CoChat
Sign Up
Cited by 4Context-Aware Activity Recognition SystemsDomain Adaptation and Few-Shot LearningTime Series Analysis and Forecasting

Towards Customizable Foundation Models for Human Activity Recognition with Wearable Devices

Open access
Qiu, Minghui, Weng, Cekai, Fan, Mingming, Wu, Kaishun·2025·Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies
DOI 10.1145/3749479
Open this paper in CoChat — free accountOpen PDF

Abstract

Foundation models have achieved remarkable success across various domains by learning general representations from raw data, offering a promising paradigm for diverse applications. This concept holds great potential for advancing human activity recognition (HAR), particularly in overcoming challenges associated with collecting large-scale labeled datasets. However, the dynamic nature of HAR tasks, characterized by diverse sensing devices and activity types, results in fragmented datasets that question the feasibility of applying foundation model to this domain. In this work, we propose a novel foundation model training framework that effectively leverages heterogeneous datasets through a two-stage training strategy: (1) self-supervised learning to extract cross-domain sensor patterns, followed by (2) multi-task learning to align representations with semantic contexts. The effectiveness of the trained foundation model is demonstrated through extensive downstream experiments, with the superior fine-tuning performance across various modalities and input configurations---achieving the highest performance metric in 10 out of 12 settings---further validating the robustness and adaptability. While our model shows a performance gap compared to foundation models pre-trained on large-scale or high-quality data in zero- and few-shot scenarios, its competitive results with a more flexible architecture demonstrate the efficiency and potential of our training strategy for HAR foundation models.

Integrity signals

Registry

Crossref · 10.1145/3749479

Cites

71 of 71 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

17 of 28 cited sources checked · 8 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    UniMTS is trained with data containing synthesized IMU readings from 22 joint nodes of the human body graph.

    “In comparison, UniMTS [65] is trained with the data that contains synthesized IMU readings from 22 joint nodes of the human body graph.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [65] 65. Mi Zhang and Alexander A Sawchuk. 2012. USC-HAD: A daily activity dataset for ubiquitous activity recognition using wearable sensors. In Proceedings of the 2012 ACM conference on ubiquitous computing. 1036-1043.
  • Findingown data

    The trained HAR-FM achieved the highest performance metric in 10 out of 12 cross-user experiments across various modalities and input configurations against baselines covering state-of-the-art pretraining and/or fine-tuning mechanisms.

    “with the superior fine-tuning performance across various modalities and input configurations-achieving the highest performance metric in 10 out of 12 settings-further validating the robustness and adaptability.”
    • Also stated in: § 1 Introduction · boosted
      “The results demonstrate that the well-trained HAR-FM advances the field by winning 10 out of 12 cross-user experiments with various input settings (e.g., different modalities and time window configurations) against baselines that cover the state-of-the-art pretraining and/or fine-tuning mechanisms.”
    Quote in the paperSubstantial claim
  • Findingown data

    The HAR-FM shows a performance gap compared to foundation models pre-trained on large-scale or high-quality data in zero- and few-shot scenarios.

    “While our model shows a performance gap compared to foundation models pre-trained on large-scale or high-quality data in zero-and few-shot scenarios, its competitive results with a more flexible architecture demonstrate the efficiency and potential of our training strategy for HAR foundation models.”
    • Also stated in: § 1 Introduction · plain
      “Though the performance gap with present foundation model instances persists, the relatively (few times) smaller datasets used for training and more flexible model structure (section 6) of the HAR-FM should further demonstrate the advancement of the proposed training strategy.”
    Quote in the paperSubstantial claim
  • Findingreasoning

    Despite different task definitions, the model learns to organize similar activities into coherent clusters across datasets.

    “Despite these datasets having different task definitions, the model learns to organize similar activities into coherent clusters across datasets.”
    Quote in the paperSubstantial claim
  • Findingown data

    YUAN-SSL(capture24) generalizes to the other two modalities (gyroscope and magnetometer) under 3-second and 5-second window size settings.

    “It is also interesting to observe the generalizability of YUAN-SSL(capture24) on the other two modalities (under 3-second and 5-second window size settings)”
    Quote in the paperNot a substantial claim
  • Classified as probably not a substantial claim. 19
  • May combine two claims. 23
  • The quote only partly states this claim. 1

  • Classified as probably not a substantial claim.

    Premisecitation

    Foundation models learn general-purpose representations from large-scale data that can be adapted to diverse downstream tasks.

    “Foundation Models learn general-purpose representations from large-scale data that can be adapted to diverse downstream tasks [6,8,13,19,25,28,30,35,45,60,62,69].”
    Quote in the paperSubstantial claim7 of 10 sources checked
    • cites [6] 6. Defu Cao, Furong Jia, Sercan O Arik, Tomas Pfister, Yixiang Zheng, Wen Ye, and Yan Liu. 2023. Tempo: Prompt-based generative pre-trained transformer for time series forecasting. arXiv preprint arXiv:2310.04948 (2023). source
    • cites [8] 8. Ching Chang, Wen-Chih Peng, and Tien-Fu Chen. 2023. Llm4ts: Two-stage fine-tuning for time-series forecasting with pre-trained llms. arXiv preprint arXiv:2308.08469 (2023). source
    • cites [13] 13. Azul Garza and Max Mergenthaler-Canseco. 2023. TimeGPT-1. arXiv preprint arXiv:2310.03589 (2023). source
    • cites [19] 19. Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, et al. 2023. Time-llm: Time series forecasting by reprogramming large language models. arXiv preprint arXiv:2310.01728 (2023). source

      The Introduction (page unavailable) says pre-trained models in computer vision and NLP can be fine-tuned for various downstream tasks, and that foundation language models perform diverse NLP tasks. It also discusses transferable representations for time series, but the passages do not establish the claim that foundation models generally learn those representations from large-scale data. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [25] 25. Xu Liu, Junfeng Hu, Yuan Li, Shizhe Diao, Yuxuan Liang, Bryan Hooi, and Roger Zimmermann. 2024. Unitime: A language-empowered unified model for cross-domain time series forecasting. In Proceedings of the ACM on Web Conference 2024. 4095-4106. source
    • cites [28] 28. Mika A Merrill and Tim Althoff. 2023. Self-supervised pretraining and transfer learning enable\titlebreak flu and covid-19 predictions in small mobile sensing datasets. In Conference on Health, Inference, and Learning. PMLR, 191-206. source

      The Methods (Section 3.2, p. 3) say self-supervised pretraining learns generalizable representations from unlabeled data, and the Introduction (p. 2) reports transfer across flu and COVID-19 prediction. The paper does not characterize its model as a foundation model or establish the broader claim that foundation models adapt to diverse downstream tasks. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [30] 30. Girish Narayanswamy, Xin Liu, Kumar Ayush, Yuzhe Yang, Xuhai Xu, Shun Liao, Jake Garrison, Shyam Tailor, Jake Sunshine, Yun Liu, et al. 2024. Scaling Wearable Foundation Models. arXiv preprint arXiv:2410.13638 (2024). source
    • cites [35] 35. Kashif Rasul, Arjun Ashok, Andrew Robert Williams, Arian Khorasani, George Adamopoulos, Rishika Bhagwatkar, Marin Biloš, Hena Ghonia, Nadhir Hassen, Anderson Schneider, et al. 2023. Lag-llama: Towards foundation models for time series forecasting. In R0-FoMo: Robustness of Few-shot and Zero-shot Learning in Large Foundation Models.

      Not checked in the free run.

    • cites [45] 45. Rahul Thapa, Bryan He, Magnus Ruud Kjaer, Hyatt Moore, Gauri Ganjoo, Emmanuel Mignot, and James Zou. [n. d.]. Sleepfm: Multi-modal representation learning for sleep across brain activity, ecg and respiratory signals. 5 2024. arXiv preprint arXiv:2405.17766 ([n. d.]). source
    • cites [60] 60. Shuochao Yao, Shaohan Hu, Yiran Zhao, Aston Zhang, and Tarek Abdelzaher. 2017. Deepsense: A unified deep learning framework for time-series mobile sensing data processing. In Proceedings of the 26th international conference on world wide web. 351-360. source

  • May combine two claims.

    Premisecitation

    Data fragmentation stems from varying sensor characteristics across devices, placement-dependent data variations, diverse sensing specifications, and inconsistent activity taxonomies.

    “This fragmentation stems from varying sensor characteristics across devices (e.g., watches, phones, wristbands), placement-dependent data variations, diverse sensing specifications [38,41,43], and inconsistent activity taxonomies [43].”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [38] 38. Jorge-L Reyes-Ortiz, Luca Oneto, Albert Samà, Xavier Parra, and Davide Anguita. 2016. Transition-aware human activity recognition using smartphones. Neurocomputing 171 (2016), 754-767. source
    • cites [41] 41. Muhammad Shoaib, Stephan Bosch, Ozlem Durmaz Incel, Hans Scholten, and Paul JM Havinga. 2014. Fusion of smartphone motion sensors for physical activity recognition. Sensors 14, 6 (2014), 10146-10176. source

      The Data Collection section (page unavailable) examines five phone placements and reports a 50 Hz sampling rate while noting that other studies use lower rates. The paper therefore supports placement-dependent variation and differences in sensing specifications, but it does not establish that fragmentation stems from differences among watches, phones, and wristbands or from inconsistent activity taxonomies; it describes its selected activities as commonly used in related studies. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [43] 43. Allan Stisen, Henrik Blunck, Sourav Bhattacharya, Thor Siiger Prentow, Mikkel Baun Kjærgaard, Anind Dey, Tobias Sonne, and Mads Møller Jensen. 2015. Smart devices are different: Assessing and mitigatingmobile sensing heterogeneities for activity recognition. In Proceedings of the 13th ACM conference on embedded networked sensor systems. 127-140. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    DDLEARN aims to improve the generalization of the tuned model through adopting Multi-task Learning solutions.

    “DDLEARN aims to improve the generalization of the tuned model through adopting Multi-task Learning solutions.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [33] 33. Xin Qin, Jindong Wang, Shuo Ma, Wang Lu, Yongchun Zhu, Xing Xie, and Yiqiang Chen. 2023. Generalizable Low-Resource Activity Recognition with Diverse and Discriminative Representation Learning. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Long Beach, CA, USA) (KDD '23). Association for Computing Machinery, New York, NY, USA, 1943-1953. doi:10.1145/3580305.3599360 source

  • May combine two claims.

    Interpretationown data

    Incorporating larger datasets and more fine-grained activity descriptions could lead to significant performance improvements.

    “However, the gap between HAR-FM and the other foundation model candidates like UniMTS and YUAN-SSL suggests that incorporating larger datasets (evidenced by the gaps between baseline YUAN-SSL (capture24) and YUAN-SSL w10 ) and more fine-grained activity descriptions (evidenced by the significant performance of UniMTS) could lead to significant performance improvements.”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [62] 62. Yinggang Yu, Dong Wang, Run Zhao, and Qian Zhang. 2019. RFID based real-time recognition of ongoing gesture with adversarial learning. In Proceedings of the 17th Conference on Embedded Networked Sensor Systems. 298-310. source

      The Results (Ablation studies; page not supplied) report that recognition performance increased as the number of unlabelled pre-training subjects grew, supporting the larger-dataset part of the claim. The Discussion credits data volume and diverse activities for improved performance, but our copy does not test finer-grained activity descriptions or compare YUAN-SSL with UniMTS or HAR-FM. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [65] 65. Mi Zhang and Alexander A Sawchuk. 2012. USC-HAD: A daily activity dataset for ubiquitous activity recognition using wearable sensors. In Proceedings of the 2012 ACM conference on ubiquitous computing. 1036-1043.

      Section 4.5 (page unavailable) reports that removing text augmentation reduced UniMTS’s zero-shot performance, supporting the idea that activity descriptions can improve results. But Section 3.3.2 describes adding diverse paraphrases, not testing more fine-grained descriptions. The larger-dataset comparison in the citing sentence concerns YUAN-SSL, not UniMTS. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    LIMU-BERT is a foundation model training framework that proposes to make use of unlabeled data with masked language modeling solution.

    “LIMU-BERT is a foundation model training framework that proposes to make use of the ever common unlabeled data with masked language modeling solution.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [57] 57. Huatao Xu, Pengfei Zhou, Rui Tan, and Mo Li. 2023. Practically Adopting Human Activity Recognition. In Proceedings of the 29th Annual International Conference on Mobile Computing and Networking (Madrid, Spain) (ACM MobiCom '23). Association for Computing Machinery, New York, NY, USA, Article 85, 15 pages. doi:10.1145/3570361.3613299 source

      Section 8 (Related Work) describes LIMU-BERT as an early foundation model for extracting features from unlabeled IMU data. Section 5.1.1 says the authors embed LIMU-BERT in their design, which reconstructs randomly masked IMU readings. The paper does not call this a masked language modeling solution or clearly attribute that specific training method to LIMU-BERT itself. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationcitation

    In zero- and few-shot experiments, larger volume or more precise annotations could contribute to model performance.

    “Secondly, in the zero-and few-shot experiments, Yuan et al. [62] and Zhang et al. [65] demonstrates either larger volume or more precise annotations could contribute to the model performance”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [62] 62. Yinggang Yu, Dong Wang, Run Zhao, and Qian Zhang. 2019. RFID based real-time recognition of ongoing gesture with adversarial learning. In Proceedings of the 17th Conference on Embedded Networked Sensor Systems. 298-310. source

      The Results (Ablation studies; page unavailable) report that downstream performance increased with more unlabelled pre-training subjects, and that additional labelled subjects improved performance most when labels were scarce. This supports the larger-volume finding attributed to Yuan et al., but the paper describes limited-data experiments rather than demonstrating a zero-shot result. The alternative about more precise annotations is not specifically assigned to Yuan et al. in the two-citation sentence. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [65] 65. Mi Zhang and Alexander A Sawchuk. 2012. USC-HAD: A daily activity dataset for ubiquitous activity recognition using wearable sensors. In Proceedings of the 2012 ACM conference on ubiquitous computing. 1036-1043.

      The Introduction (section 1; page unavailable) says data insufficiency negatively affects supervised-model performance, and section 4.3 evaluates few-shot fine-tuning with varying numbers of samples. These passages support the point about data volume, but the paper does not show that more precise annotations improve performance. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Interpretationreasoning

    UniMTS's superior performance against HAR-FM and YUAN-SSL, which are pretrained with wrist-only data, suggests benefits in integrating all placement-wise models.

    “Its superior performance against HAR-FM and YUAN-SSL [62] that are pretrained with data from wrist only, suggests the benefits in integrating all placement-wise models.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [65] 65. Mi Zhang and Alexander A Sawchuk. 2012. USC-HAD: A daily activity dataset for ubiquitous activity recognition using wearable sensors. In Proceedings of the 2012 ACM conference on ubiquitous computing. 1036-1043.

      The Introduction describes UniMTS as trained with coverage of different body locations, and Section 4.2 reports strong generalization across device locations. But our copy does not compare UniMTS with HAR-FM or YUAN-SSL or establish that those models were pretrained on wrist-only data; its single-location comparison concerns ImageBind and IMU2CLIP instead. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [62] 62. Yinggang Yu, Dong Wang, Run Zhao, and Qian Zhang. 2019. RFID based real-time recognition of ongoing gesture with adversarial learning. In Proceedings of the 17th Conference on Embedded Networked Sensor Systems. 298-310. source
    • Rests on an assumption: The modalities share an underlying representation of motion.

  • Classified as probably not a substantial claim. May combine two claims.

    Findingnone given

    The pretrained foundation model is made publicly available along with usage tutorials.

    “The pretrained foundation model is made publicly available along with usage tutorials 1 , which we hope it could facilitate customized HAR applications across different wearable devices in this domain.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Among baseline methods, TSFCN and TRAIN-FROM-SCRATCH demonstrate competitive performance, while the remaining four baselines exhibit significant limitations.

    “Among the baseline methods, while TSFCN and TRAIN-FROM-SCRATCH demonstrate competitive performance, the remaining four baselines exhibit significant limitations due to various shortcomings.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The modular design achieves both parameter and computation efficiency while supporting dynamic input configurations.

    “This approach, as quantified in Table 9 for an 18-class single-IMU case study, achieves both parameter and computation efficiency while supporting dynamic input configurations”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The HAR-FM framework implements the vision of a customizable foundation model through a two-stage training strategy and demonstrates effectiveness in extensive experiments.

    “Our HAR-FM framework implements this vision through a two-stage training strategy, demonstrating effectiveness in extensive experiments”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    TSFCN is unexpectedly competitive on gyroscope data, outperforming all but UniMTS in 1-/2-shot.

    “Interestingly, TSFCN's unexpected competitiveness on gyroscope data (outperforming all but UniMTS in 1-/2-shot)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    HAR-FM achieves the highest averaged metrics in the majority of groups, with top accuracy on 9 and top Macro F1-score on 10 out of 12 groups in the Average panel.

    “In general, HAR-FM achieves the highest averaged metrics in the majority of groups (top accuracy on 9 and top Macro F1-score on 10 out of 12 groups in the Average panel).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Both HAR-FM and UniMTS surpass their Rand-init counterparts on modalities seen during the training phase (all modalities for HAR-FM and accelerometer data for UniMTS).

    “Both HAR-FM and UniMTS, the models surpass their Rand-init counterparts on modalities seen during the training phase (i.e., in all modalities for HAR-FM and accelerometer data for UniMTS).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    UniMTS achieves partial transfer to magnetometer data but fails on gyroscope data.

    “Notably, UniMTS achieves partial transfer to magnetometer data but fails on gyroscoped data”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The customizable modular design supports variant modality input format (e.g., channels and frequency configuration) and demonstrates flexibility to realize more resource-efficient real-world deployment.

    “The customizable modular design (a dataset-wise encoder upon the pretrained context encoder) supports variant modality input format (e.g., channels and frequency configuration) and demonstrates its flexibility to realize more resource-efficient real-world deployment (Table 9).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    TinyHAR and MLP-HAR trade accuracy for computational efficiency, resulting in consistently lower performance.

    “The TinyHAR and MLP-HAR could trade accuracy for computational efficiency, resulting in consistently lower performance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    DDLEARN achieves accuracy comparable to HAR-FM in the accelerometer group under the 1-second window size setting, but this performance fails to generalize across other settings.

    “DDLEARN, which augments the fine-tuning process, achieves accuracy comparable to HAR-FM in the accelerometer group under the 1-second window size setting; however, this performance fails to generalize across other settings.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    TinyHAR is one of the state-of-the-art lightweight deep learning model designs for HAR applications.

    “It refers to one of the state-of-the-art lightweight deep learning model designs for HAR applications.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [71] 71. Yexu Zhou, Tobias King, Haibin Zhao, Yiran Huang, Till Riedel, and Michael Beigl. 2024. MLP-HAR: Boosting Performance and Efficiency of HAR Models on Edge Devices with Purely Fully Connected Layers. In Proceedings of the 2024 ACM International Symposium on Wearable Computers (Melbourne VIC, Australia) (ISWC '24). Association for Computing Machinery, New York, NY, USA, 133-139. doi:10.1145/3675095.3676624 source

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Existing open-source datasets for HAR rarely share identical data collection settings.

    “Specifically, existing open-source datasets for HAR rarely share identical data collection settings.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The diversity in HAR datasets is inherently unavoidable due to the complex nature of human activities and varying sensor placements in real-world scenarios.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    TSFCN is a strong baseline that does not need pretraining.

    “The original paper claims to be a strong baseline with only a few hyper-parameters tuning.It can be considered as one important baseline that does not need pretraining.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [51] 51. Wei Wang, Alex X Liu, Muhammad Shahzad, Kang Ling, and Sanglu Lu. 2015. Understanding and modeling of wifi signal based human activity recognition. In Proceedings of the 21st annual international conference on mobile computing and networking. 65-76. source

  • May combine two claims.

    Interpretationown data

    The HAR-FM's competitive results with a more flexible architecture demonstrate the efficiency and potential of the proposed training strategy for HAR foundation models.

    “its competitive results with a more flexible architecture demonstrate the efficiency and potential of our training strategy for HAR foundation models.”
    • Also stated in: § 1 Introduction · hedged
      “the relatively (few times) smaller datasets used for training and more flexible model structure (section 6) of the HAR-FM should further demonstrate the advancement of the proposed training strategy.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The study adopted a 5-fold cross-user validation setting.

    “Throughout the experiments, we adopt a 5-fold cross-user validation setting.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    The semantic alignment task bridges gaps across different contexts and datasets.

    “To address this limitation, we introduce a semantic alignment task to bridge gaps across different contexts and datasets.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims. The quote only partly states this claim.

    Study design and conductown data

    Five datasets (Dataset No. 5-9) were not used in training the foundation model and were used only for fine-tuning to simulate downstream tasks.

    “We prepared five datasets (Dataset No. 5-9 in Table 1, which were not used in training the foundation model) to simulate downstream tasks and evaluate the performance of the foundation model.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given

    The construction of downstream tasks aligns with each dataset's intended purpose, ensuring precise context information preservation from the dataset contributors' perspective.

    “In supervised learning, the construction of downstream tasks aligns with each dataset's intended purpose.This alignment ensures precise context information preservation from the dataset contributors' perspective.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    Placement-aware training ensures that the model performs robustly across different configurations and deployment scenarios.

    “By accounting for device placement variability during training, placement-aware training ensures that the model performs robustly across different configurations and deployment scenarios.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Wearable sensors located on the wrists provide consistent and comparable data across devices.

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationnone given

    The foundation model could be directly deployed with a simple fine-tuning pipeline and equipped with a zero-shot pipeline.

    “Our foundation model could be directly deployed with simple fine-tuning pipeline and equipped with zero-shot pipeline listed as follows.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The deployment of the foundation model supports full customization with one's own device and scenes related task definition.

    “The deployment of the foundation model (C) supports full customization with one's own device and scenes related task definition (detailed in Section 4.3).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given

    The proposed 5-fold cross-user validation setting closely resembles the practical downstream deployment pipeline, making it a reliable reflection of real-world performance.

    “The proposed 5-fold cross-user validation setting closely resembles the practical downstream deployment pipeline of related techniques, making it a reliable reflection of real-world performance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The accelerometer data's richer and lower-variance nature, combined with increased subject diversity, enables all models to achieve better performance, thus narrowing the performance gap.

    “This suggests that the accelerometer data's richer and lower-variance nature, combined with increased subject diversity, enables all models to achieve better performance, thus narrowing the performance gap.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Incorporating downstream task alignment is significant and facilitates the model's adaptation to real-world applications.

    “emphasizing the significance of incorporating downstream task alignment, which facilitates the model's adaptation to real-world applications.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    MLP-HAR is another lightweight deep learning model that adopts purely fully connected layers for efficiency of HAR models on edge devices.

    “It refers to the another lightweight deep learning model that adopts purely fully connected layers for efficiency of HAR models on edge devices.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [70] 70. Tian Zhou, Peisong Niu, Liang Sun, Rong Jin, et al. 2023. One fits all: Power general time series analysis by pretrained lm. Advances in neural information processing systems 36 (2023), 43322-43355. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationown data

    The main configuration settings demonstrate the customizability of the HAR-FM learning framework.

    “The main configuration settings are listed to demonstrate the customizability of our learning framework.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Subject-wise data partitioning for 5-fold cross-validation was used across benchmark datasets.

    “Subject-wise data partitioning for 5-fold cross-validation across benchmark datasets.”
    Quote in the paperNot a substantial claim
35 of 35

  • The paper relies on this without support.

    HAR systems fall short in generalization and often require extensive task-specific data collection and retraining.

    “Human Activity Recognition (HAR) systems fall short in generalization and often require extensive task-specific data collection and retraining.”

  • The paper relies on this without support.

    Wearable sensors located on the wrists provide consistent and comparable data across devices.

    “For instance, wearable sensors located on the wrists provide consistent and comparable data across devices.”

  • The paper relies on this without support.

    The diversity in HAR datasets is inherently unavoidable due to the complex nature of human activities and varying sensor placements in real-world scenarios.

    “Such diversity is inherently unavoidable due to the complex nature of human activities and varying sensor placements in real-world scenarios.”

  • The paper relies on this without support.

    Models trained on one type of dataset often perform suboptimally on others.

    “This intrinsic complexity and variability in HAR applications presents a significant challenge, as models trained on one type of dataset often perform suboptimally on others.”

  • The paper relies on this without support.

    Foundation models for human activity recognition remain at an early developmental stage.

    “while acknowledging that foundation models for HAR remain at an early developmental stage.”

  • The paper relies on this without support.

    The modalities share an underlying representation of motion.

    “Other than better explanation (that the modalities might share underlying representation of motion)”

  • The paper relies on this without support.

    Imbalanced data is a well-known challenge in the research community, often exacerbated by the sensitivity of sensor data or by the novelty of sensing techniques.

    “Imbalanced data is a well-known challenge in the research community, often exacerbated by the sensitivity of sensor data, which limits public sharing, or by the novelty of sensing techniques, which restricts widespread deployment.”
7 of 7

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Yuan-SSL sampling rate is given as both 20 Hz and 30 Hz

    Minor

    The paper describes Yuan-SSL's pretrained inputs as 30 Hz, but the baseline setup says its 10-second input requires 20 Hz.

  • Yuan-SSL citation points to a different paper

    Minor

    The text cites reference 62 for Yuan's wearable-data foundation model, but entry 62 is a paper by Yu and colleagues about RFID gesture recognition.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 3

  1. Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33 (2020), 12449--12460.
  2. Kashif Rasul, Arjun Ashok, Andrew Robert Williams, Arian Khorasani, George Adamopoulos, Rishika Bhagwatkar, Marin Biloš, Hena Ghonia, Nadhir Hassen, Anderson Schneider, et al. 2023. Lag-llama: Towards foundation models for time series forecasting. In R0-FoMo: Robustness of Few-shot and Zero-shot Learning in Large Foundation Models.
  3. 10.5555/3295222.3295349 · DOI

Cited by

Sorted by citations

Related

Authors
Qiu, M., Weng, C., Fan, M., & Wu, K.
Affiliations
The Hong Kong University of Science and Technology (Guangzhou) (CN)
ORCID
Venue
Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies
Publisher
Association for Computing Machinery (ACM)
Published
2025-09-03
Volume
9
Issue
3
Pages
1-29
ISSN
2474-9567 (ISSN-L)
DOI
10.1145/3749479
OpenAlex
W4414041781
Cited by
4
References
54
Citations by year
2026: 4
FWCI
1.41
Type
journal article
Language
en
Open access
Licence
public-domain
Locations
  • Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies (publishedVersion, public-domain)
  • Rare & Special e-Zone (The Hong Kong University of Science and Technology) (submittedVersion)
Funders
  • Guangdong Provincial Key Lab of Integrated Communication, Sensing and Computation for Ubiquitous Internet of Things (2023B1212010007)
  • China NSFC Grant (62472366)
  • the Project of DEGP (2024GCZX003, 2023KCXTD042)
  • 111 Center (D25008)
  • Shenzhen Science and Technology Foundation (ZDSYS20190902092853047)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Physical Sciences › Computer Science › Signal Processing › Time Series Analysis and Forecasting
Keywords
human activity recognition, foundation models, wearable devices, heterogeneous sensor data, self-supervised learning, multi-task learning, fine-tuning, few-shot learning, zero-shot learning, sensor modalities

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-11.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-11.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review