Pith. sign in

REVIEW 4 major objections 4 minor 61 references

Ensemble Distribution Distillation for Self-Supervised Human Activity Recognition

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A single prior network, trained by ensemble distribution distillation, matches a 10-member ensemble's accuracy and robustness on human activity recognition while running one forward pass.

desk verdict Solid applied paper that transfers EDD to self-supervised HAR; the robustness claim is real but thinner than advertised because it's FGSM-only. read the letter →

arxiv 2509.08225 v1 pith:OPNHV5KO submitted 2025-09-10 cs.LG

Matthew Nolan , Lina Yao , Robert Davidson This is my paper
classification cs.LG
keywords ensembledistributiondistillationself-supervisedlearninghumanactivityrecognitionuncertaintyquantificationadversarialrobustnesspriornetworksdataaugmentationwearablesensors
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 argues that ensemble-level reliability can be distilled into a single, cheap network for human activity recognition. It trains a 50-member ensemble of semi-supervised models on a small labeled set plus unlabeled sensor data, then uses ensemble distribution distillation to train one prior network to imitate the ensemble's predictive distribution. On four public HAR datasets, the distilled network matches or approaches ensemble accuracy on clean data and under FGSM adversarial perturbation (epsilon=0.1), with total-uncertainty quality close to a 10-member ensemble. Because the result is a single model run once, the method delivers these gains at the same inference cost as the baseline single model. A sympathetic reader should care because uncertainty quantification normally requires multiple forward passes, which is often impractical on wearable and edge devices.

What carries the argument

The central mechanism is ensemble distribution distillation into a prior network: rather than a point prediction, the network outputs Dirichlet parameters over activity classes, which define a distribution over class probabilities. Training minimizes the KL divergence between this Dirichlet distribution and the distribution implied by the ensemble's outputs, allowing one forward pass to reproduce the ensemble's mean prediction, total uncertainty, and separable epistemic and aleatoric estimates. Two data-augmentation ingredients carry the argument: the eight self-supervised transforms extend the distillation dataset into invariant directions, and weighted combinations of samples fill the betw

What would settle it

Run the same evaluation under a stronger white-box attack, such as projected gradient descent at epsilon=0.1, or an attack optimized directly against the distilled prior network, and compare the distilled model's accuracy and uncertainty AUC-ROC to those of the 10- and 50-member ensembles. If the distilled model drops to single-model-level accuracy while the ensembles remain robust, the central robustness claim is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that ensemble diversity, not just ensemble accuracy, is transferable to a single model through ensemble distribution distillation in the self-supervised HAR setting. The authors train each ensemble member with the same architecture and partially supervised procedure used for the baseline single model, differing only in random initialization, batch sampling, and hidden-unit scaling. The distilled prior network, whose outputs parameterize a Dirichlet distribution over activity classes, is trained by minimizing KL divergence to the ensemble's output distribution on an augmented distillation dataset. Their results show that this single network retains close to the pr

Load-bearing premise

The load-bearing assumption is that robustness measured against FGSM at epsilon=0.1 generalizes to adversarial perturbation in general; if a stronger or adaptive attack erases the distilled model's advantage over the single model, the headline robustness claim fails.

Editorial extensions

If this is right

  • Single-pass deployment: activity-recognition models on phones and wearables could get ensemble-level accuracy and uncertainty without running multiple models, enabling uncertainty-based gating or fall detection in real time.
  • Adversarial robustness without adversarial training: the robustness gain comes from distilling ensemble diversity, not from training on adversarial examples, suggesting a path to robustness that does not require knowing the attack in advance.
  • Small-label regime: with only 50 labeled instances per class, distilled ensembles outperform the fully supervised single-model baseline on several datasets, so reliable HAR may be possible where labels are scarce.
  • Uncertainty as a quality signal: the distilled model's total uncertainty remains a useful predictor of misclassification under perturbation, supporting its use to trigger human review or alternative sensors.

Reading between the lines

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

  • The paper evaluates robustness only under FGSM at epsilon=0.1; whether the distilled model retains ensemble-level accuracy under stronger or adaptive attacks, such as PGD or attacks designed for the distilled model, is an open extension.
  • The weighted-combination augmentation, designed to create between-class sensor samples, could plausibly transfer to other time-series domains with few labels, but that transfer is not shown here.
  • The separable epistemic and aleatoric uncertainties the model outputs are measured but not exploited; a natural next step is using epistemic uncertainty for active learning or out-of-distribution detection on wearable data.
  • Because the architecture is unchanged from the baseline, the same distillation recipe could be applied to any small HAR model to improve calibration and robustness, though gains may depend on dataset diversity, as the UCI results suggest.
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 / 4 minor

Summary. The paper proposes an Ensemble Distribution Distillation (EDD) pipeline for semi-supervised human activity recognition. It first trains an ensemble of CNNs following Saeed et al.'s self-supervised pretext task with a small labeled downstream set, augments unlabeled data with eight transformations and weighted combinations of samples, and distills the ensemble's predictive distribution into a single prior network by minimizing KL divergence. The manuscript reports accuracy, FGSM robustness (epsilon=0.1), and total-uncertainty quality (accuracy-vs-uncertainty curves, AUC-ROC) on HHAR, UCI HAR, MotionSense, and PAMAP2, comparing a single model, 10- and 50-member ensembles, and the distilled model. The central claim is that the distilled model retains ensemble-level accuracy, robustness, and uncertainty at no extra inference cost.

Significance. If the claims hold, the paper addresses an important deployment gap: obtaining ensemble-grade predictive accuracy and uncertainty estimates at single-model inference cost. The evaluation uses four public datasets, matched architectures, 10-run averages, and the method does not appear to fit constants to the test sets. These are genuine strengths. The main weaknesses are the narrow adversarial evaluation protocol and the incomplete reporting of training hyperparameters; both are addressable and do not invalidate the overall direction.

major comments (4)
  1. [§3.2; §4.3.2, Table 2] The robustness claim in the Abstract and Section 4.3.2 is supported only by FGSM with epsilon=0.1. FGSM is a single-step attack and can overstate robustness. More importantly, the manuscript does not state whether perturbations are computed white-box per model or transferred from a single model. If a single perturbation from the baseline model is applied to all models, Table 2 compares white-box single-model accuracy against black-box transfer accuracy for the ensembles and distilled model, invalidating the comparison. Please specify the attack protocol and add an iterative attack (e.g., PGD) or a per-model white-box FGSM baseline. This is load-bearing because 'substantial increases in robustness against adversarial perturbation' is a headline contribution.
  2. [Algorithm 1; §4.1] Algorithm 1 lists E, t0, t_max, v_t, v_c, N_C, r, and N_F as inputs, but Section 4.1 provides only the architecture, ensemble sizes, and hidden-unit scaling factor r. No values or schedules are given for the temperature annealing, combo-depth schedule, number of epochs, or combo weight distribution. Consequently the experiments cannot be reproduced and the role of the augmentation procedure cannot be isolated. Please report these hyperparameters explicitly, or provide a reproducibility appendix with the values used for each dataset.
  3. [Tables 3–6; §4.3.3] AUC-ROC scores are reported as point estimates even though the paper states that all results are averaged over 10 runs. The uncertainty-quality comparisons, such as the claim that the distilled model achieves the highest AUC-ROC on PAMAP2, lack error bars or significance tests. Without variance estimates it is unclear whether the observed differences are meaningful. Please report mean±std or confidence intervals for AUC-ROC and, where ranking claims are made, provide a paired statistical test.
  4. [§4.3.2, final paragraph] The text states that the observed adversarial resilience comes 'solely as a result of the ensemble diversity.' This attribution is not supported by the experiments, because the distilled model also receives self-supervised pretraining, transformed data, weighted-combination augmentation, and temperature annealing. A comparison against the single-model baseline does not isolate ensemble diversity. Please either add ablations that separate these components or soften the attribution claim.
minor comments (4)
  1. [§2.2, Eq. (2)] Bayes' theorem is written as p(θ|D)=p(D|θ)p(θ); the normalizing constant p(D) is missing. It should be p(θ|D)=p(D|θ)p(θ)/p(D), or the right-hand side should be marked as proportional.
  2. [Algorithm 1; §4.1] The symbol r is overloaded: in Algorithm 1 and Eq. (7) it denotes the weighted-combination parameter, while in Section 4.1 it denotes the hidden-unit scaling factor. Please use distinct symbols.
  3. [§2.2] Typo: 'we can be use this approach' should read 'we can use this approach'.
  4. [§4.1] The relationship between the 50-member ensemble used for distillation and the reported 10-member ensemble is not stated explicitly. Is Ensemble-10 a subset of the 50 members, or a separate group? Please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the distilled model's accuracy, robustness, and uncertainty are empirical outcomes evaluated on held-out data and perturbations not used in training.

full rationale

The paper's central chain is: (1) train an ensemble of 50 semi-supervised models following Saeed et al. [48]; (2) build a distillation dataset by running the ensemble on augmented inputs; (3) train a prior network to match the ensemble's predictive Dirichlet distribution via KL/NLL (Eq. 5 and Algorithm 1); (4) evaluate accuracy, FGSM robustness, and uncertainty quality on held-out participants (HHAR/UCI/MotionSense/PAMAP2) against single-model and ensemble baselines. No step reduces to its own inputs: the distilled model's outputs are not fitted to the reported test-set accuracies or AUC-ROC values, and the FGSM perturbations (Sections 3.2 and 4.3.2) were not used as training data. The method builds on prior external work [38,39] and an external training pipeline [48], with no author self-citations carrying a load-bearing argument. The only noteworthy caveat—robustness being measured solely with FGSM at epsilon=0.1—is a threat-model limitation, not circularity. The selection of augmentations based on validation performance is a model-selection concern, but the reported metrics are measurements, not predictions implied by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central method rests on two borrowed building blocks (EDD and self-supervised pretraining) plus an interpolation augmentation whose key hyperparameters (r, temperature schedule, combo depth schedule) are never reported. No invented entities or new physical objects are introduced. The four axioms listed are load-bearing: if any fails, the corresponding claim (distillation quality, transferability, uncertainty calibration, or robustness generality) loses support.

free parameters (5)
  • combo weight parameter r = not specified in paper
    Defined in Eq. 7 and used in Algorithm 1 to weight combinations of samples; no value or search range is reported, yet it controls the interpolation that the authors claim is essential for uncertainty capture.
  • temperature schedule (t0, t_max, v_t) = not specified
    Algorithm 1 and Section 3.1.2 rely on an annealing schedule but give no concrete values, preventing replication.
  • combo depth schedule (v_c, N_C) = not specified
    The maximum number of combined samples grows with epochs via v_c and N_C, but neither is reported.
  • number of epochs E = not specified
    Training duration is an input to Algorithm 1 but never reported.
  • ensemble hidden-unit scaling range = 0.75 to 1.25
    Random per-member scaling used to diversify ensemble members; reported in Section 4.1.
assumptions (4)
  • domain assumption EDD as formalized by Malinin et al. [39] correctly distills an ensemble's predictive distribution into a Dirichlet prior network.
    Adopted wholesale in Section 2.3 and Section 3; the paper does not re-derive or verify this.
  • domain assumption The self-supervised transform-prediction pretraining of Saeed et al. [48] yields transferable representations for HAR with limited labels.
    Used to train all ensemble members (Section 3); the paper inherits this method without independent validation.
  • ad hoc to paper Weighted combinations of samples (Eq. 7) create points with genuinely high epistemic uncertainty that a trustworthy ensemble labels meaningfully.
    Proposed in Section 3.1.1; no analysis shows the ensemble's labels on these interpolated points are well-calibrated.
  • domain assumption FGSM with epsilon=0.1 is representative of adversarial perturbation for the robustness claims.
    Section 3.2 selects FGSM for simplicity; Table 2 and Section 4.3.2 generalize from it to 'adversarial perturbation'.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ensemble Distribution Distillation for Self-Supervised Human Activity Recognition." pith.science (2026). https://pith.science/paper/OPNHV5KO

@misc{pith2026250908225,
  author       = {Pith},
  title        = {Pith review of: Ensemble Distribution Distillation for Self-Supervised Human Activity Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OPNHV5KO}},
  note         = {Machine review of arXiv:2509.08225}
}
read the original abstract

Human Activity Recognition (HAR) has seen significant advancements with the adoption of deep learning techniques, yet challenges remain in terms of data requirements, reliability and robustness. This paper explores a novel application of Ensemble Distribution Distillation (EDD) within a self-supervised learning framework for HAR aimed at overcoming these challenges. By leveraging unlabeled data and a partially supervised training strategy, our approach yields an increase in predictive accuracy, robust estimates of uncertainty, and substantial increases in robustness against adversarial perturbation; thereby significantly improving reliability in real-world scenarios without increasing computational complexity at inference. We demonstrate this with an evaluation on several publicly available datasets. The contributions of this work include the development of a self-supervised EDD framework, an innovative data augmentation technique designed for HAR, and empirical validation of the proposed method's effectiveness in increasing robustness and reliability.

Figures

Figures reproduced from arXiv: 2509.08225 by the authors.

Figure 1
Figure 1. Schematic representation of the semi-supervised the approach, reproduced from [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. The Ensemble Distribution Distillation procedure. First, an ensemble of models [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Comparison of model accuracies on HHAR with no adversarial perturbation at [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Comparison of model accuracies on HHAR with heavy adversarial perturbation [PITH_FULL_IMAGE:figures/full_fig_p023_4.png]
Figure 5
Figure 5. Figure 5: Comparison of model accuracies on UCI with no adversarial perturbation at [PITH_FULL_IMAGE:figures/full_fig_p025_5.png]
Figure 6
Figure 6. Figure 6: Comparison of model accuracies on UCI with heavy adversarial perturbation [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: Comparison of model accuracies on MotionSense with no adversarial perturbation [PITH_FULL_IMAGE:figures/full_fig_p026_7.png]
Figure 8
Figure 8. Figure 8: Comparison of model accuracies on MotionSense with heavy adversarial pertur [PITH_FULL_IMAGE:figures/full_fig_p027_8.png]
Figure 9
Figure 9. Figure 9: Comparison of model accuracies on PAMAP2 with no adversarial perturbation [PITH_FULL_IMAGE:figures/full_fig_p028_9.png]
Figure 10
Figure 10. Figure 10: Comparison of model accuracies on PAMAP2 with heavy adversarial perturba [PITH_FULL_IMAGE:figures/full_fig_p028_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 48 canonical work pages

  1. [1]

    Rajendra Acharya, Vladimir Makarenkov, and Saeid Nahavandi

    Moloud Abdar, Farhad Pourpanah, Sadiq Hussain, Dana Rezazadegan, Li Liu, Mohammad Ghavamzadeh, Paul Fieguth, Xiaochun Cao, Ab- bas Khosravi, U. Rajendra Acharya, Vladimir Makarenkov, and Saeid Nahavandi. A review of uncertainty quantification in deep learning: Techniques, applications and challenges.Information Fusion, 76:243–297, 2021

    work page 2021

  2. [2]

    Large scale distributed neural network training through online distillation

    Rohan Anil, Gabriel Pereyra, Alexandre Tachard Passos, Robert Or- mandi, George Dahl, and Geoffrey Hinton. Large scale distributed neural network training through online distillation. InInternational Conference on Learning Representations (ICLR) Workshop Track, 2018

    work page 2018

  3. [3]

    wav2vec 2.0: A framework for self-supervised learning of speech representations.arXiv preprint, abs/2006.11477, 2020

    Alexei Baevski, Henry Zhou, Abdel-rahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations.arXiv preprint, abs/2006.11477, 2020

    arXiv 2006

  4. [4]

    Recent advances in adversarial training for adversarial robustness

    Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI-21), pages 4312–4321, 2021. Survey Track

    work page 2021

  5. [5]

    Blei, Alp Kucukelbir, and Jon D

    David M. Blei, Alp Kucukelbir, and Jon D. McAuliffe. Variational inference: A review for statisticians.Journal of the American Statistical Association, 112(518):859–877, 2017

    2017

  6. [6]

    Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods.Machine Learning, 110(5):1309–1354, 2021

    Michael Blum and Martin Riedmiller. Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods.Machine Learning, 110(5):1309–1354, 2021

    work page 2021

  7. [7]

    Weight uncertainty in neural networks

    Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wier- stra. Weight uncertainty in neural networks. InProceedings of the 32nd International Conference on Machine Learning (ICML), pages 1613–1622, 2015

    work page 2015

  8. [8]

    Curse of dimensionality in adversarial examples

    Nandish Chattopadhyay, Anupam Chattopadhyay, Sourav Sen Gupta, and Michael Kasper. Curse of dimensionality in adversarial examples. In2019 International Joint Conference on Neural Networks (IJCNN), pages 1–8, 2019. 30

    work page 2019

Show all 61 references
  1. [9]

    Deeplearningensembleforhyperspectralimageclassification

    Yushi Chen, Ying Wang, Yanfeng Gu, Xin He, Pedram Ghamisi, and XiupingJia. Deeplearningensembleforhyperspectralimageclassification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(6):1882–1897, 2019

    2019

  2. [10]

    An inves- tigation into smartphone based weakly supervised activity recognition systems.Pervasive and Mobile Computing, 56:45–56, 2019

    William Duffy, Kevin Curran, Daniel Kelly, and Tom Lunney. An inves- tigation into smartphone based weakly supervised activity recognition systems.Pervasive and Mobile Computing, 56:45–56, 2019

    2019

  3. [11]

    Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar

    Tommaso Furlanello, Zachary C. Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar. Born again neural networks, 2018

    2018

  4. [12]

    Bayesian convolutional neural networks with bernoulli approximate variational inference.arXiv preprint, abs/1506.02158, 2015

    Yarin Gal and Zoubin Ghahramani. Bayesian convolutional neural networks with bernoulli approximate variational inference.arXiv preprint, abs/1506.02158, 2015

    2015 arXiv

  5. [13]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. InProceedings of The 33rd International Conference on Machine Learning, pages 1050–1059. PMLR, 2016

    2016

  6. [14]

    Sufficient conditions for robustness to adversarial examples: a theoretical and empirical study with bayesian neural networks, 2019

    Yarin Gal and Lewis Smith. Sufficient conditions for robustness to adversarial examples: a theoretical and empirical study with bayesian neural networks, 2019. OpenReview preprint, https://openreview. net/forum?id=B1eZRiC9YX

    2019

  7. [15]

    Daniel Garcia-Gonzalez, Daniel Rivero, Enrique Fernandez-Blanco, and Miguel R. Luaces. A public domain dataset for real-life human activity recognition using smartphone sensors.Sensors, 20(8):2200, 2020

    2020

  8. [16]

    Goodfellow, Jonathon Shlens, and Christian Szegedy

    Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples.CoRR, abs/1412.6572, 2014

    2014 arXiv

  9. [17]

    Maybank, and Dacheng Tao

    Jianping Gou, Baosheng Yu, Stephen J. Maybank, and Dacheng Tao. Knowledge distillation: A survey.International Journal of Computer Vision, 129(6):1789–1819, 2021

    2021

  10. [18]

    Practical variational inference for neural networks

    Alex Graves. Practical variational inference for neural networks. In Advances in Neural Information Processing Systems, volume 24, 2011. 31

    2011

  11. [19]

    Hammerla, Shane Halloran, and Thomas Plötz

    Nils Y. Hammerla, Shane Halloran, and Thomas Plötz. Deep, con- volutional, and recurrent models for human activity recognition using wearables. InProceedings of the Twenty-Fifth International Joint Con- ference on Artificial Intelligence (IJCAI), pages 1533–1540, 2016

    2016

  12. [20]

    Large language models memorize sensor datasets! implications on human activity recognition research, 2024

    Harish Haresamudram, Hrudhai Rajasekhar, Nikhil Murlidhar Shanbhogue, and Thomas Ploetz. Large language models memorize sensor datasets! implications on human activity recognition research, 2024

    2024

  13. [21]

    Using self-supervised learning can improve model robustness and uncer- tainty

    Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncer- tainty. InProceedings of the 33rd International Conference on Neural Information Processing Systems, pages 1403–1414, Red Hook, NY, USA,

  14. [22]

    Hinton, Oriol Vinyals, and Jeffrey Dean

    Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network.arXiv preprint, abs/1503.02531, 2015

    2015 arXiv

  15. [23]

    Detecting adversarial examples using data manifolds

    Susmit Jha, Uyeong Jang, Somesh Jha, and Brian Jalaian. Detecting adversarial examples using data manifolds. InMILCOM 2018 - 2018 IEEE Military Communications Conference (MILCOM), pages 547–552, 2018

    2018

  16. [24]

    HARGPT: Are LLMs zero-shot human activity recognizers?, 2024

    Sijie Ji, Xinzhe Zheng, and Chenshu Wu. HARGPT: Are LLMs zero-shot human activity recognizers?, 2024

    2024

  17. [25]

    Sampling- based self-supervised augmented contrast learning for movie recommen- dation systems

    Huixin Jiang, Lingyu Yan, Chunzhi Wang, and Donghua Liu. Sampling- based self-supervised augmented contrast learning for movie recommen- dation systems. In2023 11th International Conference on Information Technology: IoT and Smart City (ITIoTSC), pages 144–148, 2023

    2023

  18. [26]

    Evaluating uncertainty estimation methods on 3d semantic segmentation of point clouds, 2020

    Swaroop K, Nico Hochgeschwender, Paul Plöger, Frank Kirchner, and Matias Valdenegro. Evaluating uncertainty estimation methods on 3d semantic segmentation of point clouds, 2020

    2020

  19. [27]

    Distributional actor-critic ensemble for uncertainty-aware continuous control

    Takuya Kanazawa, Haiyan Wang, and Chetan Gupta. Distributional actor-critic ensemble for uncertainty-aware continuous control. In2022 International Joint Conference on Neural Networks (IJCNN), pages 1–10, 2022. 32

    2022

  20. [28]

    On the geometry of adversarial examples.arXiv preprint, abs/1811.00525, 2018

    Marc Khoury and Dylan Hadfield-Menell. On the geometry of adversarial examples.arXiv preprint, abs/1811.00525, 2018

    2018 arXiv

  21. [29]

    Large language models are zero-shot reasoners.arXiv preprint, abs/2205.11916, 2022

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners.arXiv preprint, abs/2205.11916, 2022

    2022 arXiv

  22. [30]

    Goodfellow, and Samy Bengio

    Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial machine learning at scale. InInternational Conference on Learning Representations (ICLR), 2017

    2017

  23. [31]

    Towards unsuper- vised image captioning with shared multimodal embeddings

    Iro Laina, Christian Rupprecht, and Nassir Navab. Towards unsuper- vised image captioning with shared multimodal embeddings. In2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7413–7423, 2019

    2019

  24. [32]

    Sim- ple and scalable predictive uncertainty estimation using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Sim- ple and scalable predictive uncertainty estimation using deep ensembles. InAdvances in Neural Information Processing Systems, 2017

    2017

  25. [33]

    Safe end-to- end imitation learning for model predictive control.arXiv preprint, abs/1803.10231, 2018

    Keuntaek Lee, Kamil Saigol, and Evangelos Theodorou. Safe end-to- end imitation learning for model predictive control.arXiv preprint, abs/1803.10231, 2018

    2018 arXiv

  26. [34]

    On norm-agnostic robustness of adversarial training.arXiv preprint, abs/1905.06455, 2019

    Bai Li, Changyou Chen, Wenlin Wang, and Lawrence Carin. On norm-agnostic robustness of adversarial training.arXiv preprint, abs/1905.06455, 2019

    1905 arXiv

  27. [35]

    Zechen Li, Shohreh Deldari, Linyao Chen, Hao Xue, and Flora D. Salim. Sensorllm: Aligning large language models with motion sensors for human activity recognition, 2024

    2024

  28. [36]

    A general framework for ensemble distribution distillation.2020 IEEE 30th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1–6, 2020

    Jakob Lindqvist, Amanda Olmin, Fredrik Lindsten, and Lennart Svens- son. A general framework for ensemble distribution distillation.2020 IEEE 30th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1–6, 2020

    2020

  29. [37]

    Clegg, Andrea Cavallaro, and Hamed Haddadi

    Mohammad Malekzadeh, Richard G. Clegg, Andrea Cavallaro, and Hamed Haddadi. Protecting sensory data against sensitive inferences. InProceedings of the 1st Workshop on Privacy by Design in Distributed Systems, W-P2DS’18, pages 2:1–2:6, New York, NY, USA, 2018. ACM. 33

    2018

  30. [38]

    Predictive uncertainty estimation via prior networks

    Andrey Malinin and Mark John Francis Gales. Predictive uncertainty estimation via prior networks. InNeural Information Processing Systems, 2018

    2018

  31. [39]

    Ensemble distribution distillation.arXiv preprint, abs/1905.00076, 2019

    Andrey Malinin, Bruno Mlodozeniec, and Mark John Francis Gales. Ensemble distribution distillation.arXiv preprint, abs/1905.00076, 2019

    1905 arXiv

  32. [40]

    Berkay Celik

    Patrick McDaniel, Nicolas Papernot, and Z. Berkay Celik. Machine learning in adversarial settings.IEEE Security & Privacy, 14(3):68–72, 2016

    2016

  33. [41]

    On the validity of bayesian neural networks for uncertainty estimation

    John Mitros and Brian Mac Namee. On the validity of bayesian neural networks for uncertainty estimation. InIrish Conference on Artificial Intelligence and Cognitive Science (AICS), 2019

    2019

  34. [42]

    Deep convolu- tional feature transfer across mobile activity recognition domains, sensor modalities and locations

    Francisco Javier Ordóñez Morales and Daniel Roggen. Deep convolu- tional feature transfer across mobile activity recognition domains, sensor modalities and locations. InProceedings of the 2016 ACM International Symposium on Wearable Computers (ISWC), pages 92–99, 2016

    2016

  35. [43]

    Pong, Glen Berseth, and Sergey Levine

    Ashvin Nair, Shikhar Bahl, Alexander Khazatsky, Vitchyr H. Pong, Glen Berseth, and Sergey Levine. Contextual imagined goals for self-supervised robotic learning.arXiv preprint, abs/1910.11670, 2019

    1910 arXiv

  36. [44]

    Henry Friday Nweke, Ying Wah Teh, Mohammed Ali Al-garadi, and Uzoma Rita Alo. Deep learning algorithms for human activity recognition using mobile and wearable sensor networks: State of the art and research challenges.Expert Systems with Applications, 105:233–261, 2018

    2018

  37. [45]

    Sculley, Sebas- tian Nowozin, Joshua V

    Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, D. Sculley, Sebas- tian Nowozin, Joshua V. Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift. InAdvances in Neural Information Proc...

    2019

  38. [46]

    Crafting adversarial input sequences for recurrent neural networks

    Nicolas Papernot, Patrick McDaniel, Ananthram Swami, and Richard Ha- rang. Crafting adversarial input sequences for recurrent neural networks. InMILCOM 2016 - 2016 IEEE Military Communications Conference, pages 49–54, 2016. 34

    2016

  39. [47]

    PAMAP2 Physical Activity Monitoring

    Attila Reiss. PAMAP2 Physical Activity Monitoring. UCI Machine Learning Repository, 2012. DOI: https://doi.org/10.24432/C5NW2H

    2012 doi

  40. [48]

    Multi-task self- supervised learning for human activity detection.Proc

    Aaqib Saeed, Tanir Ozcelebi, and Johan Lukkien. Multi-task self- supervised learning for human activity detection.Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., 3(2):61:1–61:30, 2019

    2019

  41. [49]

    A comprehensive guide to bayesian convolutional neural network with variational inference

    Kumar Shridhar, Felix Laumann, and Marcus Liwicki. A comprehensive guide to bayesian convolutional neural network with variational inference. arXiv preprint, abs/1901.02731, 2019

    1901 arXiv

  42. [50]

    Improving the generalization of adversarial training with domain adaptation, 2018

    Chuanbiao Song, Kun He, Liwei Wang, and John Hopcroft. Improving the generalization of adversarial training with domain adaptation, 2018

    2018

  43. [51]

    Joshua S. Speagle. A conceptual introduction to markov chain monte carlo methods, 2020

    2020

  44. [52]

    Smart devices are different: Assessing and mitigating mobile sensing heterogeneities for activity recognition

    Allan Stisen, Henrik Blunck, Sourav Bhattacharya, Thor Siiger Prentow, Mikkel Baun Kjærgaard, Anind Dey, Tobias Sonne, and Mads Møller Jensen. Smart devices are different: Assessing and mitigating mobile sensing heterogeneities for activity recognition. InProceedings of the 13...

    2015

  45. [53]

    Smartphone-based human activity pattern identification using unsupervised learning

    Dipanwita Thakur and Suparna Biswas. Smartphone-based human activity pattern identification using unsupervised learning. In Mukesh Saraswat, Sarbani Roy, Chandreyee Chowdhury, and Amir H. Gandomi, editors,Proceedings of International Conference on Data Science and Applications...

    2022

  46. [54]

    J.D. Tygar. Adversarial machine learning.IEEE Internet Computing, 15(5):4–6, 2011

    2011

  47. [55]

    Transferable adversarial robustness using adversarially trained autoen- coders.CoRR, abs/1909.05921, 2019

    Pratik Vaishnavi, Kevin Eykholt, Atul Prakash, and Amir Rahmati. Transferable adversarial robustness using adversarially trained autoen- coders.CoRR, abs/1909.05921, 2019

    1909 arXiv

  48. [56]

    Deep sub-ensembles for fast uncertainty esti- mation in image classification.arXiv preprint, abs/1910.08168, 2019

    Matias Valdenegro-Toro. Deep sub-ensembles for fast uncertainty esti- mation in image classification.arXiv preprint, abs/1910.08168, 2019. 35

    1910 arXiv

  49. [57]

    Deep learning for sensor-based activity recognition: A survey.Pattern Recognition Letters, 119:3–11, 2019

    Jindong Wang, Yiqiang Chen, Shuji Hao, Xiaohui Peng, and Lisha Hu. Deep learning for sensor-based activity recognition: A survey.Pattern Recognition Letters, 119:3–11, 2019

    2019

  50. [58]

    Improving single-step adversarial training by local smoothing

    Shaopeng Wang, Yanhong Huang, Jianqi Shi, Yang Yang, and Xin Guo. Improving single-step adversarial training by local smoothing. In2023 International Joint Conference on Neural Networks (IJCNN), pages 1–8, 2023

    2023

  51. [59]

    Deep convolutional neural networks on multi- channel time series for human activity recognition

    Jian Bo Yang, Minh Nhut Nguyen, Phyo Phyo San, Xiao Li Li, and Shonali Krishnaswamy. Deep convolutional neural networks on multi- channel time series for human activity recognition. InProceedings of the 24th International Conference on Artificial Intelligence (IJCAI), pages 39...

    2015

  52. [60]

    Strength-adaptive adversarial training

    Chaojian Yu, Dawei Zhou, Li Shen, Jun Yu, Bo Han, Mingming Gong, Nannan Wang, and Tongliang Liu. Strength-adaptive adversarial training. arXiv preprint, abs/2210.01288, 2022. 36

    2022 arXiv

  53. [2019]

    Curran Associates Inc

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.