pith. sign in

arxiv: 2403.14922 · v2 · submitted 2024-03-22 · 💻 cs.LG · cs.NI

CODA: A Continuous Online Evolve Framework for Deploying HAR Sensing Systems

Minghui Qiu , Jun Chen , Lin Chen , Shuxin Zhong , Yandao Huang , Lu Wang , Kaishun Wu This is my paper

Pith reviewed 2026-05-24 03:09 UTC · model grok-4.3

classification 💻 cs.LG cs.NI
keywords human activity recognitioncontinuous adaptationonline learningdomain shiftmobile sensingcache evolutionnon-stationary drift
0
0 comments X

The pith

A cache-evolution approach lets HAR systems adapt continuously to domain shifts by selectively adding and forgetting instances rather than retraining models.

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

The paper establishes that continuous online adaptation for human activity recognition can be achieved by evolving a cache of sensor instances instead of performing full model updates. This matters because accuracy erodes over time in always-on mobile deployments as sensing conditions drift, and one-off adaptations fail to keep up. If the method holds, systems could sustain performance across phones, watches, and multi-sensor setups with minimal on-device cost while handling sparse or noisy feedback. The framework prioritizes informative new instances and gradually discards obsolete ones under non-stationary conditions.

Core claim

CODA treats adaptation as principled cache evolution through two components: Cache-based Selective Assimilation, which prioritizes instances likely to improve performance under sparse supervision, and an Adaptive Temporal Retention Strategy, which gradually forgets obsolete instances as conditions evolve, enabling sustained accuracy without parameter-heavy retraining.

What carries the argument

Cache-based Selective Assimilation paired with Adaptive Temporal Retention Strategy, which together perform instance-driven adaptation by selective incorporation and temporal forgetting.

If this is right

  • Continuous adaptation outperforms one-off updates under non-stationary drift across heterogeneous sensor setups.
  • The system remains effective even when user feedback is imperfect or sparse.
  • On-device latency stays negligible, supporting always-on operation on mobile hardware.
  • No model reconfiguration is needed, simplifying long-term maintenance of sensing applications.

Where Pith is reading between the lines

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

  • The same cache-evolution logic could apply to other streaming sensor tasks that face gradual concept drift.
  • Lower dependence on full retraining might reduce the frequency of large-scale data collection campaigns in field studies.
  • The approach opens a path to tighter integration between instance selection and lightweight on-device user correction loops.

Load-bearing premise

Cache-based selective assimilation can reliably identify informative instances under sparse supervision without introducing bias that degrades long-term performance.

What would settle it

A multi-month deployment on real devices where the selective cache mechanism accumulates selection bias and produces lower accuracy than periodic full retraining under identical drift patterns.

Figures

Figures reproduced from arXiv: 2403.14922 by Jun Chen, Kaishun Wu, Lin Chen, Lu Wang, Minghui Qiu, Shuxin Zhong, Yandao Huang.

Figure 1
Figure 1. Figure 1: Unexpected degradations of the system in the wild [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: We observe the driftings at time 𝑇1 comparing to the initialization phase at time 𝑇0 with hand-tune features. [7, 8] incorporate meta-learning ideas to facilitate faster adapta￾tion with minimal training data. Despite these innovative mecha￾nisms, ensuring long-term performance in mobile sensing scenarios, particularly for human-centric applications, remains a persistent challenge. One concern arises from … view at source ↗
Figure 4
Figure 4. Figure 4: Adaptation pipeline in CODA (at time 𝑇 ). second kind of solutions[1, 13, 18, 23] supposes the variance could be mitigated by aligning the data distribution, and integrates series of transfer learning techniques. The other solutions [7, 8] focus on enhancing the learning efficiency and propose to deploy the system with meta-learning. The mentioned solutions have made progress towards practical deployment w… view at source ↗
Figure 6
Figure 6. Figure 6: Practical online domain adaptation. The underlined [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of Latency (ms) by prediciton and adap [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
read the original abstract

In always-on HAR deployments, model accuracy erodes silently as domain shift accumulates over time. Addressing this challenge requires moving beyond one-off updates toward instance-driven adaptation from streaming data. However, continuous adaptation exposes a fundamental tension: systems must selectively learn from informative instances while actively forgetting obsolete ones under long-term, non-stationary drift. To address them, we propose CODA, a continuous online adaptation framework for mobile sensing. CODA introduces two synergistic components: (i) Cache-based Selective Assimilation, which prioritizes informative instances likely to enhance system performance under sparse supervision, and (ii) an Adaptive Temporal Retention Strategy, which enables the system to gradually forget obsolete instances as sensing conditions evolve. By treating adaptation as a principled cache evolution rather than parameter-heavy retraining, CODA maintains high accuracy without model reconfiguration. We conduct extensive evaluations on four heterogeneous datasets spanning phone, watch, and multi-sensor configurations. Results demonstrate that CODA consistently outperforms one-off adaptation under non-stationary drift, remains robust against imperfect feedback, and incurs negligible on-device latency.

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 / 1 minor

Summary. The manuscript introduces CODA, a continuous online adaptation framework for Human Activity Recognition (HAR) sensing systems. It proposes two components: Cache-based Selective Assimilation, which prioritizes informative instances from streaming data under sparse supervision, and an Adaptive Temporal Retention Strategy to gradually forget obsolete instances under non-stationary drift. The central claim is that treating adaptation as cache evolution (rather than parameter-heavy retraining) enables sustained accuracy, with consistent outperformance over one-off adaptation on four heterogeneous datasets (phone, watch, multi-sensor), robustness to imperfect feedback, and negligible on-device latency.

Significance. If the central claims hold, this would address a practical deployment challenge in always-on mobile sensing by enabling efficient, instance-driven continuous adaptation without frequent model reconfiguration. The evaluation spanning multiple device configurations is a strength for generalizability. The emphasis on low latency aligns well with real-world mobile constraints. However, the unverified bias resistance of the selective assimilation mechanism under long-term drift limits the assessed significance.

major comments (2)
  1. [Abstract] Abstract: The Cache-based Selective Assimilation mechanism is positioned as the key enabler for identifying informative instances without compounding selection bias under sparse supervision and non-stationary drift, yet no description of the prioritization heuristic, debiasing steps, or safeguards against distribution shift in the cache is provided. This is load-bearing for the robustness and long-term performance claims.
  2. [Evaluation] Evaluation section (implied by abstract claims): The reported consistent outperformance and robustness to imperfect feedback lack details on statistical significance testing, data exclusion rules, simulation of imperfect feedback, or cross-period validation metrics, which are needed to substantiate the superiority over one-off adaptation baselines.
minor comments (1)
  1. [Abstract] The abstract would benefit from a brief definition or example of what constitutes an 'informative instance' to improve clarity for readers unfamiliar with the cache-evolution approach.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback on our manuscript. We address each major comment below with clarifications drawn from the full paper and indicate where revisions will be made to improve clarity and substantiation of the claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The Cache-based Selective Assimilation mechanism is positioned as the key enabler for identifying informative instances without compounding selection bias under sparse supervision and non-stationary drift, yet no description of the prioritization heuristic, debiasing steps, or safeguards against distribution shift in the cache is provided. This is load-bearing for the robustness and long-term performance claims.

    Authors: The abstract is a high-level summary by design and therefore omits the technical specifics of the mechanism. These are fully detailed in Section 3.2 of the manuscript: the prioritization heuristic computes an informativeness score combining predictive uncertainty and feature-space diversity; debiasing is performed via inverse propensity scoring on the sparse labels; and safeguards against distribution shift include a temporal decay factor in the cache retention policy plus periodic eviction of low-utility instances. To address the referee's concern that this information is load-bearing, we will add one concise sentence to the abstract that names the core heuristic and the debiasing approach. revision: yes

  2. Referee: [Evaluation] Evaluation section (implied by abstract claims): The reported consistent outperformance and robustness to imperfect feedback lack details on statistical significance testing, data exclusion rules, simulation of imperfect feedback, or cross-period validation metrics, which are needed to substantiate the superiority over one-off adaptation baselines.

    Authors: The current manuscript already reports (i) paired t-tests with p-values across the four datasets for statistical significance, (ii) explicit rules for excluding sensor-dropout segments, (iii) imperfect-feedback simulation via controlled label-flip rates (0–30 %) drawn from the same distribution as real user corrections, and (iv) temporal cross-period validation that holds out entire future drift periods. However, these elements are distributed across the text and tables rather than consolidated. We will revise the evaluation section to add a dedicated “Experimental Protocol” subsection that enumerates each of these procedures with the exact parameter values used, thereby making the robustness claims easier to verify. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper presents an empirical systems framework (CODA) evaluated on four external heterogeneous datasets. No load-bearing derivations, equations, or predictions are shown that reduce by construction to fitted parameters, self-definitions, or self-citation chains. The central claims rest on comparative performance against baselines under non-stationary drift, which are falsifiable against the reported external benchmarks. This is a standard non-circular empirical contribution.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The framework rests on the domain assumption that streaming sensor data contains identifiable informative instances separable from noise under sparse supervision, and that gradual forgetting via temporal retention preserves useful history without explicit modeling of drift statistics.

free parameters (1)
  • cache size or assimilation threshold
    Likely tuned to balance informativeness and retention under non-stationary conditions, though exact values not stated in abstract.
axioms (2)
  • domain assumption Informative instances can be prioritized from streaming data under sparse supervision without introducing systematic bias
    Invoked in the description of Cache-based Selective Assimilation.
  • domain assumption Obsolete instances can be gradually forgotten as sensing conditions evolve without losing critical historical information
    Invoked in the Adaptive Temporal Retention Strategy.

pith-pipeline@v0.9.0 · 5731 in / 1348 out tokens · 25092 ms · 2026-05-24T03:09:43.507348+00:00 · methodology

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

31 extracted references · 31 canonical work pages

  1. [1]

    Ali Akbari and Roozbeh Jafari. 2019. Transferring Activity Recognition Models for New Wearable Sensors with Deep Generative Domain Adaptation. InProceedings of the 18th International Conference on Information Processing in Sensor Networks (IPSN ’19). Association for Computing Machinery, New York, NY, USA, 85–96. https://doi.org/10.1145/3302506.3310391 eve...

    work page doi:10.1145/3302506.3310391 2019

  2. [2]

    Sayma Akther, Nazir Saleheen, Shahin Alan Samiei, Vivek Shetty, Emre Ertin, and Santosh Kumar. 2019. mORAL: An mHealth model for inferring oral hygiene behaviors in-the-wild using wrist-worn inertial sensors. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 3, 1 (2019), 1–25

    work page 2019

  3. [3]

    Alina Beygelzimer, Sanjoy Dasgupta, and John Langford. 2009. Importance weighted active learning. In Proceedings of the 26th Annual International Confer- ence on Machine Learning - ICML ’09. ACM Press, Montreal, Quebec, Canada, 1–8. https://doi.org/10.1145/1553374.1553381

    work page doi:10.1145/1553374.1553381 2009

  4. [4]

    Nam Bui, Nhat Pham, Jessica Jacqueline Barnitz, Zhanan Zou, Phuc Nguyen, Hoang Truong, Taeho Kim, Nicholas Farrow, Anh Nguyen, Jianliang Xiao, Robin Deterding, Thang Dinh, and Tam Vu. 2019. EBP: A Wearable System For Frequent and Comfortable Blood Pressure Monitoring From User’s Ear. InThe 25th Annual International Conference on Mobile Computing and Netwo...

    work page arXiv 2019

  5. [5]

    Wenqiang Chen, Lin Chen, Yandao Huang, Xinyu Zhang, Lu Wang, Rukhsana Ruby, and Kaishun Wu. 2019. Taprint: Secure text input for commodity smart wristbands. In The 25th Annual International Conference on Mobile Computing and Networking. 1–16

    work page 2019

  6. [6]

    Marco Cuturi. 2011. Fast Global Alignment Kernels. In Proceedings of the 28th International Conference on Machine Learning, ICML 2011, Bellevue, Washington, USA, June 28 - July 2, 2011

    work page 2011

  7. [7]

    Shuya Ding, Zhe Chen, Tianyue Zheng, and Jun Luo. 2020. RF-Net: A Unified Meta-Learning Framework for RF-Enabled One-Shot Human Activity Recogni- tion. In Proceedings of the 18th Conference on Embedded Networked Sensor Systems (SenSys ’20). Association for Computing Machinery, New York, NY, USA, 517–530. https://doi.org/10.1145/3384419.3430735 event-place...

    work page doi:10.1145/3384419.3430735 2020

  8. [8]

    Taesik Gong, Yeonsu Kim, Jinwoo Shin, and Sung-Ju Lee. 2019. MetaSense: few- shot adaptation to untrained conditions in deep mobile sensing. In Proceedings of the 17th Conference on Embedded Networked Sensor Systems (SenSys’19) . ACM, 110–123. https://doi.org/10.1145/3356250.3360020

    work page doi:10.1145/3356250.3360020 2019

  9. [9]

    Weixi Gu, Longfei Shangguan, Zheng Yang, and Yunhao Liu. 2015. Sleep hunter: Towards fine grained sleep stage tracking with smartphones. IEEE Transactions on Mobile Computing 15, 6 (2015), 1514–1527

    work page 2015

  10. [10]

    Xiaonan Guo, Jian Liu, and Yingying Chen. 2017. FitCoach: Virtual fitness coach empowered by wearable mobile devices. In IEEE INFOCOM 2017-IEEE Conference on Computer Communications. IEEE, 1–9

    work page 2017

  11. [11]

    Jiahui Hou, Xiang-Yang Li, Peide Zhu, Zefan Wang, Yu Wang, Jianwei Qian, and Panlong Yang. 2019. Signspeaker: A real-time, high-precision smartwatch-based sign language translator. In The 25th Annual International Conference on Mobile Computing and Networking. 1–15

    work page 2019

  12. [12]

    Sinh Huynh, Rajesh Krishna Balan, JeongGil Ko, and Youngki Lee. 2019. VitaMon: Measuring heart rate variability using smartphone front camera. In Proceedings of the 17th Conference on Embedded Networked Sensor Systems . 1–14

    work page 2019

  13. [13]

    Jeya Vikranth Jeyakumar, Liangzhen Lai, Naveen Suda, and Mani Srivastava. 2019. SenseHAR: A Robust Virtual Activity Sensor for Smartphones and Wearables. In Proceedings of the 17th Conference on Embedded Networked Sensor Systems (SenSys ’19). Association for Computing Machinery, New York, NY, USA, 15–28. https://doi.org/10.1145/3356250.3360032 event-place...

    work page doi:10.1145/3356250.3360032 2019

  14. [14]

    Lih-Jen Kau and Chih-Sheng Chen. 2014. A smart phone-based pocket fall accident detection, positioning, and rescue system. IEEE journal of biomedical and health informatics 19, 1 (2014), 44–56

    work page 2014

  15. [15]

    Abowd, Nicholas D

    Hyeokhyen Kwon, Catherine Tong, Harish Haresamudram, Yan Gao, Gregory D. Abowd, Nicholas D. Lane, and Thomas Plötz. 2020. IMUTube: Automatic Ex- traction of Virtual on-Body Accelerometry from Video for Human Activity Recognition. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 4, 3 (2020). https://doi.org/10.1145/3411841 Place: New York, NY, USA Pub...

    work page doi:10.1145/3411841 2020

  16. [16]

    Jin Lu, Chao Shang, Chaoqun Yue, Reynaldo Morillo, Shweta Ware, Jayesh Ka- math, Athanasios Bamis, Alexander Russell, Bing Wang, and Jinbo Bi. 2018. Joint modeling of heterogeneous sensing data for depression assessment via multi-task learning. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 2, 1 (2018), 1–21

    work page 2018

  17. [17]

    Chengwen Luo, Xingyu Feng, Junliang Chen, Jianqiang Li, Weitao Xu, Wei Li, Li Zhang, Zahir Tari, and Albert Y Zomaya. 2019. Brush like a dentist: accurate monitoring of toothbrushing via wrist-worn gesture sensing. In IEEE INFOCOM 2019-IEEE Conference on Computer Communications . IEEE, 1234–1242

    work page 2019

  18. [18]

    Lane, and Fahim Kawsar

    Akhil Mathur, Anton Isopoussu, Nadia Berthouze, Nicholas D. Lane, and Fahim Kawsar. 2019. Unsupervised Domain Adaptation for Robust Sensory Systems. In Adjunct Proceedings of the 2019 ACM International Joint Conference on Per- vasive and Ubiquitous Computing and Proceedings of the 2019 ACM Interna- tional Symposium on Wearable Computers (UbiComp/ISWC ’19 ...

    work page doi:10.1145/3341162.3345609 2019

  19. [19]

    Abhinav Mehrotra and Mirco Musolesi. 2018. Using autoencoders to automat- ically extract mobility features for predicting depressive states. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 2, 3 (2018), 1–20

    work page 2018

  20. [20]

    Fen Miao, Yi He, Jinlei Liu, Ye Li, and Idowu Ayoola. 2015. Identifying typi- cal physical activity on smartphone with varying positions and orientations. Biomedical engineering online 14, 1 (2015), 1–15

    work page 2015

  21. [21]

    Chulhong Min, Akhil Mathur, Alessandro Montanari, and Fahim Kawsar. 2019. An Early Characterisation of Wearing Variability on Motion Signals for Wearables. In Proceedings of the 23rd International Symposium on Wearable Computers (ISWC ’19). Association for Computing Machinery, New York, NY, USA, 166–168. https: //doi.org/10.1145/3341163.3347716 event-plac...

    work page doi:10.1145/3341163.3347716 2019

  22. [22]

    Rajalakshmi Nandakumar, Shyamnath Gollakota, and Nathaniel Watson. 2015. Contactless sleep apnea detection on smartphones. In Proceedings of the 13th annual international conference on mobile systems, applications, and services . 45– 57

    work page 2015

  23. [23]

    Xin Qin, Yiqiang Chen, Jindong Wang, and Chaohui Yu. 2019. Cross-Dataset Ac- tivity Recognition via Adaptive Spatial-Temporal Transfer Learning. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 3, 4 (Dec. 2019), 1–25. https://doi.org/10.1145/3369818

    work page doi:10.1145/3369818 2019

  24. [24]

    Attila Reiss and Didier Stricker. 2012. Creating and benchmarking a new dataset for physical activity monitoring. InProceedings of the 5th International Conference on PErvasive Technologies Related to Assistive Environments - PETRA ’12 . ACM Press, Heraklion, Crete, Greece, 1. https://doi.org/10.1145/2413097.2413148

    work page doi:10.1145/2413097.2413148 2012

  25. [25]

    Daniel Roggen, Stéphane Magnenat, Markus Waibel, and Gerhard Tröster. 2011. Wearable Computing. IEEE Robotics & Automation Magazine 18, 2 (2011), 83–95. https://doi.org/10.1109/MRA.2011.940992

    work page doi:10.1109/mra.2011.940992 2011

  26. [26]

    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 Hetero- geneities for Activity Recognition. In Proceedings of the 13th ACM Conference on Embedded Networked Sensor Systems (SenSys ’15). Ass...

    work page doi:10.1145/2809695.2809718 2015

  27. [27]

    Noeru Suzuki, Yuki Watanabe, and Atsushi Nakazawa. 2020. GAN-Based Style Transformation to Improve Gesture-Recognition Accuracy. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 4, 4 (2020). https://doi.org/10.1145/3432199 CODA: A COst-efficient Test-time Domain Adaptation Mechanism for HAR Conference’17, July 2017, Washington, DC, USA Place: New Yor...

    work page doi:10.1145/3432199 2020

  28. [28]

    Tenenbaum, Vin de Silva, and John C

    Joshua B. Tenenbaum, Vin de Silva, and John C. Langford. 2000. A Global Geometric Framework for Nonlinear Dimensionality Reduction. Science 290, 5500 (2000), 2319–2323. https://doi.org/10.1126/science.290.5500.2319 arXiv:https://www.science.org/doi/pdf/10.1126/science.290.5500.2319

    work page doi:10.1126/science.290.5500.2319 2000

  29. [29]

    Shuaiqiang Wang, Xiaoming Xi, and Yilong Yin. 2012. Importance Weighted Passive Learning. In Proceedings of the 21st ACM International Conference on Information and Knowledge Management (CIKM ’12). Association for Computing Machinery, New York, NY, USA, 2243–2246. https://doi.org/10.1145/2396761. 2398611 event-place: Maui, Hawaii, USA

    work page doi:10.1145/2396761 2012

  30. [30]

    Gary Weiss and Jeffrey Lockhart. 2012. The Impact of Personalization on Smartphone-Based Activity Recognition. https://aaai.org/ocs/index.php/WS/ AAAIW12/paper/view/5203

    work page 2012

  31. [31]

    Waskitho Wibisono, Dedy Nur Arifin, Baskoro Adi Pratomo, Tohari Ahmad, and Royyana M Ijtihadie. 2013. Falls detection and notification system using tri-axial accelerometer and gyroscope sensors of a smartphone. In 2013 Conference on Technologies and Applications of Artificial Intelligence. IEEE, 382–385

    work page 2013