Pith. sign in

REVIEW 5 major objections 5 minor 4 cited by

SETransformer: A Hybrid Attention-Based Architecture for Robust Human Activity Recognition

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SETransformer, a hybrid Transformer with channel and temporal attention, reaches 84.68% validation accuracy and 84.64% macro F1 on the WISDM activity-recognition benchmark, surpassing LSTM, GRU, BiLSTM, and CNN baselines.

desk verdict A routine composition of off-the-shelf attention modules whose central temporal modeling claim is contradicted by the architecture as written; the headline result never appears in the results table. read the letter →

arxiv 2505.19369 v1 pith:M5ZMX64B submitted 2025-05-25 cs.LG cs.AI

Yunbo Liu , Xukui Qin , Yifan Gao , Xiang Li , Chengwei Feng This is my paper
classification cs.LGcs.AI
keywords HumanActivityRecognitionwearablesensorsTransformerself-attentionsqueeze-and-excitationtemporalattentionpoolingWISDMtime-seriesclassification
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

The paper proposes SETransformer, a hybrid architecture for wearable human activity recognition that combines a two-layer Transformer encoder with channel-wise squeeze-and-excitation attention and a learnable temporal attention pooling layer. The central claim is that this combination captures both long-range temporal dependencies and the differential relevance of sensor axes and time steps, and that it outperforms LSTM, BiLSTM, GRU, and CNN baselines on the WISDM dataset. On the reported validation split the model reaches an accuracy of 84.68% and a macro F1-score of 84.64%. If the claim holds, attention-based pooling and channel recalibration are a practical alternative to recurrent and convolutional HAR models, with interpretable attention weights as a by-product.

What carries the argument

The central machinery is the SETransformer stack: a linear projection from 3 accelerometer channels to a model dimension of 128, two Transformer encoder layers with 4-head self-attention and feed-forward blocks, a squeeze-and-excitation module that pools each channel over time, gates it with two fully connected layers at reduction ratio 16, and multiplies the gated weights back onto the features, and a temporal attention pooling layer that assigns each time step a normalized score $\alpha_t = \exp(v^\top \tanh(W_a H^{SE}[t])) / \sum_k \exp(v^\top \tanh(W_a H^{SE}[k]))$ and forms the final context vector as the weighted sum over time. The Transformer gives the model a global receptive field, the SE module reweights sensor channels, and the pooling layer selects salient time steps. The paper explicitly omits positional encodings, saying the structure of sensor data and sequential windows retains implicit temporal order.

What would settle it

Permute the time steps within each validation window, keeping the per-channel marginal distributions unchanged, and rerun the trained SETransformer; if validation accuracy stays near 84.68%, the model is not using temporal order and the temporal-dynamics claim is falsified, while if accuracy collapses the architecture must be using order through some mechanism the paper does not describe.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a model built from raw triaxial accelerometer windows of 200 time steps can achieve strong classification on WISDM without recurrent or convolutional feature extraction: global self-attention in two Transformer layers models the motion dynamics, a squeeze-and-excitation module reweights the three sensor channels according to their global temporal context, and temporal attention pooling replaces fixed pooling by learning which time steps matter for the activity label. The reported result, 84.68% validation accuracy and 84.64% macro F1, is presented as surpassing all four baselines by a clear margin, with stable convergence and interpretable attention behavior.

Load-bearing premise

The entire temporal-modeling story depends on the premise that the order of the 200 time steps remains available to the network after attention and pooling, even though the architecture injects no positional information anywhere.

Editorial extensions

If this is right

  • If the reported numbers hold up, a pure attention-based model with no convolutional or recurrent layers is sufficient to beat established HAR baselines on WISDM by more than 13 accuracy points over the best CNN baseline.
  • The SE and temporal attention components give the model a built-in interpretation route: channel weights show which accelerometer axes matter, and attention scores show which time segments drive the decision.
  • Because the model is trained end-to-end on z-score-normalized raw windows, the pipeline removes hand-crafted feature engineering from the loop for this benchmark.
  • The same modular design can be extended to additional sensor modalities, such as gyroscope or magnetometer channels, without changing the architecture's logic.

Reading between the lines

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

  • A test the paper does not run: shuffling the time steps within each window should not change the model's predictions, because the architecture contains no positional encoding and no convolution; if accuracy stays near 84.68%, the reported result reflects orderless summary statistics rather than temporal dynamics.
  • Adding a learned or sinusoidal positional encoding and comparing accuracy would settle whether time order carries signal the current model discards: a meaningful gain would show the temporal-modeling claim is untested as written, while no gain would mean WISDM windows of this length can be classified from their multiset statistics.
  • The paper describes the evaluation as subject-independent, but the experimental section only states an 80/20 stratified split; if that split is at the window level rather than the subject level, the comparison may be easier than a user-independent deployment scenario.
  • If replicated on other HAR benchmarks with more subjects or cross-device shifts, the margin over CNN baselines may shrink; the WISDM phone-only setting is a single dataset, not evidence of general superiority.
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

5 major / 5 minor

Summary. The paper proposes SETransformer, a hybrid architecture for human activity recognition from triaxial accelerometer windows. The design combines a Transformer encoder without positional encodings, a squeeze-and-excitation (SE) channel attention module, and a learnable temporal attention pooling mechanism. The authors report a validation accuracy of 84.68% and a macro F1-score of 84.64% on the WISDM dataset, claiming to outperform LSTM, GRU, BiLSTM, and CNN baselines. The manuscript also claims to provide ablation studies validating each module, though no such studies are presented. The paper is an empirical benchmark submission without code release or detailed reproducibility artifacts beyond hyperparameter listings.

Significance. If the architecture and results were sound, the paper would contribute a competitive and modular attention-based HAR model with interpretable channel recalibration and temporal aggregation. The authors provide a clear problem formulation and a detailed hyperparameter table (Table 1), which are useful presentation elements. However, the central modeling claim is invalidated by the permutation-invariance of the described architecture with respect to time, and the reported results are not verifiable because the main comparison table omits the proposed model, no error bars or random-seed information are given, and the promised ablations are absent. As presented, the contribution cannot be assessed and the reported performance is unexplained.

major comments (5)
  1. [Section 3.4 and Section 4.5] The described model is permutation-invariant with respect to the time axis, so the claim that it 'leverages global self-attention to capture activity-specific motion dynamics over extended time windows' (Abstract) is unsupported. Section 3.4 explicitly states that positional encodings are omitted, and the architecture contains no convolution or other order-sensitive operation: the input projection in Section 3.3 applies the same linear map per time step; the self-attention and feed-forward layers are permutation-equivariant; and the temporal attention pooling in Section 4.5 computes a weighted sum with weights α_t that are themselves invariant to time-step order. Consequently, the entire function from X∈R^{200×3} to the context vector c is invariant to reordering of the 200 time steps. The statement that 'the structure of sensor data and sequential convolution of windows' retains implicit temporal order is not a substitute for an order-sensitive mechanism; the overlapping window construction in Section 3.1 cannot inject order information into the network. This is a load-bearing internal inconsistency that invalidates the central contribution.
  2. [Table 2 and Section 4.3] The central performance comparison is unverifiable because Table 2 contains only the LSTM, BiLSTM, GRU, and CNN baselines; the SETransformer row is absent. The text in Section 4.3 refers to 'Table 2' for comparison but does not report the proposed model's numbers there, and the abstract's claimed 84.68% accuracy and 84.64% F1 are not tied to any table entry. Without this row, the claim that SETransformer 'surpasses all baseline architectures by a notable margin' cannot be checked. Additionally, no variance information (e.g., repeated runs with different seeds) is provided for any model, so the reported improvements cannot be distinguished from random variation.
  3. [Section 3.1 and Section 3.7] The dataset description is internally inconsistent. Section 3.1 states that 51 subjects each performed 18 tasks, and that the data include 'gyroscope values,' but Table 1 and Section 3.7 fix the classification output at 6 classes and the input shape at R^{200×3} (accelerometer only). The preprocessing description also says windows are retained only if all 200 time steps share the same activity label, which, for a 20 Hz signal with a 10-second window, would discard most windows in WISDM because activity transitions occur at boundaries; this explanation does not match the standard WISDM protocol. These inconsistencies prevent reproduction of the dataset construction.
  4. [Section 5 and Section 3.1] The paper claims subject-independent evaluation, but the described split in Section 3.1 is an 80/20 stratified split over windows, not over subjects. Section 5 states that the model is 'trained and evaluated in a subject-independent but device-consistent setting,' but the method does not partition by user ID. This is a methodological mismatch: with window-level splitting, windows from the same subject appear in both training and validation, which inflates performance measures and undermines the generalization claim.
  5. [Section 6] The conclusion states that 'Ablation results further validate the individual contributions of the SE and temporal attention modules,' but no ablation experiments are presented anywhere in the manuscript. Section 3.7 mentions 'its ablation variants' in the experimental setup, yet no ablation table, figure, or analysis appears in the results. This is a claim of support that the manuscript does not contain.
minor comments (5)
  1. [Section 4.7] Section 4.7 ('Architectural Overview and Design Motivation') is numbered as part of the results chapter and is positioned between Sections 3.6 and 3.7; the temporal attention pooling is presented in Section 4.5, which is also inside the results chapter. The section numbering and ordering should be corrected so that all methodology components appear before the results.
  2. [Figure 2] The caption of Figure 2 is the placeholder 'Enter Caption' and the figure is not described in the text with a proper caption; a detailed caption explaining the training and validation loss curves is needed.
  3. [Figure 1 and Section 3.7] The confusion matrix in Figure 1 is said to be for the test set, but the experimental setup only defines a validation set; clarify whether the reported metrics and confusion matrix are computed on a held-out test set or on the 20% validation split.
  4. [Throughout] The model name alternates between 'SETransformer' and 'SE-Transformer' (e.g., Figure 1 caption and Section 5); please use a single consistent name.
  5. [References] Several references are tangential to the HAR contribution (e.g., [25] on credit risk detection, [23] on RFID heartbeat monitoring, [15] on instructional video understanding) and do not directly support the specific claims made; consider replacing them with HAR-specific citations.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning found; the reported accuracy is an empirical benchmark result and the self-citations are not load-bearing.

full rationale

The paper makes no first-principles derivation whose inputs coincide with its outputs. Its central claim is an empirical validation accuracy of 84.68% and macro F1 of 84.64% on the WISDM dataset, obtained by training the described network end-to-end and evaluating on a held-out split; that is a standard benchmark claim, not a quantity fitted and then renamed as a prediction. The self-citations in the introduction (refs [3] and [4]) concern pedometer and rehabilitation work and do not supply any premise for the SETransformer architecture, so they are not load-bearing. The paper states that ablation results validate the SE and temporal attention modules, but no ablation table is shown; missing evidence is not circularity. The separate concern that Section 3.4 omits positional encodings while the paper claims temporal modeling is an internal correctness or consistency issue, not a circularity: it does not make the result equivalent to an input by construction. There is no quotable step where an equation equals its input, or where a fitted parameter is renamed as a prediction, or where a self-citation chain forces the conclusion, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claims rest on hand-chosen windowing and architecture hyperparameters (listed), plus three questionable assumptions about temporal order, dataset composition, and subject independence. The paper introduces no invented entities. The architecture-level assumption that self-attention without positional encoding preserves time order is false, which is the largest single burden on the reported results.

free parameters (6)
  • Window size T = 200 time steps (about 10 s), stride 100
    Hand-chosen windowing scheme; the central sequence length and label-consistency filter depend on it.
  • Transformer model dimension and depth = d=128, 2 layers, 4 heads
    Architecture hyperparameters set without reported search or sensitivity analysis.
  • SE reduction ratio = r=16
    Gating bottleneck ratio chosen by hand; affects channel recalibration behavior.
  • Temporal attention hidden dimension = d'=64
    Attention pooling projection size chosen by hand.
  • Training settings = lr=0.001, batch=64, epochs=65, Adam
    Standard but hand-fixed; single-run results depend on these choices.
  • Train/validation split = 80/20 stratified by class
    Split choice determines reported accuracy; not a user-grouped split despite the subject-independent claim.
assumptions (4)
  • ad hoc to paper Transformer self-attention without positional encodings can retain implicit temporal order in sensor windows
    Stated in Section 3.4, but self-attention and the temporal attention pooling are permutation equivariant/invariant, so temporal order is lost; the temporal modeling claim depends on this false premise.
  • domain assumption The WISDM records used correspond to the described 18-task corpus with a six-class labeling problem
    Section 3.1 says 18 tasks and 18 fields, while Table 1 specifies six output classes; the exact activity subset and device (phone vs watch) are never pinned down.
  • ad hoc to paper An 80/20 stratified split over windows produces a subject-independent evaluation
    Section 3.1 describes the split and Section 5 claims subject independence, but without grouping by user, the same subject's windows can appear in both training and validation, causing leakage.
  • domain assumption Window labels are consistent and reliable
    Only windows whose 200 time steps share the same label are kept; this assumes the dataset's labels are correct at 20 Hz and that discarding mixed windows does not bias the remaining classes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SETransformer: A Hybrid Attention-Based Architecture for Robust Human Activity Recognition." pith.science (2026). https://pith.science/paper/M5ZMX64B

@misc{pith2026250519369,
  author       = {Pith},
  title        = {Pith review of: SETransformer: A Hybrid Attention-Based Architecture for Robust Human Activity Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M5ZMX64B}},
  note         = {Machine review of arXiv:2505.19369}
}
read the original abstract

Human Activity Recognition (HAR) using wearable sensor data has become a central task in mobile computing, healthcare, and human-computer interaction. Despite the success of traditional deep learning models such as CNNs and RNNs, they often struggle to capture long-range temporal dependencies and contextual relevance across multiple sensor channels. To address these limitations, we propose SETransformer, a hybrid deep neural architecture that combines Transformer-based temporal modeling with channel-wise squeeze-and-excitation (SE) attention and a learnable temporal attention pooling mechanism. The model takes raw triaxial accelerometer data as input and leverages global self-attention to capture activity-specific motion dynamics over extended time windows, while adaptively emphasizing informative sensor channels and critical time steps. We evaluate SETransformer on the WISDM dataset and demonstrate that it significantly outperforms conventional models including LSTM, GRU, BiLSTM, and CNN baselines. The proposed model achieves a validation accuracy of 84.68\% and a macro F1-score of 84.64\%, surpassing all baseline architectures by a notable margin. Our results show that SETransformer is a competitive and interpretable solution for real-world HAR tasks, with strong potential for deployment in mobile and ubiquitous sensing applications.

Figures

Figures reproduced from arXiv: 2505.19369 by the authors.

Figure 1
Figure 1. Confusion matrix of the SE-Transformer model on [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Enter Caption The evolution of both training and validation loss over 65 epochs is shown in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

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

  1. Analysis of Anonymous User Interaction Relationships and Prediction of Advertising Feedback Based on Graph Neural Network

    cs.IR 2025-06 reject novelty 5.0 of 10

    DTH-GNN, a graph neural network with temporal edge decomposition, heterogeneous transformers, and contrastive learning, is claimed to improve ad feedback prediction AUC by 8.2% on the Coveo dataset, but evidence is in...

  2. Predicting Asphalt Pavement Friction Using Texture-Based Image Indicator

    cs.CV 2025-07 conditional novelty 4.0 of 10

    The paper defines aggregate protrusion area from 2D pavement images as a texture indicator and reports linear models with adjusted R2 above 0.90 for predicting DFT friction for three asphalt surface types.

  3. Robust Anomaly Detection in Network Traffic: Evaluating Machine Learning Models on CICIDS2017

    cs.CR 2025-06 reject novelty 3.0 of 10

    A benchmark finds supervised MLP and CNN detect known attacks almost perfectly but miss most novel attacks, while OCSVM generalizes better to unseen threats.

  4. Gated Multimodal Graph Learning for Personalized Recommendation

    cs.IR 2025-05 reject novelty 2.0 of 10

    RLMultimodalRec combines gated image/text fusion with LightGCN, but its own table shows it is not better than LayerGCN on half the reported metrics.

Reference graph

Works this paper leans on

31 extracted references · 27 canonical work pages · cited by 4 Pith papers

  1. [1]

    A survey on deep learning architectures in human activities recog- nition application in sports science, healthcare, and secu- rity

    Basant Adel, Asmaa Badran, Nada E Elshami, Ahmad Salah, Ahmed Fathalla, and Mahmoud Bekhit. A survey on deep learning architectures in human activities recog- nition application in sports science, healthcare, and secu- rity. In The International Conference on Innovations in Computing Research, pages 121–134. Springer, 2022

    work page 2022

  2. [2]

    The appli- cations of metaheuristics for human activity recognition and fall detection using wearable sensors: A comprehen- sive analysis

    Mohammed AA Al-Qaness, Ahmed M Helmi, Abdel- ghani Dahou, and Mohamed Abd Elaziz. The appli- cations of metaheuristics for human activity recognition and fall detection using wearable sensors: A comprehen- sive analysis. Biosensors, 12(10):821, 2022

    work page 2022

  3. [3]

    Jy61 imu sensor external validity: A framework for advanced pe- dometer algorithm personalisation

    Boris Ba ˇci´c, Chengwei Feng, and Weihua Li. Jy61 imu sensor external validity: A framework for advanced pe- dometer algorithm personalisation. ISBS Proceedings Archive, 42(1):60, 2024

    work page 2024

  4. [4]

    Towards nation-wide analytical healthcare infrastructures: A privacy-preserving aug- mented knee rehabilitation case study

    Boris Ba ˇci´c, Claudiu Vasile, Chengwei Feng, and Marian G Ciuc ˘a. Towards nation-wide analytical healthcare infrastructures: A privacy-preserving aug- mented knee rehabilitation case study. arXiv preprint arXiv:2412.20733, 2024

    arXiv 2024

  5. [5]

    A comparative analysis of k-nearest neighbor, genetic, support vector machine, decision tree, and long short term memory algorithms in machine learning

    Malti Bansal, Apoorva Goyal, and Apoorva Choudhary. A comparative analysis of k-nearest neighbor, genetic, support vector machine, decision tree, and long short term memory algorithms in machine learning. Decision Analytics Journal, 3:100071, 2022

    work page 2022

  6. [6]

    Is space-time attention all you need for video understand- ing? In ICML, volume 2, page 4, 2021

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understand- ing? In ICML, volume 2, page 4, 2021

    work page 2021

  7. [7]

    An overview of indoor localization system for human activity recognition (har) in healthcare

    Luigi Bibb `o, Riccardo Carotenuto, and Francesco Della Corte. An overview of indoor localization system for human activity recognition (har) in healthcare. Sen- sors, 22(21):8119, 2022

    work page 2022

  8. [8]

    Unidoc: A universal large multimodal model for simultaneous text detection, recognition, spotting and understanding.arXiv preprint arXiv:2308.11592, 2023

    Hao Feng, Zijian Wang, Jingqun Tang, Jinghui Lu, Wen- gang Zhou, Houqiang Li, and Can Huang. Unidoc: A universal large multimodal model for simultaneous text detection, recognition, spotting and understanding.arXiv preprint arXiv:2308.11592, 2023

    arXiv 2023

Show all 31 references
  1. [9]

    Deep learning for computer vision based activity recognition and fall detection of 7 Journal of Emerging Applied Artificial Intelligence (JEAAI) the elderly: a systematic review

    F Xavier Gaya-Morey, Cristina Manresa-Yee, and Jos´e M Buades-Rubio. Deep learning for computer vision based activity recognition and fall detection of 7 Journal of Emerging Applied Artificial Intelligence (JEAAI) the elderly: a systematic review. Applied Intelligence, 54(19):...

    2024

  2. [10]

    Hang- time har: a benchmark dataset for basketball activity recognition using wrist-worn inertial sensors

    Alexander Hoelzemann, Julia Lee Romero, Marius Bock, Kristof Van Laerhoven, and Qin Lv. Hang- time har: a benchmark dataset for basketball activity recognition using wrist-worn inertial sensors. Sensors, 23(13):5879, 2023

    2023

  3. [11]

    Squeeze-and-excitation networks

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 7132– 7141, 2018

    2018

  4. [12]

    Per- ceiver: General perception with iterative attention

    Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Per- ceiver: General perception with iterative attention. In International conference on machine learning , pages 4651–4664. PMLR, 2021

    2021

  5. [13]

    Human activity recognition: A survey

    Charmi Jobanputra, Jatna Bavishi, and Nishant Doshi. Human activity recognition: A survey. Procedia Com- puter Science, 155:698–703, 2019

    2019

  6. [14]

    Transformer-based approaches for sensor-based human activity recognition: Opportuni- ties and challenges

    Clayton Souza Leite, Henry Mauranen, Aziza Zhan- abatyrova, and Yu Xiao. Transformer-based approaches for sensor-based human activity recognition: Opportuni- ties and challenges. arXiv preprint arXiv:2410.13605 , 2024

    2024 arXiv

  7. [15]

    To- wards visual-prompt temporal answer grounding in in- structional video

    Shutao Li, Bin Li, Bin Sun, and Yixuan Weng. To- wards visual-prompt temporal answer grounding in in- structional video. IEEE transactions on pattern analysis and machine intelligence, 46(12):8836–8853, 2024

    2024

  8. [16]

    Foundation models for time series analysis: A tutorial and survey

    Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shirui Pan, and Qingsong Wen. Foundation models for time series analysis: A tutorial and survey. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining , pages 6555–6565, 2024

    2024

  9. [17]

    Deep learning for intelligent human– computer interaction

    Zhihan Lv, Fabio Poiesi, Qi Dong, Jaime Lloret, and Houbing Song. Deep learning for intelligent human– computer interaction. Applied Sciences, 12(22):11457, 2022

    2022

  10. [18]

    U-har: A convolutional ap- proach to human activity recognition combining head and eye movements for context-aware smart glasses

    Johannes Meyer, Adrian Frank, Thomas Schlebusch, and Enkelejda Kasneci. U-har: A convolutional ap- proach to human activity recognition combining head and eye movements for context-aware smart glasses. Proceedings of the ACM on Human-Computer Interac- tion, 6(ETRA):1–19, 2022

    2022

  11. [19]

    Ahar: Adaptive cnn for energy-efficient human activity recognition in low- power edge devices

    Nafiul Rashid, Berken Utku Demirel, and Moham- mad Abdullah Al Faruque. Ahar: Adaptive cnn for energy-efficient human activity recognition in low- power edge devices. IEEE Internet of Things Journal , 9(15):13041–13051, 2022

    2022

  12. [20]

    Distributed sec- ondary frequency control scheme with a-symmetric time varying communication delays and switching topology

    Muhammad Asfandyar Rustam, Muhammad Yasir Ali Khan, Tasawar Abbas, and Bilal Khan. Distributed sec- ondary frequency control scheme with a-symmetric time varying communication delays and switching topology. e-Prime-Advances in Electrical Engineering, Electronics and Energy, 9:...

    2024

  13. [21]

    Advanced sensing and hu- man activity recognition in early intervention and reha- bilitation of elderly people

    Lisa Schrader, Agust ´ın Vargas Toro, Sebastian Koni- etzny, Stefan R ¨uping, Barbara Sch ¨apers, Martina Steinb¨ock, Carmen Krewer, Friedemann M ¨uller, J ¨org G¨uttler, and Thomas Bock. Advanced sensing and hu- man activity recognition in early intervention and reha- bilitat...

    2020

  14. [22]

    Spiking neural net- works with temporal attention-guided adaptive fusion for imbalanced multi-modal learning

    Jiangrong Shen, Yulin Xie, Qi Xu, Gang Pan, Hua- jin Tang, and Badong Chen. Spiking neural net- works with temporal attention-guided adaptive fusion for imbalanced multi-modal learning. arXiv preprint arXiv:2505.14535, 2025

    2025 arXiv

  15. [23]

    Fine- grained heartbeat waveform monitoring with rfid: A la- tent diffusion model

    Yiting Wang, Tianya Zhao, and Xuyu Wang. Fine- grained heartbeat waveform monitoring with rfid: A la- tent diffusion model. In Proceedings of the 3rd Interna- tional Workshop on Human-Centered Sensing, Modeling, and Intelligent Systems, pages 86–91, 2025

    2025

  16. [24]

    Fully-connected spatial-temporal graph for multivariate time-series data

    Yucheng Wang, Yuecong Xu, Jianfei Yang, Min Wu, Xi- aoli Li, Lihua Xie, and Zhenghua Chen. Fully-connected spatial-temporal graph for multivariate time-series data. In Proceedings of the AAAI conference on artificial intel- ligence, volume 38, pages 15715–15724, 2024

    2024

  17. [25]

    Appli- cation of ai in real-time credit risk detection

    Zhuqi Wang, Qinghe Zhang, and Zhuopei Cheng. Appli- cation of ai in real-time credit risk detection. Preprints, February 2025

    2025

  18. [26]

    Wisdm smartphone and smartwatch activity and biometrics dataset

    Gary M Weiss. Wisdm smartphone and smartwatch activity and biometrics dataset. UCI Machine Learn- ing Repository: WISDM Smartphone and Smartwatch Activity and Biometrics Dataset Data Set , 7(133190- 133202):5, 2019

    2019

  19. [27]

    V-petl bench: A uni- fied visual parameter-efficient transfer learning bench- mark

    Yi Xin, Siqi Luo, Xuyang Liu, Haodi Zhou, Xinyu Cheng, Christina E Lee, Junlong Du, Haozhe Wang, MingCai Chen, Ting Liu, et al. V-petl bench: A uni- fied visual parameter-efficient transfer learning bench- mark. Advances in Neural Information Processing Sys- tems, 37:80522–80535, 2024

    2024

  20. [28]

    Comet: A communication-efficient and performant approximation for private transformer infer- ence

    Xiangrui Xu, Qiao Zhang, Rui Ning, Chunsheng Xin, and Hongyi Wu. Comet: A communication-efficient and performant approximation for private transformer infer- ence. arXiv preprint arXiv:2405.17485, 2024

    2024 arXiv

  21. [29]

    Wcdt: World-centric diffusion trans- former for traffic scene generation

    Chen Yang, Yangfan He, Aaron Xuxiang Tian, Dong Chen, Jianhui Wang, Tianyu Shi, Arsalan Heydarian, and Pei Liu. Wcdt: World-centric diffusion trans- former for traffic scene generation. arXiv preprint arXiv:2404.02082, 2024. 8 Journal of Emerging Applied Artificial Intelligenc...

    2024 arXiv

  22. [30]

    A computational ethology approach for characterizing behavioral dynamics in bipolar disorder

    Zhanqi Zhang, Chi K Chou, Holden Rosberg, William Perry, Jared W Young, Arpi Minassian, Gal Mishne, and Mikio Aoi. A computational ethology approach for characterizing behavioral dynamics in bipolar disorder. medRxiv, 2024. Preprint, not peer-reviewed

    2024

  23. [31]

    Continuous human activity recognition with distributed radar sensor net- works and cnn–rnn architectures

    Simin Zhu, Ronny Gerhard Guendel, Alexander Yarovoy, and Francesco Fioranelli. Continuous human activity recognition with distributed radar sensor net- works and cnn–rnn architectures. IEEE Transactions on Geoscience and Remote Sensing, 60:1–15, 2022. 9

    2022

Pith tools

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