. Using these signals to route updates only at probable transitions lets the model refine itself on unlabeled streams while keeping memory and compute low enough for direct edge deployment.

Core claim

The paper claims that temporal continuity and observation-induced feature deviations supply complementary cues for deciding when to preserve or release temporal inertia in WHAR streams. SIGHT implements this by estimating predictive surprise through direct comparison of the current feature against a prototype-based expected state, then routing any prediction refinement with geometry-aware alignment and stream-level marginal habit tracking, all without backpropagation.

What carries the argument

SIGHT, a backpropagation-free TTA framework that estimates predictive surprise by prototype comparison and routes adaptation via feature-deviation-guided geometry-aware transition and marginal habit tracking.

If this is right

  • Models can be updated continuously on new users without any labeled test data.
  • Memory and compute stay low enough for direct execution on wearable hardware.
  • Adaptation occurs mainly at likely activity changes rather than at every window.
  • Performance holds under cross-user distribution shifts typical of real-world use.

Where Pith is reading between the lines

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

  • The same surprise-plus-deviation logic could apply to other streaming sensor tasks such as gesture recognition or physiological monitoring.
  • If prototypes begin to drift over long sessions, an occasional lightweight reset mechanism would be needed to keep the reference states accurate.
  • The geometry-aware routing step may extend naturally to multi-sensor fusion where different modalities have their own transition statistics.

Load-bearing premise

Temporal continuity together with feature deviations from prototypes give reliable cues for when to adapt and when to preserve the current model state in unlabeled streams.

What would settle it

On a dataset of real wearable streams where activity transitions are abrupt or sensor noise causes rapid prototype drift, SIGHT would lose its accuracy advantage over standard TTA baselines.

Figures

Figures reproduced from arXiv: 2605.04617 by Xuanyao Jie, Zaipeng Xie, Zishu Zhou.

Figure 1
Figure 1. Figure 1: Scenario Illustration. Sensor streams naturally form temporal structure through sustained activity segments and infrequent transitions between them over time. Test-time adaptation (TTA) has gained attention as a promising approach to address distribution shifts by adapt￾ing models during inference using only test data (Wang et al. 2021). Compared to UDA and SFDA, TTA is more practical for real-world applic… view at source ↗
Figure 2
Figure 2. Figure 2: Motivation of SIGHT. (a) Output-space transition statistics show strong diagonal dominance, reflecting activ￾ity persistence and limited transition flexibility. (b) Feature￾space deviation offers an observation-conditioned direction for routing refinement toward aligned activity prototypes. class-level geometry. Let the linear head be parameterized by class weights W = [w1, . . . , wK] ⊤, where the logit o… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of SIGHT. It has predictive transition modulation, geometric state routing, and memory update. Expected State Projection. Predictive transition modula￾tion requires a reference feature describing what the cur￾rent observation should look like if the previous activity state persists. A simple choice is to use zˆt−1 as this refer￾ence. However, zˆt−1 is only a noisy observation and may vary within t… view at source ↗
Figure 5
Figure 5. Figure 5: Efficiency comparison. (a) Per-sample inference time (ms), measured on a single CPU thread. (b) Memory overhead (KB), defined as the exact byte size of additional tensor state each method stores beyond the shared backbone. Adaptability to Backbone Architectures (RQ5). SIGHT can be seamlessly applied to various deep model architec￾tures. As view at source ↗
Figure 4
Figure 4. Figure 4: Sensitivity Analysis. ∆ MF1-score over Source￾only under different parameter values. Efficiency Analysis (RQ4). As WHAR models are typ￾ically deployed on resource-constrained edge devices, we evaluate SIGHT in inference time and memory usage. Spa￾tially, SIGHT only maintains a prototype bank and a habit vector, with memory cost O(Kd) for K activity classes. Computationally, its surprise estimation, geometr… view at source ↗
read the original abstract

Wearable human activity recognition (WHAR) models often suffer from performance degradation under real-world cross-user distribution shifts. Test-time adaptation (TTA) mitigates this degradation by adapting models online using unlabeled test streams, yet existing methods largely inherit assumptions from vision tasks and underexploit the inherent inter-window temporal structure in WHAR streams. In this paper, we revisit such temporal structure as a feature-conditioned inference signal rather than merely an output-space smoothing prior. We derive the insight that temporal continuity and observation-induced feature deviations provide complementary cues for determining when to preserve or release temporal inertia and where to route prediction refinement during likely transitions. Building upon this insight, we propose SIGHT, a lightweight and backpropagation-free TTA framework for WHAR, enabling real-time edge deployment. SIGHT estimates predictive surprise by comparing the current feature with a prototype-based expected state, and then uses the resulting feature deviation to guide geometry-aware transition routing based on prototype alignment and stream-level marginal habit tracking. Evaluations on real-world datasets confirm that SIGHT outperforms existing TTA baselines while reducing computational and memory costs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes SIGHT, a lightweight backpropagation-free test-time adaptation (TTA) framework for wearable human activity recognition (WHAR) under cross-user distribution shifts. It treats temporal continuity and observation-induced feature deviations as complementary cues for preserving or releasing temporal inertia, estimating predictive surprise by comparing current features against prototype-based expected states and routing refinements via geometry-aware transition logic that incorporates prototype alignment and stream-level marginal habit tracking. The method targets real-time edge deployment and is claimed to outperform existing TTA baselines while lowering computational and memory costs, based on evaluations on real-world datasets.

Significance. If the empirical claims hold under rigorous validation, the work could be significant for practical TTA in resource-constrained wearable sensing, where backprop-free online adaptation is desirable. The reframing of temporal structure as a feature-conditioned signal (rather than output-space smoothing) is a potentially useful distinction for time-series adaptation, and the emphasis on a small number of free parameters (surprise threshold, prototype update rate) aligns with efficiency goals. No mention of machine-checked proofs or fully reproducible code artifacts is present, but the design avoids heavy reliance on fitted parameters that could circularly define performance.

major comments (2)
  1. [§4.2] §4.2 (Prototype update and surprise estimation): The central claim that feature deviation from prototype-based expected states supplies reliable complementary cues for inertia decisions and routing rests on prototypes remaining accurate in fully unlabeled cross-user streams. The update rule (exponential moving average or equivalent) is described without analysis of drift or mode collapse under activity distribution shifts; this is load-bearing because corrupted prototypes would invalidate both the surprise signal and the geometry-aware routing, yet no stability bounds, anchor mechanisms, or cross-user ablation isolating this component are provided.
  2. [§5] §5 (Experimental validation): The abstract and method summary assert outperformance with reduced costs, but the reported results lack quantitative effect sizes, error bars, dataset counts, or ablations that isolate the temporal deviation and habit-tracking components versus simpler baselines. Without these, it is impossible to determine whether the claimed superiority is robust or sensitive to post-hoc threshold choices, undermining the central efficiency-and-accuracy claim.
minor comments (2)
  1. [§4.3] The precise definition and implementation of 'geometry-aware transition routing' and 'stream-level marginal habit tracking' would benefit from an explicit equation or pseudocode block to clarify how prototype alignment is computed and combined with marginal statistics.
  2. [§5] Figure captions and axis labels in the experimental section should explicitly state the metrics (e.g., accuracy delta, FLOPs, memory footprint) and the exact baselines being compared to improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment below and commit to revisions that strengthen the empirical and analytical rigor of the work without altering its core claims.

read point-by-point responses

  1. Referee: [§4.2] §4.2 (Prototype update and surprise estimation): The central claim that feature deviation from prototype-based expected states supplies reliable complementary cues for inertia decisions and routing rests on prototypes remaining accurate in fully unlabeled cross-user streams. The update rule (exponential moving average or equivalent) is described without analysis of drift or mode collapse under activity distribution shifts; this is load-bearing because corrupted prototypes would invalidate both the surprise signal and the geometry-aware routing, yet no stability bounds, anchor mechanisms, or cross-user ablation isolating this component are provided.

    Authors: We agree that a dedicated stability analysis would strengthen the presentation. The manuscript already employs a conservative exponential moving average update rate chosen to balance adaptation and stability across users, and the reported cross-user results implicitly rely on this choice. In revision we will add (i) an ablation varying the prototype update rate, (ii) a quantitative discussion of observed prototype drift on the evaluated datasets, and (iii) a brief analysis of why mode collapse is mitigated by the geometry-aware routing term. No formal stability bounds are derived in the current work; we will therefore frame the added material as empirical robustness evidence rather than a theoretical guarantee. revision: yes

  2. Referee: [§5] §5 (Experimental validation): The abstract and method summary assert outperformance with reduced costs, but the reported results lack quantitative effect sizes, error bars, dataset counts, or ablations that isolate the temporal deviation and habit-tracking components versus simpler baselines. Without these, it is impossible to determine whether the claimed superiority is robust or sensitive to post-hoc threshold choices, undermining the central efficiency-and-accuracy claim.

    Authors: We acknowledge that the experimental section would benefit from greater statistical detail and component-wise ablations. The current evaluations are performed on multiple real-world WHAR datasets with comparisons to several TTA baselines, demonstrating both accuracy gains and lower compute/memory footprints. In the revised manuscript we will (i) report mean accuracy with standard deviations over multiple random seeds, (ii) include effect-size tables (absolute and relative improvements), (iii) explicitly list the number of subjects and windows per dataset, and (iv) add ablations that disable surprise estimation and habit tracking individually while keeping all other elements fixed. These additions will directly address sensitivity to the surprise threshold. revision: yes

Circularity Check

0 steps flagged

No significant circularity; SIGHT's core claims rest on independent design choices and external benchmarks.

full rationale

The paper's derivation begins from the stated insight that temporal continuity and feature deviations supply complementary cues for inertia decisions, then implements this via prototype-based surprise estimation, geometry-aware routing, and marginal habit tracking. These elements are introduced as methodological choices (with tunable thresholds and update rules) rather than quantities whose values are forced by the performance metrics they are later evaluated against. No equation reduces a claimed gain to a self-fit of the same unlabeled stream, no self-citation chain is invoked to establish uniqueness of the approach, and the evaluation compares against independent TTA baselines on real-world datasets. The method is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that unlabeled test streams contain stable enough temporal structure to maintain useful prototypes without external labels or supervision. No explicit free parameters are named in the abstract, but the surprise threshold and prototype update rules are implicit design choices that must be set. No new physical entities are introduced.

free parameters (2)
  • surprise threshold
    Used to decide when feature deviation triggers transition routing; value must be chosen or tuned per dataset.
  • prototype update rate
    Controls how quickly the expected state prototypes adapt to new observations.
axioms (1)
  • domain assumption Temporal continuity in activity streams provides reliable cues for when to preserve or release prediction inertia.
    Invoked in the derivation of the insight that temporal structure is a feature-conditioned signal.

pith-pipeline@v0.9.0 · 5498 in / 1481 out tokens · 32891 ms · 2026-05-12T02:49:04.059890+00:00 · methodology

Review history (2 revisions) →

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

53 extracted references · 53 canonical work pages

  1. [1]

    and Labrador, Miguel A

    Lara, Oscar D. and Labrador, Miguel A. , journal=. A Survey on Human Activity Recognition using Wearable Sensors , year=

    work page

  2. [2]

    Cross-Domain HAR: Few-Shot Transfer Learning for Human Activity Recognition , journal =

    Thukral, Megha and Haresamudram, Harish and Pl\". Cross-Domain HAR: Few-Shot Transfer Learning for Human Activity Recognition , journal =

    work page

  3. [3]

    2021 , journal =

    Chen, Kaixuan and Zhang, Dalin and Yao, Lina and Guo, Bin and Yu, Zhiwen and Liu, Yunhao , title =. 2021 , journal =

    work page 2021

  4. [4]

    Heterogeneous Mutual Knowledge Distillation for Wearable Human Activity Recognition , year=

    Xiao, Zhiwen and Xing, Huanlai and Qu, Rong and Li, Hui and Cheng, Xinzhou and Xu, Lexi and Feng, Li and Wan, Qian , journal=. Heterogeneous Mutual Knowledge Distillation for Wearable Human Activity Recognition , year=

    work page

  5. [5]

    Proceedings of the 34th ACM International Conference on Information and Knowledge Management , pages =

    Ghosh, Indrajeet and Chugh, Garvit and Faridee, Abu Zaher Md and Roy, Nirmalya , title =. Proceedings of the 34th ACM International Conference on Information and Knowledge Management , pages =. 2025 , address =

    work page 2025

  6. [6]

    2023 , address=

    Hu, Rong and Chen, Ling and Miao, Shenghuan and Tang, Xing , booktitle=. 2023 , address=

    work page 2023

  7. [7]

    Kang, Hua and Hu, Qingyong and Zhang, Qian , journal=

    work page

  8. [8]

    Proceedings of the 27th European Conference on Artificial Intelligence , pages =

    Tong Wu and Yanbin Liu and Sira Yongchareon , title =. Proceedings of the 27th European Conference on Artificial Intelligence , pages =. 2024 , month=

    work page 2024

  9. [9]

    Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume =

    Meng Xue and Yinan Zhu and Wentao Xie and Zhixian Wang and Yanjiao Chen and Kui Jiang and Qian Zhang , title =. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume =. 2025 , pages =

    work page 2025

  10. [10]

    2026 , address=

    Deng, Ruxi and Bao, Wenxuan and Wei, Tianxin and He, Jingrui , booktitle=. 2026 , address=

    work page 2026

  11. [11]

    IEEE Transactions on Image Processing , title=

    Tian, Yuntong and Li, Kang and He, Tianyang and Wan, Liang and Heng, Pheng. IEEE Transactions on Image Processing , title=. 2026 , volume=

    work page 2026

  12. [12]

    Engineering Applications of Artificial Intelligence , volume =

    Yujie Sun and Xiaolong Xu and Xincheng Tian and Lelai Zhou and Yibin Li , title =. Engineering Applications of Artificial Intelligence , volume =

    work page

  13. [13]

    Innovative Dual-Decoupling CNN With Layer-Wise Temporal-Spatial Attention for Sensor-Based Human Activity Recognition , year=

    Teng, Qi and Li, Wei and Hu, Guangwei and Shu, Yuanyuan and Liu, Yun , journal=. Innovative Dual-Decoupling CNN With Layer-Wise Temporal-Spatial Attention for Sensor-Based Human Activity Recognition , year=

    work page

  14. [14]

    Journal of Medical Systems , volume=

    An adaptive hidden Markov model for activity recognition based on a wearable multi-sensor device , author=. Journal of Medical Systems , volume=

    work page

  15. [15]

    Sabatini , title =

    Andrea Mannini and Angelo M. Sabatini , title =. Computational Intelligence and Neuroscience , volume =

    work page

  16. [16]

    Journal of Ambient Intelligence and Humanized Computing , volume=

    Online human activity recognition employing hierarchical hidden Markov models , author=. Journal of Ambient Intelligence and Humanized Computing , volume=

    work page

  17. [17]

    IEEE Journal of Biomedical and Health Informatics , volume =

    Tianyun Zhu and Yilin Dong and Yong Zhou and Changming Zhu and Lei Cao , title =. IEEE Journal of Biomedical and Health Informatics , volume =

    work page

  18. [18]

    Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies, Volume 5: HEALTHINF , pages =

    Hui Liu and Tingting Xue and Tanja Schultz , title =. Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies, Volume 5: HEALTHINF , pages =. 2023 , address =

    work page 2023

  19. [19]

    Baraka, AbdulRahman M. A. and Mohd Noor, Mohd Halim , journal=. Similarity Segmentation Approach for Sensor-Based Activity Recognition , year=

    work page

  20. [20]

    International Journal of Computer Vision , volume=

    A Comprehensive Survey on Test-Time Adaptation Under Distribution Shifts , author=. International Journal of Computer Vision , volume=

    work page

  21. [21]

    2023 , month=

    Hyesu Lim and Byeonggeun Kim and Jaegul Choo and Sungha Choi , booktitle=. 2023 , month=

    work page 2023

  22. [22]

    Proceedings of the 39th International Conference on Machine Learning , pages =

    Efficient Test-Time Model Adaptation without Forgetting , author =. Proceedings of the 39th International Conference on Machine Learning , pages =. 2022 , month =

    work page 2022

  23. [23]

    Proceedings of the AAAI Conference on Artificial Intelligence , pages=

    Decomposing the Neurons: Activation Sparsity via Mixture of Experts for Continual Test Time Adaptation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , pages=. 2026 , address=

    work page 2026

  24. [24]

    2022 , month=

    Zhang, Marvin and Levine, Sergey and Finn, Chelsea , booktitle=. 2022 , month=

    work page 2022

  25. [25]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Boudiaf, Malik and Mueller, Romain and Ben Ayed, Ismail and Bertinetto, Luca , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    work page

  26. [26]

    Proceedings of the International Conference on Learning Representations , month=

    Theoretical Analysis of Self-Training with Deep Networks on Unlabeled Data , author=. Proceedings of the International Conference on Learning Representations , month=

    work page

  27. [27]

    A Hybrid Discriminative/Generative Approach for Modeling Human Activities , author=. Proc. Int. Joint Conf. Artif. Intell. , pages =. 2005 , month=

    work page 2005

  28. [28]

    Self-Supervised Contrastive Representation Learning for Semi-Supervised Time-Series Classification , year=

    Eldele, Emadeldeen and Ragab, Mohamed and Chen, Zhenghua and Wu, Min and Kwoh, Chee-Keong and Li, Xiaoli and Guan, Cuntai , journal=. Self-Supervised Contrastive Representation Learning for Semi-Supervised Time-Series Classification , year=

    work page

  29. [29]

    2026 , eprinttype =

    Zhang, Ronghao and Niu, Shuaicheng and Deng, Qi and Dong, Yanjie and Chen, Jian and Zeng, Runhao , journal =. 2026 , eprinttype =

    work page 2026

  30. [30]

    2020 , volume =

    Chang, Youngjae and Mathur, Akhil and Isopoussu, Anton and Song, Junehwa and Kawsar, Fahim , title =. 2020 , volume =

    work page 2020

  31. [31]

    2026 , author =

    Source-free domain adaptation for unsupervised radar-based human activity recognition , journal =. 2026 , author =

    work page 2026

  32. [32]

    Sanabria, Andrea Rosales and Zambonelli, Franco and Dobson, Simon and Ye, Juan , journal=

    work page

  33. [33]

    Deep Generative Domain Adaptation with Temporal Relation Knowledge for Cross-User Activity Recognition , booktitle =

    Xiaozhou Ye and Kevin I. Deep Generative Domain Adaptation with Temporal Relation Knowledge for Cross-User Activity Recognition , booktitle =. 2024 , month =

    work page 2024

  34. [34]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Dynamic Few-Shot Visual Learning without Forgetting , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=. 2018 , month=

    work page 2018

  35. [35]

    Test-Time Classifier Adjustment Module for Model-Agnostic Domain Generalization , year =

    Iwasawa, Yusuke and Matsuo, Yutaka , booktitle =. Test-Time Classifier Adjustment Module for Model-Agnostic Domain Generalization , year =

    work page

  36. [36]

    Do We Really Need to Access the Source Data?

    Liang, Jian and Hu, Dapeng and Feng, Jiashi , booktitle =. Do We Really Need to Access the Source Data?. 2020 , month =

    work page 2020

  37. [37]

    CoRR , volume =

    Order-Aware Test-Time Adaptation: Leveraging Temporal Dynamics for Robust Streaming Inference , author=. CoRR , volume =. 2026 , eprinttype =

    work page 2026

  38. [38]

    Unsupervised domain adaptation via softmax-based prototype construction and adaptation , journal =

    Jingyao Li and Shuai L. Unsupervised domain adaptation via softmax-based prototype construction and adaptation , journal =

    work page

  39. [39]

    2021 , month=

    Dequan Wang and Evan Shelhamer and Shaoteng Liu and Bruno Olshausen and Trevor Darrell , booktitle=. 2021 , month=

    work page 2021

  40. [40]

    2022 , month=

    Gong, Taesik and Jeong, Jongheon and Kim, Taewon and Kim, Yewon and Shin, Jinwoo and Lee, Sung-Ju , booktitle =. 2022 , month=

    work page 2022

  41. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Wang, Qin and Fink, Olga and Van Gool, Luc and Dai, Dengxin , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    work page

  42. [42]

    Proceedings of the International Conference on Learning Representations , year=

    Towards Stable Test-time Adaptation in Dynamic Wild World , author=. Proceedings of the International Conference on Learning Representations , year=

    work page

  43. [43]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Yuan, Longhui and Xie, Binhui and Li, Shuang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    work page

  44. [44]

    2024 , volume =

    Wang, Shuoyuan and Wang, Jindong and Xi, Huajun and Zhang, Bob and Zhang, Lei and Wei, Hongxin , title =. 2024 , volume =

    work page 2024

  45. [45]

    2026 , author =

    Expert Systems with Applications , volume =. 2026 , author =

    work page 2026

  46. [46]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining , month =

    Gong, Peiliang and Ragab, Mohamed and Wu, Min and Chen, Zhenghua and Su, Yongyi and Li, Xiaoli and Zhang, Daoqiang , title =. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining , month =. 2025 , pages =

    work page 2025

  47. [47]

    A Prototype-Oriented Framework for Unsupervised Domain Adaptation , year =

    Tanwisuth, Korawat and Fan, Xinjie and Zheng, Huangjie and Zhang, Shujian and Zhang, Hao and Chen, Bo and Zhou, Mingyuan , booktitle =. A Prototype-Oriented Framework for Unsupervised Domain Adaptation , year =

    work page

  48. [48]

    Proceedings of the 42nd International Conference on Machine Learning , pages =

    Test-time Correlation Alignment , author =. Proceedings of the 42nd International Conference on Machine Learning , pages =. 2025 , month =

    work page 2025

  49. [49]

    Proceedings of the 42nd International Conference on Machine Learning , pages =

    Han, Jisu and Na, Jaemin and Hwang, Wonjun , title =. Proceedings of the 42nd International Conference on Machine Learning , pages =. 2025 , month =

    work page 2025

  50. [50]

    Sensors , volume=

    Logacjov, Aleksej and Bach, Kerstin and Kongsvold, Atle and B. Sensors , volume=

    work page

  51. [51]

    Chan, Shing and Hang, Yuan and Tong, Catherine and Acquah, Aidan and Schonfeldt, Abram and Gershuny, Jonathan and Doherty, Aiden , journal=

    work page

  52. [52]

    Davide Anguita and Alessandro Ghio and Luca Oneto and Xavier Parra and Jorge L. Reyes. A Public Domain Dataset for Human Activity Recognition Using Smartphones , booktitle =

    work page

  53. [53]

    , title =

    Zhang, Mi and Sawchuk, Alexander A. , title =. 2012 , booktitle =

    work page 2012