pith. sign in

arxiv: 2602.06523 · v3 · pith:EAXWJMPEnew · submitted 2026-02-06 · 💻 cs.CV · cs.HC

MicroBi-ConvLSTM: An Ultra-Lightweight Efficient Model for Human Activity Recognition on Resource Constrained Devices

Mridankan Mandal This is my paper

Pith reviewed 2026-05-21 13:20 UTC · model grok-4.3

classification 💻 cs.CV cs.HC
keywords Human Activity RecognitionUltra-lightweight modelsMicrocontrollersConvLSTMResource constrained devicesOn-device AIParameter efficiency
0
0 comments X

The pith

MicroBi-ConvLSTM achieves competitive human activity recognition on microcontrollers with only 11.4K parameters on average.

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

The paper presents MicroBi-ConvLSTM as an architecture for human activity recognition that uses far fewer parameters than existing lightweight models to fit on devices with limited memory. It combines two stages of convolutional feature extraction, 4x temporal pooling, and a single bidirectional LSTM layer to reach 11.4K parameters while keeping linear computational complexity. This allows the model to run on microcontrollers where previous approaches exceed SRAM budgets after operating system overhead. Tests across eight benchmarks demonstrate competitive accuracy, including 93.41% macro F1 on UCI-HAR, and successful on-device deployment on the Raspberry Pi Pico 2 and ESP32 under both quantized and full precision.

Core claim

MicroBi-ConvLSTM is an ultra-lightweight convolutional recurrent model that achieves an average of 11.4K parameters through two-stage convolutional feature extraction with 4x temporal pooling and a single bidirectional LSTM layer, delivering 2.9x parameter reduction compared to TinierHAR and 11.9x versus DeepConvLSTM while maintaining competitive performance on human activity recognition tasks and enabling full deployment on resource-constrained hardware.

What carries the argument

The two-stage convolutional feature extraction with 4x temporal pooling combined with a single bidirectional LSTM layer that extracts features efficiently before recurrent processing.

If this is right

  • Full 8/8 dataset coverage on both Raspberry Pi Pico 2 and ESP32 under INT8 quantization.
  • 72.8 ms average latency on the Pico 2 with INT8.
  • 97.9% PyTorch parity on ESP32 under INT8 and 100% under FP32 on successful runs.
  • Bidirectionality provides benefits mainly for episodic event detection tasks rather than periodic ones.
  • The architecture itself has no inherent limitation causing fidelity loss, as all degradation comes from quantization.

Where Pith is reading between the lines

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

  • Similar parameter reduction techniques could apply to other real-time sensor processing tasks on edge devices beyond activity recognition.
  • Lower parameter counts like this may enable longer battery life in always-on wearable monitoring systems.
  • Testing the model on additional microcontroller platforms could reveal broader hardware compatibility.
  • The task-dependent role of bidirectionality suggests tailoring recurrent components based on activity type for further optimization.

Load-bearing premise

Prior lightweight models exceed available SRAM on microcontrollers once operating system overhead is taken into account.

What would settle it

Running MicroBi-ConvLSTM and competing models on an ESP32 to verify if only this model achieves complete 8/8 dataset coverage under INT8 quantization.

Figures

Figures reproduced from arXiv: 2602.06523 by Mridankan Mandal.

Figure 1
Figure 1. Figure 1: µBi-ConvLSTM architecture overview. Input sensor signals (C channels × T timesteps) pass through two convolutional blocks with batch normalization, ReLU activation, and 2× max pooling each, achieving 4× total temporal compression. A single bidirectional LSTM (hidden dimension 24) processes the compressed sequence, with the final timestep representation feeding the classification head. Parameter count varie… view at source ↗
Figure 2
Figure 2. Figure 2: Parameters, MACs, FLOPs, Model Size (in KB), F1-score per million MACs, and F1-score per thousand parameters distributions across architectures and datasets. Box plots show mean, and standard deviations across five random seeds. µBi-ConvLSTM (leftmost in each group) maintains competitive variance despite 2.9× fewer parameters than TinierHAR. TABLE VI: INT8 Quantization Impact on µBi-ConvLSTM Dataset FP32 F… view at source ↗
Figure 5
Figure 5. Figure 5: FP32 versus INT8 F1-scores for uBi-ConvLSTM across datasets. The near diagonal alignment demonstrates quantization robustness, with average degradation of only 0.21%. Temporal Compression Value: Removing max pooling (A1) increases MACs by 3.1× with inconsistent accuracy effects, validating the aggressive temporal compression strat￾egy. The pooling layers provide regularization that benefits generalization … view at source ↗
Figure 4
Figure 4. Figure 4: Efficiency heatmap: F1-score per thousand parameters across datasets. Darker cells indicate higher parameter efficiency. µBi-ConvLSTM advantage is consistent across benchmarks rather than dataset-specific. minimum), whereas TinyHAR increases to 305 KB (411% increase) due to attention’s quadratic channel scaling. This characteristic makes µBi-ConvLSTM particularly suitable for multi-sensor wearable platform… view at source ↗
Figure 6
Figure 6. Figure 6: Ablation results across variants (A0–A4) and datasets. A0: Base configuration, A1: No pooling, A2: Unidirectional LSTM, A3: Single conv block, and A4: Mean pooling aggregation [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: F1-score differences from base configuration (A0) across ablation variants and datasets. The base configuration achieves best or near best performance on 5 of 8 datasets. B. Limitations Class Imbalance Sensitivity: On PAMAP2, µBi￾ConvLSTM shows a 13.3% F1-score gap as compared to TinierHAR due to extreme class imbalance where rare activities (rope jumping, cycling) constitute <10% of samples, causing ultra… view at source ↗
read the original abstract

Human Activity Recognition (HAR) on resource constrained wearables requires models that balance accuracy against strict memory and computational budgets. State of the art lightweight architectures such as TinierHAR (34K parameters) and TinyHAR (55K parameters) achieve strong accuracy, but exceed memory budgets of microcontrollers with limited SRAM once operating system overhead is considered. We present MicroBi-ConvLSTM, an ultra-lightweight convolutional recurrent architecture achieving 11.4K parameters on average through two stage convolutional feature extraction with 4x temporal pooling, and a single bidirectional LSTM layer. This represents 2.9x parameter reduction versus TinierHAR and 11.9x versus DeepConvLSTM while preserving linear O(N) complexity. Evaluation across eight diverse HAR benchmarks shows that MicroBi-ConvLSTM maintains competitive performance within the ultra-lightweight regime: 93.41% macro F1 on UCI-HAR, 94.46% on SKODA assembly gestures, and 88.98% on Daphnet gait freeze detection. Systematic ablation reveals task dependent component contributions where bidirectionality benefits episodic event detection, but provides marginal gains on periodic locomotion. On-device deployment on the Raspberry Pi Pico 2 and ESP32 validates hardware viability under both INT8 quantized and FP32 full-precision paths. Under INT8 quantization, MicroBi-ConvLSTM is the only architecture achieving full 8/8 dataset coverage on both platforms, with 72.8 ms average latency on Pico 2 and 97.9% PyTorch parity on ESP32. Under FP32 deployment, it achieves 100.0% parity on all successful configurations (8/8 Pico 2, 7/8 ESP32), confirming that all INT8 fidelity degradation is a quantization artifact rather than an architectural limitation.

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

1 major / 2 minor

Summary. The manuscript presents MicroBi-ConvLSTM, an ultra-lightweight convolutional recurrent architecture for human activity recognition on resource-constrained devices. It reports an average of 11.4K parameters via two-stage convolutional feature extraction with 4x temporal pooling and a single bidirectional LSTM layer, claiming 2.9x parameter reduction versus TinierHAR and 11.9x versus DeepConvLSTM while preserving O(N) complexity. Across eight HAR benchmarks it achieves competitive accuracies (e.g., 93.41% macro F1 on UCI-HAR, 94.46% on SKODA, 88.98% on Daphnet) and, on Raspberry Pi Pico 2 and ESP32, is the only model to reach full 8/8 dataset coverage under INT8 quantization with reported latencies and PyTorch parity; FP32 results confirm quantization as the source of any fidelity loss.

Significance. If the on-device memory and coverage claims hold, the work supplies a concrete, deployable ultra-lightweight HAR architecture that fits within tight microcontroller SRAM limits where prior models reportedly do not. The ablation findings on bidirectionality, the INT8/FP32 parity comparison, and the multi-platform, multi-dataset evaluation add practical value for resource-constrained wearable applications.

major comments (1)
  1. [On-device deployment results] On-device deployment results: the central claim that MicroBi-ConvLSTM is the sole architecture achieving full 8/8 dataset coverage on Pico 2 and ESP32 under INT8 because TinierHAR (34K) and TinyHAR (55K) exceed SRAM budgets once OS overhead is included is unsupported. No tabulated peak SRAM measurements (including overhead) for the reproduced baselines, no explicit overhead value, and no side-by-side memory profiling under identical compilation/quantization settings are provided. This measurement gap directly weakens the uniqueness and necessity arguments that motivate the 11.4K-parameter design.
minor comments (2)
  1. [Evaluation] Evaluation: the reported macro F1 scores and accuracy figures are presented without error bars, standard deviations across runs, or statistical significance tests, making it difficult to judge whether observed differences versus baselines are reliable.
  2. [Methods] Methods: full training details (optimizer, learning-rate schedule, batch size, epoch count, exact preprocessing and augmentation per dataset) are not supplied, which limits reproducibility of the accuracy and ablation results.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the practical value of the multi-platform, multi-dataset evaluation. We address the single major comment below and will revise the manuscript to strengthen the supporting evidence for the on-device claims.

read point-by-point responses

  1. Referee: On-device deployment results: the central claim that MicroBi-ConvLSTM is the sole architecture achieving full 8/8 dataset coverage on Pico 2 and ESP32 under INT8 because TinierHAR (34K) and TinyHAR (55K) exceed SRAM budgets once OS overhead is included is unsupported. No tabulated peak SRAM measurements (including overhead) for the reproduced baselines, no explicit overhead value, and no side-by-side memory profiling under identical compilation/quantization settings are provided. This measurement gap directly weakens the uniqueness and necessity arguments that motivate the 11.4K-parameter design.

    Authors: We agree that the current manuscript would be strengthened by more explicit documentation of the memory measurements. In the revised version we will add a dedicated table reporting peak SRAM usage (including OS overhead) for MicroBi-ConvLSTM, TinierHAR and TinyHAR on both the Raspberry Pi Pico 2 and ESP32 under identical INT8 quantization and compilation settings. The table will also list the specific overhead value applied and describe the profiling procedure used to obtain the figures. These additions will directly substantiate the full 8/8 coverage claim and the motivation for the 11.4 K parameter design. revision: yes

Circularity Check

0 steps flagged

No circularity; results are direct empirical measurements with no self-referential reductions.

full rationale

The paper defines MicroBi-ConvLSTM via explicit architectural choices (two-stage convolution with 4x temporal pooling plus one bidirectional LSTM), reports parameter counts by direct enumeration of that structure, and evaluates accuracy plus hardware coverage through standard training runs and on-device deployments on public benchmarks. No equation, fitted constant, or self-citation reduces any reported accuracy, latency, or uniqueness claim back to a quantity defined by the paper's own outputs. The SRAM-overhead premise for baselines is an external assumption rather than a load-bearing internal derivation, leaving the central empirical claims self-contained.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard neural-network assumptions for time-series modeling and on the representativeness of the eight chosen benchmarks; no new entities are postulated and free parameters are limited to ordinary architectural hyperparameters.

free parameters (1)
  • Convolutional filter counts and LSTM hidden size
    Chosen to reach the target 11.4K parameter budget while preserving accuracy.
axioms (2)
  • domain assumption Two-stage convolution followed by 4x temporal pooling retains sufficient information for activity classification.
    Invoked in the feature-extraction stage of the architecture description.
  • domain assumption Bidirectional LSTM improves detection of episodic events over unidirectional processing.
    Stated in the ablation analysis of task-dependent contributions.

pith-pipeline@v0.9.0 · 5863 in / 1550 out tokens · 72450 ms · 2026-05-21T13:20:02.079047+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

45 extracted references · 45 canonical work pages · 2 internal anchors

  1. [1]

    Deep convolutional and LSTM recurrent neural networks for multimodal wearable activity recognition,

    F. J. Ord ´o˜nez and D. Roggen, “Deep convolutional and LSTM recurrent neural networks for multimodal wearable activity recognition,”Sensors, vol. 16, no. 1, p. 115, 2016

    work page 2016

  2. [2]

    TinyHAR: A lightweight deep learning model designed for human activity recognition,

    Y . Zhou, H. Zhao, Y . Huang, T. Radu, M. Constantinides, and S. Mehro- tra, “TinyHAR: A lightweight deep learning model designed for human activity recognition,” inProc. ACM Int. Symp. Wearable Comput., 2022, pp. 89–93

    work page 2022

  3. [3]

    TinierHAR: Towards ultra-lightweight deep learning models for efficient human activity recognition on edge devices,

    S. Bian, M. Liu, V . F. Rey, D. Geissler, and P. Lukowicz, “TinierHAR: Towards ultra-lightweight deep learning models for efficient human activity recognition on edge devices,” inProc. ACM Int. Joint Conf. Pervasive Ubiquitous Comput., 2025

    work page 2025

  4. [4]

    Ensembles of deep LSTM learners for activity recognition using wearables,

    Y . Guan and T. Pl ¨otz, “Ensembles of deep LSTM learners for activity recognition using wearables,”Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 1, no. 2, pp. 1–28, 2017

    work page 2017

  5. [5]

    Deep, convolutional, and recurrent models for human activity recognition using wearables,

    N. Y . Hammerla, S. Halloran, and T. Pl ¨otz, “Deep, convolutional, and recurrent models for human activity recognition using wearables,” in Proc. Int. Joint Conf. Artif. Intell., 2016, pp. 1533–1540

    work page 2016

  6. [6]

    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 Comput. Surv., vol. 54, no. 4, pp. 1–40, 2021

    work page 2021

  7. [7]

    Deep learning for health informatics,

    D. Ravi, C. Wong, B. Lo, and G.-Z. Yang, “Deep learning for health informatics,”IEEE J. Biomed. Health Inform., vol. 21, no. 1, pp. 4–21, 2017

    work page 2017

  8. [8]

    Real-time human activity recognition from accelerometer data using convolutional neural networks,

    A. Ignatov, “Real-time human activity recognition from accelerometer data using convolutional neural networks,”Appl. Soft Comput., vol. 62, pp. 915–922, 2018

    work page 2018

  9. [9]

    A tutorial on human activity recognition using body-worn inertial sensors,

    A. Bulling, U. Blanke, and B. Schiele, “A tutorial on human activity recognition using body-worn inertial sensors,”ACM Comput. Surv., vol. 46, no. 3, pp. 1–33, 2014

    work page 2014

  10. [10]

    Activity recognition using cell phone accelerometers,

    J. R. Kwapisz, G. M. Weiss, and S. A. Moore, “Activity recognition using cell phone accelerometers,”ACM SIGKDD Explor. Newsl., vol. 12, no. 2, pp. 74–82, 2011

    work page 2011

  11. [11]

    A public domain dataset for human activity recognition using smartphones,

    D. Anguita, A. Ghio, L. Oneto, X. Parra, and J. L. Reyes-Ortiz, “A public domain dataset for human activity recognition using smartphones,” in Proc. Eur. Symp. Artif. Neural Netw., 2013, pp. 437–442

    work page 2013

  12. [12]

    Mobile sensor data anonymization,

    M. Malekzadeh, R. G. Clegg, A. Cavallaro, and H. Haddadi, “Mobile sensor data anonymization,” inProc. ACM/IEEE Int. Conf. Internet Things Design Implement., 2019, pp. 49–58

    work page 2019

  13. [13]

    The Opportunity challenge: A bench- mark database for on-body sensor-based activity recognition,

    R. Chavarriaga, H. Sagha, A. Calatroni, S. T. Digumarti, G. Tr ¨oster, J. d. R. Mill ´an, and D. Roggen, “The Opportunity challenge: A bench- mark database for on-body sensor-based activity recognition,”Pattern Recognit. Lett., vol. 34, no. 15, pp. 2033–2042, 2013

    work page 2033

  14. [14]

    Wearable activity tracking in car manufacturing,

    T. Stiefmeier, D. Roggen, G. Ogris, P. Lukowicz, and G. Tr ¨oster, “Wearable activity tracking in car manufacturing,”IEEE Pervasive Comput., vol. 7, no. 2, pp. 42–50, 2008

    work page 2008

  15. [15]

    Wearable assistant for Parkinson’s disease patients with the freezing of gait symptom,

    M. B ¨achlin, M. Plotnik, D. Roggen, I. Maidan, J. M. Hausdorff, N. Giladi, and G. Tr ¨oster, “Wearable assistant for Parkinson’s disease patients with the freezing of gait symptom,”IEEE Trans. Inf. Technol. Biomed., vol. 14, no. 2, pp. 436–446, 2010

    work page 2010

  16. [16]

    Introducing a new benchmarked dataset for activity monitoring,

    A. Reiss and D. Stricker, “Introducing a new benchmarked dataset for activity monitoring,” inProc. Int. Symp. Wearable Comput., 2012, pp. 108–109

    work page 2012

  17. [17]

    UniMiB SHAR: A dataset for human activity recognition using acceleration data from smartphones,

    D. Micucci, M. Mobilio, and P. Napoletano, “UniMiB SHAR: A dataset for human activity recognition using acceleration data from smartphones,”Appl. Sci., vol. 7, no. 10, p. 1101, 2017

    work page 2017

  18. [18]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inProc. Int. Conf. Learn. Represent., 2019

    work page 2019

  19. [19]

    Optuna: A next-generation hyperparameter optimization framework,

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A next-generation hyperparameter optimization framework,” inProc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Mining, 2019, pp. 2623–2631

    work page 2019

  20. [20]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,”Neural Comput., vol. 9, no. 8, pp. 1735–1780, 1997

    work page 1997

  21. [21]

    Speech recognition with deep recurrent neural networks,

    A. Graves, A.-r. Mohamed, and G. Hinton, “Speech recognition with deep recurrent neural networks,” inProc. IEEE Int. Conf. Acoust. Speech Signal Process., 2013, pp. 6645–6649

    work page 2013

  22. [22]

    Bidirectional recurrent neural net- works,

    M. Schuster and K. K. Paliwal, “Bidirectional recurrent neural net- works,”IEEE Trans. Signal Process., vol. 45, no. 11, pp. 2673–2681, 1997

    work page 1997

  23. [23]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” inProc. Int. Conf. Mach. Learn., 2015, pp. 448–456

    work page 2015

  24. [24]

    Dropout: A simple way to prevent neural networks from over- fitting,

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhut- dinov, “Dropout: A simple way to prevent neural networks from over- fitting,”J. Mach. Learn. Res., vol. 15, no. 1, pp. 1929–1958, 2014

    work page 1929

  25. [25]

    Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification,

    K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification,” in Proc. IEEE Int. Conf. Comput. Vis., 2015, pp. 1026–1034

    work page 2015

  26. [26]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inProc. Int. Conf. Learn. Represent., 2015

    work page 2015

  27. [27]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference,

    B. Jacob, S. Kligys, B. Chen, M. Zhu, M. Tang, A. Howard, H. Adam, and D. Kalenichenko, “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2018, pp. 2704–2713

    work page 2018

  28. [28]

    Quantizing deep convolutional networks for efficient inference: A whitepaper

    R. Krishnamoorthi, “Quantizing deep convolutional networks for effi- cient inference,”arXiv preprint arXiv:1806.08342, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  29. [29]

    MLPerf Tiny benchmark,

    C. R. Banbury, V . J. Reddi, M. Lam, W. Fu, A. Faber, M. Mattina, P. Whatmough, L. Lee, H. Tiber, D. Wijayasinghe,et al., “MLPerf Tiny benchmark,” inProc. NeurIPS Datasets Benchmarks Track, 2021

    work page 2021

  30. [30]

    Warden and D

    P. Warden and D. Situnayake,TinyML: Machine Learning with Tensor- Flow Lite on Arduino and Ultra-Low-Power Microcontrollers. O’Reilly Media, 2019

    work page 2019

  31. [31]

    MCUNet: Tiny deep learning on IoT devices,

    J. Lin, W.-M. Chen, Y . Lin, J. Cohn, C. Gan, and S. Han, “MCUNet: Tiny deep learning on IoT devices,” inProc. Adv. Neural Inf. Process. Syst., 2020, pp. 11711–11722

    work page 2020

  32. [32]

    CMSIS-NN: Efficient Neural Network Kernels for Arm Cortex-M CPUs

    L. Lai, N. Suda, and V . Chandra, “CMSIS-NN: Efficient neural networks on ARM Cortex-M CPUs,”arXiv preprint arXiv:1801.06601, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  33. [33]

    Attend and discriminate: Beyond the state-of-the-art for human activity recognition using wearable sensors,

    A. Abedin, M. Ehsanpour, Q. Shi, H. Rezatofighi, and D. C. Ranasinghe, “Attend and discriminate: Beyond the state-of-the-art for human activity recognition using wearable sensors,”Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 5, no. 1, pp. 1–22, 2021

    work page 2021

  34. [34]

    GlobalFusion: A global attentional deep learning framework for multisensor information fusion,

    S. Liu, S. Yao, J. Li, D. Liu, T. Wang, H. Shao, and T. Abdelza- her, “GlobalFusion: A global attentional deep learning framework for multisensor information fusion,”Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 4, no. 1, pp. 1–27, 2020

    work page 2020

  35. [35]

    AttnSense: Multi-level attention mechanism for multimodal human activity recognition,

    H. Ma, W. Li, X. Zhang, S. Gao, and S. Lu, “AttnSense: Multi-level attention mechanism for multimodal human activity recognition,” in Proc. Int. Joint Conf. Artif. Intell., 2019, pp. 3109–3115

    work page 2019

  36. [36]

    On attention models for human activ- ity recognition,

    V . S. Murahari and T. Pl ¨otz, “On attention models for human activ- ity recognition,” inProc. ACM Int. Symp. Wearable Comput., 2018, pp. 100–103

    work page 2018

  37. [37]

    MLP- HAR: Boosting performance and efficiency of HAR models on edge devices with purely fully connected layers,

    Y . Zhou, T. King, H. Zhao, Y . Huang, T. Riedel, and M. Beigl, “MLP- HAR: Boosting performance and efficiency of HAR models on edge devices with purely fully connected layers,” inProc. ACM Int. Symp. Wearable Comput., 2024, pp. 133–139

    work page 2024

  38. [38]

    LHAR: Lightweight human activity recognition on knowledge distilla- tion,

    S. Deng, J. Chen, D. Teng, C. Yang, D. Chen, T. Jia, and H. Wang, “LHAR: Lightweight human activity recognition on knowledge distilla- tion,”IEEE J. Biomed. Health Inform., 2023

    work page 2023

  39. [39]

    A human activity recognition method based on lightweight feature extraction combined with pruned and quantized CNN for wearable device,

    M.-K. Yi, W.-K. Lee, and S. O. Hwang, “A human activity recognition method based on lightweight feature extraction combined with pruned and quantized CNN for wearable device,”IEEE Trans. Consum. Elec- tron., vol. 69, no. 3, pp. 657–670, 2023

    work page 2023

  40. [40]

    Efficient human activity recognition using lookup table-based neural architecture search for mobile devices,

    W.-S. Lim, W. Seo, D.-W. Kim, and J. Lee, “Efficient human activity recognition using lookup table-based neural architecture search for mobile devices,”IEEE Access, vol. 11, pp. 71727–71738, 2023

    work page 2023

  41. [41]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” inProc. Adv. Neural Inf. Process. Syst., 2017, pp. 5998–6008

    work page 2017

  42. [42]

    Are Transformers a useful tool for tiny devices in human activity recognition?,

    E. Lattanzi, L. Calisti, and C. Contoli, “Are Transformers a useful tool for tiny devices in human activity recognition?,” inProc. 8th Int. Conf. Advances Artif. Intell., 2024, pp. 339–344

    work page 2024

  43. [43]

    Improv- ing deep learning for HAR with shallow LSTMs,

    M. Bock, A. H ¨olzemann, M. Moeller, and K. Van Laerhoven, “Improv- ing deep learning for HAR with shallow LSTMs,” inProc. Int. Symp. Wearable Comput., 2021, pp. 7–12

    work page 2021

  44. [44]

    A lightweight framework for human activity recognition on wearable devices,

    Y . L. Coelho, F. de Assis Souza dos Santos, A. Frizera-Neto, and T. Freire Bastos-Filho, “A lightweight framework for human activity recognition on wearable devices,”IEEE Sensors J., vol. 21, no. 21, pp. 24471–24481, 2021

    work page 2021

  45. [45]

    Human activity recognition with smart- phone sensors using deep learning neural networks,

    C. A. Ronao and S.-B. Cho, “Human activity recognition with smart- phone sensors using deep learning neural networks,”Expert Syst. Appl., vol. 59, pp. 235–244, 2016

    work page 2016