Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

In-sensor 24 classes HAR under 850 Bytes

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A 24-class human activity recognition model can run directly inside an intelligent sensor's processing unit using less than 850 bytes of stack memory while holding about 85 percent accuracy.

desk verdict A credible sub-850-byte ISPU deployment demo, but the 85% accuracy is a post-hoc estimate, not a held-out result. read the letter →

arxiv 2502.17472 v1 pith:F4NMZ6HY submitted 2025-02-13 eess.SP

classification eess.SP
keywords HumanActivityRecognitionTinyMLIntelligentSensorProcessingUniton-sensorinferenceincrementalclassinjectionstatisticalfeatureextractionIMUaccelerometergyroscopememory-constrainedmodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that a full 24-class human activity recognition (HAR) system can be embedded directly inside an intelligent sensor processing unit (ISPU), the small digital core built into a 6-axis accelerometer-gyroscope chip. The authors report a model that classifies gestures such as writing, cleaning, sports, and kitchen tasks at roughly 85% accuracy while using under 850 bytes of stack memory and drawing about 0.5 mA. They achieve this by resampling raw 104 Hz sensor streams to 26 Hz, cutting them into 1.5-second windows, extracting 78 statistical features, and then growing the classifier one class at a time so overlapping classes are merged or discarded. If the claim holds, activity recognition can happen at the sensor itself, with no data sent to a microcontroller or cloud, which would reduce power, latency, and privacy risk in wearables and smart environments. The paper also releases a 12.5-hour, 24-gesture dataset for further work.

What carries the argument

The mechanism that carries the argument is a preprocessing-plus-training pipeline whose load-bearing piece is incremental class injection. Raw 104 Hz six-axis data are downsampled to 26 Hz, windowed at 1.5 seconds, and reduced to 78 statistical features per window (max, min, mean, standard deviation, range, zero-crossing rate, and similar quantities on each axis). Training then starts from a small class set and adds classes one at a time, measuring overlap against already-trained classes and either merging or dropping each candidate class; this is what lets the model reach 24 classes while staying under the 2 kB memory target. The deployed model is a decision-tree ensemble restricted to the 16 most important features, and its stack footprint of about 0.82-0.85 kB is the concrete evidence for the headline claim.

What would settle it

Retrain the same pipeline on the released dataset with classes fixed in advance and feature importance computed separately inside each cross-validation training fold; if the held-out accuracy drops materially below 84.7%, the headline number is inflated by selection on the full data.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that a 24-class activity classifier can be deployed on an ISPU within an 850-byte stack memory budget without giving up usable accuracy. The final model is a gradient-boosted decision-tree ensemble (an XGBoost model) trained on only the top 20% most important statistical features, derived from accelerometer and gyroscope windows; this reduces stack usage to 0.82 kB while preserving 84.7% accuracy on the preprocessed data. The authors attribute the scalability to an incremental class injection procedure: adding classes sequentially, checking overlap, and either merging or discarding a new class depending on how confused the current model is. With this procedure the system grew from 12-16 classes to 24, and the released dataset records 24 distinct gestures over 750 minutes with 2-5 participants per class. They also report that the model runs in real time, uses about 0.5 mA, and keeps raw data on the sensor, which is the practical payoff of the whole exercise.

Load-bearing premise

The load-bearing premise is that the reported 85% accuracy reflects how the model will behave on genuinely new recordings, even though feature-importance selection and the final choice of 24 classes were made using the entire dataset, including the evaluation data.

Editorial extensions

If this is right

  • A 24-class activity classifier can be embedded directly in the sensor die, so an IMU can output activity labels instead of raw time series, cutting host-MCU load.
  • Raw acceleration and gyroscope data never have to leave the sensor, strengthening privacy in healthcare and home monitoring.
  • At about 0.5 mA, always-on activity recognition becomes plausible for coin-cell or energy-harvesting wearables.
  • The released 24-gesture dataset gives the tinyML community a public benchmark for ultra-constrained HAR.
  • If the class-injection method transfers, other multi-class sensor tasks such as vibration diagnostics or gesture recognition could fit under 8 kB program and 2 kB data limits.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the feature-importance ranking and the final class selection were computed using the full dataset (Table 1 footnote), the 84.7% figure is likely an optimistic estimate of generalization; re-selecting features inside each training fold would give a cleaner number.
  • The comparison against the 1D ResNet baseline stacks an unconstrained 1M-parameter network against the memory-limited models; a same-budget or quantized ResNet would be a more direct test of what is lost by the feature-based approach.
  • The 24-class set itself was chosen after looking at model behavior on the full data, so the 'scaled to 24 classes' claim is about a curated 24-class problem; fixing classes before seeing the data would test whether the injection strategy truly scales to any 24 classes.
  • One testable extension is to apply the same downsampling-windowing-feature pipeline to other periodic signals, such as ECG or vibration, and measure whether a similarly tiny gradient-boosted classifier can match larger models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper reports a human-activity-recognition (HAR) pipeline for STMicroelectronics ISPUs. The authors collect 12.5 hours of accelerometer/gyroscope data, preprocess it (cleaning, downsampling to 26 Hz, 1.5 s windows, 78 statistical features), and compare 1D ResNet, LightGBM, XGBoost, a basic neural network, and Neuton AutoML. The headline result is an XGBoost model with 84.7% accuracy and a 0.82 KB stack footprint, obtained after selecting the top 20% features and after a class-selection procedure that reduces 31 initial gestures to 24. The paper claims this demonstrates in-sensor deployment of a 24-class HAR model under 850 bytes of stack memory.

Significance. The engineering contribution is timely: it shows that gradient-boosted trees can be compiled to ISPU-compatible C and run with a sub-kilobyte stack, and the proposed preprocessing and feature set may be useful for other ultra-constrained HAR tasks. The paper also contains a useful comparison of several model families on the same pipeline. However, because the feature subset and the 24-class label set are both selected using the full dataset, the reported 84.7% accuracy is not a clean held-out estimate of the deployed system's accuracy. The memory claim for the resulting configuration may be valid, but the accuracy attached to it is not yet credible. No code, binary, or dataset URL is provided, so the central numbers cannot be independently verified.

major comments (4)
  1. [Table 1, footnote * and Section 6] The reported XGBoost* accuracy of 84.7% and the associated 0.82 KB stack footprint are obtained after selecting the top 20% features using the feature importance of a model trained on the entire dataset. This makes the feature subset a function of the evaluation data, so the 84.7% figure cannot be read as an unbiased estimate of the accuracy of the deployed configuration. Please provide a nested evaluation (feature selection performed inside each training fold only) or a separate validation set, and report the resulting accuracy for the final fixed configuration.
  2. [Sections 5 and 7] The 24-class label set is not pre-specified: 'Brushing Teeth' and 'Driving' are discarded and 'Writing'/'Using Computer' are merged based on observed confusions during incremental class injection on the full dataset. Consequently, the reported accuracy is for a taxonomy selected post hoc, not for a fixed 24-class task. The evaluation should treat class-set selection as part of the model-selection procedure, for example by fixing the taxonomy before evaluation or by using nested validation over the class-selection rule.
  3. [Section 6] The text reports '5-fold cross-validation' but does not specify whether the folds split by window, by recording session, or by participant. If windows from the same recording appear in both training and test folds, temporally correlated samples will inflate accuracy. Please describe the split and, if participant-level generalization is intended, report leave-one-participant-out results and per-class accuracy or a confusion matrix.
  4. [Section 3 and Table 1] The headline 'under 850 bytes' refers only to the stack footprint of one XGBoost configuration; the same row reports a 13 Kb program footprint, and Section 3 notes that the ISPU has no ROM and requires a host MCU to load the program. The paper should state the full memory budget (data memory, program memory, and host overhead) and provide the compiled artifacts, since the memory claim is as central as the accuracy claim.
minor comments (6)
  1. [Section 4.1] The first paragraph says the data collection encompassed '31 distinct classes', then says only 24 will be used; the second paragraph immediately says the dataset comprises '24 distinct classes'. This contradiction should be resolved.
  2. [Section 5] The heading 'EXPIREMENTS' and the repeated 'Expirement 1/2/3' and 'Experement 3' are typos and should be corrected to 'Experiments'.
  3. [Table 1] Units are used inconsistently across the table and text: 'KB', 'Kb', and 'Bytes' appear interchangeably (e.g., 0.82KB, 13Kb, 21Kb). Define the units and use them consistently, and verify that '13Kb' is not meant to be '13KB'.
  4. [Abstract / Section 8] The abstract and conclusion claim '85% accuracy' while Table 1 reports 84.7%. Reconcile the exact number or round consistently.
  5. [Section 7] The class 'hand-still' is mentioned in the driving-class discussion but is not listed among the 24 classes in Section 4.1; clarify which class this refers to.
  6. [Abstract / Data availability] The paper states the dataset is 'publicly available' but provides no URL or repository. Add a data availability statement and, if possible, link to the code or compiled ISPU binaries to support reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline 85% accuracy is not a clean held-out estimate: the feature subset and the 24-class taxonomy are both selected using the full dataset, so the reported accuracy is co-determined by the evaluation data.

  1. fitted input called prediction [Table 1 footnote (*) and Section 6 (Results)]
    "The model was trained using the top 20% most important features, as determined from the feature importance matrix of a model trained on the entire dataset. Additionally, to further reduce stack memory consumption, we retrained the XGBoost model using only the top 20% most important features, as determined by the feature importance matrix from the previous training."

    The 84.7% accuracy in the XGBoost* row is the basis for the abstract's 85% figure. The 16-feature subset is selected by training a model on the entire dataset, which includes the same data later used to report the accuracy. Consequently, the test folds participate in choosing the input features, so the reported accuracy is an optimistic in-sample estimate rather than an independent held-out prediction. The feature set is a fitted input to the evaluation, not a configuration fixed before seeing the evaluation data.

  2. other [Section 5 (Experiments) and Section 7 (Discussion)]
    "Using this approach, we scaled the model to 24 distinct classes, selected from an initial set of 31 gestures. ... After obtaining the final set of classes, we train and evaluate models on both raw and preprocessed data using the following architectures. ... During the incremental class injection process, certain classes were either discarded or merged to enhance model performance and maintain memory efficiency."

    The 24-class task itself is not fixed before evaluation. Classes are discarded ("Brushing Teeth", "Driving") or merged ("Writing", "Using Computer") based on the model's observed confusion on the same dataset, and only then are accuracy figures reported on the resulting taxonomy. The reported accuracy therefore applies to a class set chosen to improve performance on that same data, rather than to a pre-specified 24-class benchmark. The problem definition and the reported performance are co-selected, undermining the headline accuracy as an unbiased estimate.

full rationale

The central accuracy claim is partially circular/selection-dependent. The XGBoost* model in Table 1 uses the top 20% features determined from a model trained on the entire dataset, and the 24-class set is obtained by discarding or merging classes based on model confusion on the same data. The reported 85% (84.7%) accuracy is therefore not an independent held-out estimate; it is for a configuration and taxonomy selected using the evaluation data. The 850-byte stack footprint, however, is a compilation and deployment property and is not derived from the accuracy estimate, so the feasibility portion of the claim retains independent content. There is no load-bearing self-citation chain: reference [4] is a self-citation used only to motivate downsampling, and no self-authored uniqueness theorem is invoked. The score reflects partial circularity in the accuracy claim, not in the memory claim.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim depends on data-fitted design choices (feature subset, window size, downsampling rate, class selection) and on the assumption that the reported stack footprint reflects actual ISPU usage. No new physical entities or first-principles constants are introduced.

free parameters (4)
  • Top 20% feature subset = 16 out of 78 features
    Selected based on feature importance from a model trained on the full dataset, then used to retrain the final small XGBoost model. The choice directly affects the 0.82KB stack footprint and 84.7% accuracy.
  • Window size = 1.5 seconds at 26 Hz (39 samples)
    Chosen based on average time between two consecutive major peaks in the recordings; this is a hand-selected design parameter that affects the feature statistics and classification.
  • Downsampling factor = 104 Hz to 26 Hz
    Applied to reduce memory and 'denoise'; the factor of 4 is a design choice not independently validated.
  • Class set = 24 classes selected from 31
    Classes were discarded or merged based on model performance (e.g., Brushing Teeth and Driving discarded; Writing and Writing on Board merged). This selection is data-driven and affects the reported accuracy.
assumptions (4)
  • domain assumption The ISPU stack memory footprint reported (e.g., 0.82KB) corresponds to actual worst-case stack usage during inference on the LSM6DSO16IS sensor.
    The paper states stack footprints but does not describe the measurement method (static analysis, instrumentation, or profiling). The feasibility claim depends on this number being accurate.
  • domain assumption Downsampling to 26 Hz preserves the information needed for HAR classification.
    The paper claims downsampling acts as denoising and improves accuracy, but no experiment compares 26 Hz vs 104 Hz in a controlled way for the final model; it is assumed the morphology is sufficient.
  • domain assumption The statistical features (13 per axis) are sufficient to separate the 24 classes.
    The UMAP plot shows clusters, but the sufficiency of the feature set is assumed; the authors note that the ResNet on preprocessed data plateaus at 93.65%, suggesting some information loss.
  • domain assumption The dataset, recorded by 2-5 participants per class, is representative enough to support claims of generalizable accuracy.
    No participant-independent evaluation is described; the reported accuracies may reflect overfitting to the small participant pool.

how reviews work

0 comments
Cite this review

Pith. "Pith review of In-sensor 24 classes HAR under 850 Bytes." pith.science (2026). https://pith.science/paper/F4NMZ6HY

@misc{pith2026250217472,
  author       = {Pith},
  title        = {Pith review of: In-sensor 24 classes HAR under 850 Bytes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F4NMZ6HY}},
  note         = {Machine review of arXiv:2502.17472}
}
read the original abstract

The year 2023 was a key year for tinyML unleashing a new age of intelligent sensors pushing intelligence from the MCU into the source of the data at the sensor level, enabling them to perform sophisticated algorithms and machine learning models in real-time. This study presents an innovative approach to Human Activity Recognition (HAR) using Intelligent Sensor Processing Units (ISPUs), demonstrating the feasibility of deploying complex machine learning models directly on ultra-constrained sensor hardware. We developed a 24-class HAR model achieving 85\% accuracy while operating within an 850-byte stack memory limit. The model processes accelerometer and gyroscope data in real time, reducing latency, enhancing data privacy, and consuming only 0.5 mA of power. To address memory constraints, we employed incremental class injection and feature optimization techniques, enabling scalability without compromising performance. This work underscores the transformative potential of on-sensor processing for applications in healthcare, predictive maintenance, and smart environments, while introducing a publicly available, diverse HAR dataset for further research. Future efforts will explore advanced compression techniques and broader IoT integration to push the boundaries of TinyML on constrained devices.

Figures

Figures reproduced from arXiv: 2502.17472 by the authors.

Figure 1
Figure 1. Hierarchical Representation of Evolution of Sensors [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. ISPU Chip Architecture Operating at CPU frequencies of 5 MHz or 10 MHz, the ISPU balances processing power with minimal energy consumption. The core is an 8-kilogate Harvard RISC STRED processor that includes a floating point unit, a binary convolutional accelerator, four pipeline stages, and two branch shadows, achieving a CoreMark® score of 70 µW/MHz packaged in LGA-14 [12]. The ISPU features 8kB of data memory an… view at source ↗
Figure 4
Figure 4. Example of Outlier/Noise instances 4.3 Data Downsampling In this study, all signals were downsampled from an original sam￾pling rate of 104 Hz to 26 Hz. This downsampling approach served dual purposes: it reduced computational memory requirements by a factor of four and simultaneously enhanced the accuracy of the machine learning models by acting as a denoising mechanism. By decreasing the sampling rate, the process… view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Example of Downsampling effect: Left hand side at [PITH_FULL_IMAGE:figures/full_fig_p003_5.png]
Figure 6
Figure 6. Figure 6: example of the effect of window size on different [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 7
Figure 7. Figure 7: Umap distrubution of the 24 data classes [PITH_FULL_IMAGE:figures/full_fig_p004_7.png]
Figure 8
Figure 8. Figure 8: Visual Representation of two examples of data win [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. USAD: End-to-End Human Activity Recognition via Diffusion Model with Spatiotemporal Attention

    cs.CV 2025-07 reject novelty 4.0 of 10

    USAD combines diffusion-based data augmentation, multi-branch spatiotemporal attention, and adaptive loss weighting, reporting 98.84% on WISDM, 94.07% on PAMAP2, and 84.60% on OPPORTUNITY, though the abstract lists di...

Reference graph

Works this paper leans on

35 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [1]

    Murray, Benoit Steiner, Paul Tucker, Vijay Vasudevan, Pete Warden, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng

    Martín Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, Man- junath Kudlur, Josh Levenberg, Rajat Monga, Sherry Moore, Derek G. Murray, Benoit Steiner, Paul Tucker, Vijay Vasudevan, Pete Warden, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. 2016. TensorFlow: A system f...

  2. [2]

    Al Hawari

    Shadi Al-Sarawi, Mohammed Anbar, Rosni Abdullah, and Ahmad B. Al Hawari

  3. [3]

    Fabrizio Maria Aymone and Danilo Pietro Pau. 2024. Benchmarking In-Sensor Machine Learning Computing: An Extension to the MLCommons-Tiny Suite. Information 15, 11 (2024). https://doi.org/10.3390/info15110674

  4. [4]

    Ahmed.S Benmessaoud, Farida Medjani, Yahia Bousseloub, Khalid Bouaita, Dhia Benrahem, and Tahar Kezai. 2023. High quality ECG dataset based on MIT- BIH recordings for improved heartbeats classification. In 2023 IEEE International Conference on Omni-layer Intelligent Systems (COINS) . 1–4. https://doi.org/10. 1109/COINS57856.2023.10189299

  5. [5]

    Luigi Bibbò and Marley M. B. R. Vellasco. 2023. Human Activity Recognition (HAR) in Healthcare. Applied Sciences 13, 24 (2023). https://doi.org/10.3390/ app132413009

  6. [6]

    Yahia Bousseloub, Farida Medjani, Ahmed S Benmessaoud, Tahar Kezai, Ali Belhamra, and Issam Attoui. 2024. New method for bearing fault diagnosis based on VMD technique. Diagnostyka 25, 3 (March 2024), 123–135. https: //doi.org/10.1234/diagnostyka.v25i3.56789

  7. [7]

    Wahyu Caesarendra and Tegoeh Tjahjowidodo. 2017. A Review of Feature Extraction Methods in Vibration-Based Condition Monitoring and Its Application for Degradation Trend Estimation of Low-Speed Slew Bearing. Machines 5, 4 (2017). https://doi.org/10.3390/machines5040021

  8. [8]

    Tony Cai and Rong Ma

    T. Tony Cai and Rong Ma. 2022. Theoretical Foundations of t-SNE for Visualizing High-Dimensional Clustered Data. arXiv:2105.07536 [stat.ML] https://arxiv.org/ abs/2105.07536

Show all 35 references
  1. [9]

    Kaixuan Chen, Dalin Zhang, Lina Yao, Bin Guo, Zhiwen Yu, and Yunhao Liu

  2. [10]

    Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (San Francisco, California, USA) (KDD ’16). ACM, New York, NY, USA, 785–794. https://doi.org/1...

  3. [12]

    Mahesh Chowdhary and Swapnil Sayan Saha. 2023. On-Sensor Online Learning and Classification Under 8 KB Memory. https://doi.org/10.23919/FUSION52260. 2023.10224228

  4. [13]

    AbdElRahman ElSaid, Brandon Wild, James Higgins, and Travis Desell. 2016. Using LSTM recurrent neural networks to predict excess vibration events in aircraft engines. In 2016 IEEE 12th International Conference on e-Science (e-Science). 260–269. https://doi.org/10.1109/eScience...

  5. [14]

    Sweta Jain, Bholanath Roy, and Abhishek Deb

    Shubham Gupta, Dr. Sweta Jain, Bholanath Roy, and Abhishek Deb. 2022. A TinyML Approach to Human Activity Recognition. Journal of Physics: Conference Series 2273, 1 (may 2022), 012025. https://doi.org/10.1088/1742-6596/2273/1/ 012025

  6. [15]

    Robert Harle. 2013. A Survey of Indoor Inertial Positioning Systems for Pedes- trians. IEEE Communications Surveys & Tutorials 15, 3 (2013), 1281–1293. https://doi.org/10.1109/SURV.2012.121912.00075

  7. [16]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2015. Deep Residual Learning for Image Recognition. arXiv:1512.03385 [cs.CV] https://arxiv.org/abs/ 1512.03385

  8. [17]

    Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory. Neural Comput. 9, 8 (Nov. 1997), 1735–1780. https://doi.org/10.1162/neco.1997.9. 8.1735

  9. [18]

    Saidul Islam, Hanae Elmekki, Ahmed Elsebai, Jamal Bentahar, Najat Drawel, Gaith Rjoub, and Witold Pedrycz. 2023. A Comprehensive Survey on Applications of Transformers for Deep Learning Tasks. arXiv:2306.07303 [cs.LG] https: //arxiv.org/abs/2306.07303

  10. [19]

    Andrej Karpathy. 2019. A Recipe for Neural Networks. http://karpathy.github. io/2019/04/25/recipe/ Accessed: 2024-11-28

  11. [20]

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. 2017. LightGBM: a highly efficient gradient boosting decision tree. In Proceedings of the 31st International Conference on Neural Information Processing Systems (Long Beach, Calif...

  12. [21]

    Yann LeCun, Bernhard Boser, John S Denker, Donnie Henderson, Richard E Howard, Wayne Hubbard, and Lawrence D Jackel. 1989. Backpropagation applied to handwritten zip code recognition. Neural computation 1, 4 (1989), 541–551

  13. [22]

    Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. 1998. Gradient- based learning applied to document recognition. Proc. IEEE 86, 11 (1998), 2278– 2324

  14. [23]

    Leland McInnes, John Healy, and James Melville. 2020. UMAP: Uni- form Manifold Approximation and Projection for Dimension Reduction. arXiv:1802.03426 [stat.ML] https://arxiv.org/abs/1802.03426

  15. [24]

    Sakorn Mekruksavanich and Anuchit Jitpattanakul. 2019. Classification of Gait Pattern with Wearable Sensing Data. In 2019 Joint International Conference on Digital Arts, Media and Technology with ECTI Northern Section Conference on Electrical, Electronics, Computer and Telecom...

  16. [25]

    Denis Mikhaylov, Tommaso Polonelli, and Michele Magno. 2024. On-Sensor TinyML Event-Based Fault Detection Strategies on Wind Turbine Blades. In 2024 IEEE Sensors Applications Symposium (SAS) . 1–6. https://doi.org/10.1109/ SAS60918.2024.10636542

  17. [26]

    Neuton.ai. 2024. Neuton.ai tinyml platform . https://neuton.ai Accessed on 2024-11-26

  18. [27]

    Daniele Ravì, Charence Wong, Benny Lo, and Guang-Zhong Yang. 2016. A Deep Learning Approach to on-Node Sensor Data Analytics for Mobile or Wearable Devices. IEEE journal of biomedical and health informatics PP (12 2016). https: //doi.org/10.1109/JBHI.2016.2633287

  19. [28]

    de Aguirre, Lucas Compassi-Severo, and Alessandro G

    Marcelo Romanssini, Paulo César C. de Aguirre, Lucas Compassi-Severo, and Alessandro G. Girardi. 2023. A Review on Vibration Monitoring Techniques for Predictive Maintenance of Rotating Machinery. Eng 4, 3 (2023), 1797–1817. https://doi.org/10.3390/eng4030102

  20. [29]

    Andrea Ronco, Lukas Schulthess, David Zehnder, and Michele Magno. 2022. Machine Learning In-Sensors: Computation-enabled Intelligent Sensors For Next Generation of IoT. In 2022 IEEE Sensors . 01–04. https://doi.org/10.1109/ SENSORS52175.2022.9967240

  21. [30]

    STMicroelectronics. 2024. ISM330IS / LSM6DSO16IS - MEMS and Sensors. https: //www.st.com/en/mems-and-sensors/ism330is.html. Accessed on 2024-11-26

  22. [31]

    STMicroelectronics. 2024. ISPU Toolchain - Development Tools for ISPU . https: //www.st.com/en/development-tools/ispu-toolchain.html Accessed on 2024-11- 26

  23. [32]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. Attention Is All You Need. arXiv:1706.03762 [cs.CL] https://arxiv.org/abs/1706.03762

  24. [33]

    Jing Zhang and Dacheng Tao. 2021. Empowering Things With Intelligence: A Survey of the Progress, Challenges, and Opportunities in Artificial Intelligence of Things. IEEE Internet of Things Journal 8, 10 (2021), 7789–7817. https://doi. org/10.1109/JIOT.2020.3039359

  25. [34]

    Juntang Zhuang, Tommy Tang, Yifan Ding, Sekhar Tatikonda, Nicha Dvornek, Xenophon Papademetris, and James S. Duncan. 2020. AdaBelief Optimizer: Adapting Stepsizes by the Belief in Observed Gradients. arXiv:2010.07468 [cs.LG] https://arxiv.org/abs/2010.07468

  26. [2020]

    In 2020 Fourth World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4)

    Internet of Things Market Analysis Forecasts, 2020–2030. In 2020 Fourth World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4). 449–453. https://doi.org/10.1109/WorldS450073.2020.9210375

  27. [2021]

    54, 4, Article 77 (May 2021), 40 pages

    Deep Learning for Sensor-based Human Activity Recognition: Overview, Challenges, and Opportunities. 54, 4, Article 77 (May 2021), 40 pages. https: //doi.org/10.1145/3447744

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.