Pith. sign in

REVIEW 2 major objections 6 minor 65 references

Sense and Learn: Self-Supervision for Omnipresent Sensors

T0 review · 2 major / 6 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read This paper claims that self-supervised pre-training on unlabeled multisensor data lets downstream classifiers match or approach fully supervised performance, cutting the labeled data needed to as few as five examples per class.

desk verdict Solid empirical case for self-supervised pre-training on multisensor data, with the real caveat that the headline low-data gains are selected on downstream test performance. read the letter →

arxiv 2009.13233 v2 pith:5KBVNF3K submitted 2020-09-28 cs.LG stat.ML

Aaqib Saeed , Victor Ungureanu , Beat Gfeller This is my paper · ORCID
classification cs.LGstat.ML
keywords self-supervisedlearninglow-datasensorsunsupervisedactivityrecognitionsleepstagescoringneuralnetworkstime-series
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

Unlabeled data from the sensors in phones, wearables, and other devices is abundant, while labeled data is scarce and costly. This paper argues that self-supervised pre-training can convert that unlabeled abundance into a reusable feature extractor, so that downstream tasks such as activity recognition, sleep stage scoring, stress detection, and WiFi-based activity sensing can be learned from very few labeled examples. The framework, called Sense and Learn, defines eight auxiliary tasks that generate their own labels from raw multisensor signals, pre-trains a small temporal convolutional network on those tasks, and then uses the result as an initialization or frozen feature extractor for end tasks. Across eight public datasets, the authors report that fine-tuning the pre-trained network matches or approaches fully supervised performance, and that as few as five to ten labeled instances per class can yield large improvements over training from scratch. If the claim holds, it makes personalized and on-device learning practical, because only a handful of labels would be needed per user and raw data need never leave the device.

What carries the argument

At the center is a multi-stream temporal convolutional network (TCN) in which each input modality, such as accelerometer and gyroscope, has its own encoder and a shared encoder fuses the streams into a joint embedding. The pretraining phase solves eight auxiliary tasks whose labels are generated on the fly from the unlabeled data: detecting blended sensor inputs, predicting the blending magnitude, predicting summary statistics of a masked window, recognizing which transformation was applied to a signal, predicting temporal shift amounts, denoising one modality using another, spotting an odd segment spliced from another modality, and a triplet metric-learning objective. After pretraining, the task-specific heads are discarded and the encoder is either kept frozen with a linear classifier on top or fine-tuned; the low-data experiments use the encoder as an initialization for a classifier trained with only a few labeled instances per class. The design choice of a small, fixed architecture is meant to show that the gains come from the self-supervised objectives rather than from model capacity.

What would settle it

Run each of the eight pretext tasks on phase-randomized or channel-shuffled sensor inputs that preserve amplitude statistics but destroy semantic content; any task that still solves its proxy at near-perfect accuracy, while the pre-trained encoder yields no low-data transfer gain, would show the task is being solved by shortcuts rather than by learning meaningful features.

Watch

Extended reading notes

Core claim

The paper's central claim is that a temporal convolutional network pre-trained on eight self-supervised pretext tasks over unlabeled multisensor input learns generic, semantically meaningful representations, and that these representations are strong enough to support linear classifiers whose accuracy rivals fully supervised end-to-end training, to close the remaining gap when the shared encoder layers are fine-tuned, and to deliver a significant boost in the low-data regime with as few as 5 to 10 labeled instances per class. The same pre-trained features transfer across related activity-recognition datasets, improving target performance even when the source and target data come from different devices and recording protocols. The authors present this as evidence that self-supervision can replace most of the labeling effort in sensing applications.

Load-bearing premise

The framework's gains rest on the assumption that its eight self-supervised tasks force the network to learn meaningful signal features rather than being solved by trivial low-level cues that would not transfer to new data.

Editorial extensions

If this is right

  • Sensor-based classifiers for health and context monitoring could be trained with tens of labels per user instead of hundreds or thousands, making personalization practical.
  • A single unlabeled pretraining stage on one dataset can be reused for several related downstream tasks, reducing per-task data collection and engineering.
  • Because the pretext data generator runs on the fly and the architecture is small, the same pipeline is a candidate for on-device federated learning where raw sensor data stays on the user's device.
  • In several of the paper's fine-tuning experiments, self-supervised initialization improves accuracy even when all labeled data is available, suggesting it can be a drop-in replacement for random initialization in existing supervised pipelines.

Reading between the lines

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

  • Editor's inference: the paper evaluates each pretext task in isolation, but training them jointly in a multi-task objective could yield more stable representations than selecting a single task per dataset.
  • Editor's inference: the shortcut-learning concern implies a concrete diagnostic, namely measuring how much of each pretext task's accuracy can be explained by simple signal statistics such as amplitude or spectral content, and removing tasks that are trivially solvable before deployment.
  • Editor's inference: an untested extension is cross-modal transfer, for example pretraining on inertial data and evaluating on physiological signals; if those tasks share temporal structure, some of the reported transfer gains may carry over.
  • Editor's inference: the low-data results pool labeled instances across users, so a harder and more practical test is per-user personalization, where the five labels come from the same new user's device.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper proposes Sense and Learn, a self-supervised learning framework for multisensor time-series data. The framework defines eight pretext tasks (blend detection, fusion magnitude prediction, feature prediction from masked windows, transformation recognition, temporal shift prediction, modality denoising, odd segment recognition, and triplet metric learning) and uses a multi-stream temporal convolutional network with modality-specific encoders and a shared encoder. The authors evaluate the learned representations on eight public datasets across four application domains (activity recognition, sleep stage scoring, stress detection, and WiFi sensing), using linear separability, fine-tuning, low-data (5 and 10 labeled instances per class), and transfer-learning settings. The central claim is that self-supervised pre-training provides a strong initialization that significantly boosts downstream performance in the low-data regime, approaching or exceeding fully-supervised baselines trained on all labels.

Significance. If the central claim holds, the paper provides a substantial contribution to self-supervised representation learning for sensor data: it demonstrates broad applicability across heterogeneous modalities, uses a small architecture suitable for on-device learning, and shows label-efficiency gains that are practically important. The evaluation is extensive (eight datasets, 5-fold cross-validation, multiple evaluation protocols, comparisons against autoencoders and fully-supervised baselines) and the results are reported with means and standard deviations. However, the low-data and transfer-learning claims are weakened by the paper's reliance on selecting the best-performing pretext task per dataset apparently on the basis of downstream test performance, which the stress-test note correctly identifies as a load-bearing issue. The paper also explicitly acknowledges shortcut learning as a limitation but does not test whether any of the proposed pretext tasks are trivially solvable; this is a correctness risk that should be addressed before the claims can be taken at face value.

major comments (2)
  1. [Section 4.3 (low-data regime) and Figure 5] The central low-data claim that pre-training 'significantly boost[s] the performance' with as few as 5 labeled instances per class is reported only for the best-performing pretext task per dataset (e.g., temporal shift on HHAR and MobiAct). The paper never states whether this selection is made on a validation split or after inspecting the downstream test labels. Figure 5 explicitly reports 'mean results only of the best performing auxiliary task' for each source-target pair, and Section 5 concedes that 'a practitioner cannot be certain which self-supervised task will work best for a new application.' If the choice is made on test performance, the reported gains are an upper bound, not the expected performance of the framework under a fixed selection protocol. To make the central claim actionable, the authors should either define a selection rule that uses only unlabeled data and the few labeled training instances, or report the average or robust performance across all pretext tasks in the low-data setting.
  2. [Section 5 (Impact and Limitations)] The paper explicitly states that 'networks can use shortcuts to solve proxy task without learning anything useful' and that discovering a model relying on shortcuts for time-series or multisensor inputs 'is an unsolved problem and could be challenging to detect.' However, no analysis is provided to rule out shortcut solutions for the proposed tasks, such as blend detection (which might be solved by distinguishing blended from clean inputs via signal energy) or modality denoising (which might be solved by amplitude or scale cues). If any pretext task is shortcut-solvable, the claimed representation quality and low-data gains may not transfer to new data. The authors should report proxy-task accuracies on held-out data and include a diagnostic such as training the downstream classifier on features from a network pretrained with shuffled proxy labels, to confirm that the gains are due to the specific self-supervised objectives rather than low-level artifacts.
minor comments (6)
  1. [Tables 2-5 and Figure captions] The label 'Tripet Loss' appears repeatedly; it should be 'Triplet Loss.'
  2. [Algorithm 1] The mathematical symbols in Algorithm 1 are garbled (e.g., the gradient and summation signs appear as '∇𝜃𝐹' and '˝𝐵'); the algorithm should be typeset correctly.
  3. [Table 1] There is a typo in the table caption: 'experiements' should be 'experiments.'
  4. [Section 4.3] There is an unresolved reference in the sentence 'as it has been shown to improve adversarial robustness and uncertainty of deep models as compared to purely supervised methods [? ]'; the citation needs to be completed.
  5. [Section 3.2] The paper claims 'eight novel auxiliary tasks' in the introduction and Section 6, but the Transformation Recognition task is a simplification of the signal transformation recognition task of [45]. The novelty relative to [45] should be stated more precisely.
  6. [General] No code availability statement or link to an implementation is provided; releasing code would help reproducibility given the complexity of the data-generation procedures.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the downstream claims are evaluated against external public labels, and the self-cited prior work is a baseline task, not a load-bearing derivation.

full rationale

The paper's central claim is empirical: self-supervised pre-training on unlabeled sensor data improves downstream performance in a low-data regime. The pretext tasks generate their own supervisory labels from the input, which is the defining mechanism of self-supervised learning and not circular for the downstream claims, because the downstream tasks are evaluated with external public labels (HHAR, MobiAct, UCI HAR, Sleep-EDF, MIT DriverDb, WiFi CSI). No downstream F-score or kappa is used to define the pretext objectives or to construct the embeddings. The only self-citation is to [45], which proposed signal transformation recognition; the paper explicitly reformulates that task as multi-class classification and evaluates it as one of eight auxiliary tasks, while the low-data and transfer claims are supported by experiments across many tasks and datasets. The paper contains no theoretical derivation whose output reduces to a fitted parameter, no imported uniqueness theorem, and no ansatz smuggled in via citation. The reported selection of the 'best performing auxiliary task' in Figure 5 raises a possible evaluation-protocol concern about whether task selection used the test set, but that is a correctness and reporting issue, not a circularity of the kind where a predicted quantity is equivalent to an input by construction. Therefore the derivation chain is self-contained with respect to the external benchmarks, and the circularity score is 0.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the choice of pretext tasks and the fixed TCN architecture. The main 'free parameter' is the per-dataset selection of the surrogate task, which is not held out. There are no newly invented physical or mathematical entities.

free parameters (1)
  • Per-dataset pretext task selection = unknown (best task reported per dataset)
    In Figures 3 and 5, the authors report results for the 'best performing auxiliary task' or 'optimal auxiliary tasks' for each source-target combination, chosen after looking at downstream test performance. This is a model-selection choice that could inflate the reported gains if the test set was used for task selection.
assumptions (3)
  • domain assumption The two or more input modalities in each dataset are synchronized and reflect the same underlying phenomenon, so cross-modal operations (blending, denoising, odd-segment insertion) yield meaningful learning signals.
    Invoked throughout Section 3.2 for blend detection, modality denoising, and odd segment recognition; if the modalities are decoupled, these tasks may be trivially solvable or learn sensor-specific artifacts.
  • domain assumption The TCN architecture and the fixed set of hyperparameters (architecture, learning rate, L2 rate, SELU, 30 epochs) are sufficient to learn useful representations for all eight datasets.
    Section 3.3 fixes one architecture for all tasks; the paper does not ablate architecture capacity or show that the results hold with other backbones.
  • domain assumption Unlabeled pre-training data comes from the same or closely related distribution as the downstream task data.
    Pre-training uses the training split of each dataset; the transfer experiments use different but related activity-recognition datasets. The paper does not test out-of-distribution pre-training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sense and Learn: Self-Supervision for Omnipresent Sensors." pith.science (2026). https://pith.science/paper/5KBVNF3K

@misc{pith2026200913233,
  author       = {Pith},
  title        = {Pith review of: Sense and Learn: Self-Supervision for Omnipresent Sensors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5KBVNF3K}},
  note         = {Machine review of arXiv:2009.13233}
}
read the original abstract

Learning general-purpose representations from multisensor data produced by the omnipresent sensing systems (or IoT in general) has numerous applications in diverse use cases. Existing purely supervised end-to-end deep learning techniques depend on the availability of a massive amount of well-curated data, acquiring which is notoriously difficult but required to achieve a sufficient level of generalization on a task of interest. In this work, we leverage the self-supervised learning paradigm towards realizing the vision of continual learning from unlabeled inputs. We present a generalized framework named Sense and Learn for representation or feature learning from raw sensory data. It consists of several auxiliary tasks that can learn high-level and broadly useful features entirely from unannotated data without any human involvement in the tedious labeling process. We demonstrate the efficacy of our approach on several publicly available datasets from different domains and in various settings, including linear separability, semi-supervised or few shot learning, and transfer learning. Our methodology achieves results that are competitive with the supervised approaches and close the gap through fine-tuning a network while learning the downstream tasks in most cases. In particular, we show that the self-supervised network can be utilized as initialization to significantly boost the performance in a low-data regime with as few as 5 labeled instances per class, which is of high practical importance to real-world problems. Likewise, the learned representations with self-supervision are found to be highly transferable between related datasets, even when few labeled instances are available from the target domains. The self-learning nature of our methodology opens up exciting possibilities for on-device continual learning.

Figures

Figures reproduced from arXiv: 2009.13233 by the authors.

Figure 1
Figure 1. Illustration of our Sense and Learn representation learning framework. A deep neural network is pre-trained with self-supervision using input modalities from large unlabeled sensory data, such as inertial measurements (or electroencephalogram, heart rate, and channel state information). The learned network can then be utilized as a feature-extractor or initialization for rapidly solving downstream tasks of interest … view at source ↗
Figure 2
Figure 2. A multistream neural network architecture for learning representations from multiple sensory inputs. [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Contribution of self-supervised pre-training for improving end-task performance with few labeled [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Generalization of the self-supervised representations under transfer learning setting. We evaluate the [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: Contribution of self-supervised learning, and fine-tuning of the transferred networks in learning [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]
Figure 6
Figure 6. Figure 6: Class distribution of the datasets used in evaluation. [PITH_FULL_IMAGE:figures/full_fig_p029_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

65 extracted references · 65 canonical work pages

  1. [45]

    Multi-task self-supervised learning for human activity detection

    Saeed, A., Ozcelebi, T., Lukkien, J., 2019. Multi-task self-supervised learning for human activity detection. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 3, 61

    work page 2019

  2. [1]

    A public domain dataset for human activity recognition using smartphones., in: Esann

    Anguita, D., Ghio, A., Oneto, L., Parra, X., Reyes-Ortiz, J.L., 2013. A public domain dataset for human activity recognition using smartphones., in: Esann

    work page 2013

  3. [2]

    An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling

    Bai, S., Kolter, J.Z., Koltun, V., 2018. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271

    work page Pith review arXiv 2018

  4. [3]

    Using unlabeled data in a sparse-coding framework for human activity recognition

    Bhattacharya, S., Nurmi, P., Hammerla, N., Plötz, T., 2014. Using unlabeled data in a sparse-coding framework for human activity recognition. Pervasive and Mobile Computing 15, 242–262

    work page 2014

  5. [4]

    Towards Federated Learning at Scale: System Design

    Bonawitz, K., Eichner, H., Grieskamp, W., Huba, D., Ingerman, A., Ivanov, V., Kiddon, C., Konecny, J., Mazzocchi, S., McMahan, H.B., et al., 2019. Towards federated learning at scale: System design. arXiv preprint arXiv:1902.01046

    work page Pith review arXiv 2019

  6. [5]

    Cross-position activity recognition with stratified transfer learning

    Chen, Y., Wang, J., Huang, M., Yu, H., 2019. Cross-position activity recognition with stratified transfer learning. Pervasive and Mobile Computing 57, 1–13

    work page 2019

  7. [6]

    Time Series Change Point Detection with Self-Supervised Contrastive Predictive Coding

    Deldari, S., Smith, D.V., Xue, H., Salim, F.D., 2020. Time-series change point detection with self-supervised contrastive predictive coding. arXiv preprint arXiv:2011.14097

    work page Pith review arXiv 2020

  8. [7]

    Lazy Prices

    Devlin, J., Chang, M.W., Lee, K., Toutanova, K., 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805

    work page Pith review arXiv 2018

Show all 65 references
  1. [8]

    Unsupervised visual representation learning by context prediction, in: Proceedings of the IEEE International Conference on Computer Vision, pp

    Doersch, C., Gupta, A., Efros, A.A., 2015. Unsupervised visual representation learning by context prediction, in: Proceedings of the IEEE International Conference on Computer Vision, pp. 1422–1430

    2015

  2. [9]

    Real-valued (medical) time series generation with recurrent conditional gans

    Esteban, C., Hyland, S.L., Rätsch, G., 2017. Real-valued (medical) time series generation with recurrent conditional gans. arXiv preprint arXiv:1706.02633

    2017 arXiv

  3. [10]

    On the personalization of classification models for human activity recognition

    Ferrari, A., Micucci, D., Mobilio, M., Napoletano, P., 2020. On the personalization of classification models for human activity recognition. IEEE Access 8, 32066–32079. doi: 10.1109/ACCESS.2020.2973425. 26 Aaqib Saeed, Victor Ungureanu, and Beat Gfeller

    2020

  4. [11]

    Shortcut learning in deep neural networks

    Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A., 2020. Shortcut learning in deep neural networks. arXiv preprint arXiv:2004.07780

    2020 arXiv

  5. [12]

    Unsupervised representation learning by predicting image rotations

    Gidaris, S., Singh, P., Komodakis, N., 2018. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728

    2018 arXiv

  6. [13]

    Gjoreski, M., Kalabakov, S., Luštrek, M., Gams, M., Gjoreski, H., 2019. Cross-dataset deep transfer learning for activity recognition, in: Adjunct Proceedings of the 2019 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2019 ACM I...

    2019

  7. [14]

    Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals

    Goldberger, A.L., Amaral, L.A., Glass, L., Hausdorff, J.M., Ivanov, P.C., Mark, R.G., Mietus, J.E., Moody, G.B., Peng, C.K., Stanley, H.E., 2000. Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals. Circulation 101, e215–e220

    2000

  8. [15]

    Efficient sleep stage recognition system based on eeg signal using k-means clustering based feature weighting

    Güneş, S., Polat, K., Yosunkaya, Ş., 2010. Efficient sleep stage recognition system based on eeg signal using k-means clustering based feature weighting. Expert Systems with Applications 37, 7922–7928

    2010

  9. [16]

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

    Hammerla, N.Y., Halloran, S., Plötz, T., 2016. Deep, convolutional, and recurrent models for human activity recognition using wearables. arXiv preprint arXiv:1604.08880

    2016 arXiv

  10. [17]

    Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network

    Hannun, A.Y., Rajpurkar, P., Haghpanahi, M., Tison, G.H., Bourn, C., Turakhia, M.P., Ng, A.Y., 2019. Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nature medicine 25, 65

    2019

  11. [18]

    Detecting stress during real-world driving tasks using physiological sensors

    Healey, J.A., Picard, R.W., 2005. Detecting stress during real-world driving tasks using physiological sensors. IEEE Transactions on intelligent transportation systems 6, 156–166

    2005

  12. [19]

    Time series segmentation for context recognition in mobile devices, in: Proceedings 2001 IEEE International Conference on Data Mining, IEEE

    Himberg, J., Korpiaho, K., Mannila, H., Tikanmaki, J., Toivonen, H.T., . Time series segmentation for context recognition in mobile devices, in: Proceedings 2001 IEEE International Conference on Data Mining, IEEE. pp. 203–210

    2001

  13. [20]

    Distilling the knowledge in a neural network

    Hinton, G., Vinyals, O., Dean, J., 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531

    2015 arXiv

  14. [21]

    Unsupervised feature extraction by time-contrastive learning and nonlinear ica, in: Advances in Neural Information Processing Systems, pp

    Hyvarinen, A., Morioka, H., 2016. Unsupervised feature extraction by time-contrastive learning and nonlinear ica, in: Advances in Neural Information Processing Systems, pp. 3765–3773

    2016

  15. [22]

    Analysis of a sleep-dependent neuronal feedback loop: the slow-wave microcontinuity of the eeg

    Kemp, B., Zwinderman, A.H., Tuk, B., Kamphuisen, H.A., Oberye, J.J., 2000. Analysis of a sleep-dependent neuronal feedback loop: the slow-wave microcontinuity of the eeg. IEEE Transactions on Biomedical Engineering 47, 1185–1194

    2000

  16. [23]

    Adam: A method for stochastic optimization

    Kingma, D.P., Ba, J., 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  17. [24]

    Cooperative learning of audio and video models from self-supervised synchronization, in: Advances in Neural Information Processing Systems, pp

    Korbar, B., Tran, D., Torresani, L., 2018. Cooperative learning of audio and video models from self-supervised synchronization, in: Advances in Neural Information Processing Systems, pp. 7763–7774

    2018

  18. [25]

    Learning machines and sleeping brains: automatic sleep stage classification using decision-tree multi-class support vector machines

    Lajnef, T., Chaibi, S., Ruby, P., Aguera, P.E., Eichenlaub, J.B., Samet, M., Kachouri, A., Jerbi, K., 2015. Learning machines and sleeping brains: automatic sleep stage classification using decision-tree multi-class support vector machines. Journal of neuroscience methods 250, 94–105

    2015

  19. [26]

    Learning representations for automatic colorization, in: European Conference on Computer Vision, Springer

    Larsson, G., Maire, M., Shakhnarovich, G., 2016. Learning representations for automatic colorization, in: European Conference on Computer Vision, Springer. pp. 577–593

    2016

  20. [27]

    Unsupervised feature learning for human activity recognition using smartphone sensors, in: Mining intelligence and knowledge exploration

    Li, Y., Shi, D., Ding, B., Liu, D., 2014. Unsupervised feature learning for human activity recognition using smartphone sensors, in: Mining intelligence and knowledge exploration. Springer, pp. 99–107

    2014

  21. [28]

    Multivariate time series imputation with generative adversarial networks, in: Advances in Neural Information Processing Systems, pp

    Luo, Y., Cai, X., Zhang, Y., Xu, J., et al., 2018. Multivariate time series imputation with generative adversarial networks, in: Advances in Neural Information Processing Systems, pp. 1596–1607

    2018

  22. [29]

    Conv-tasnet: Surpassing ideal time–frequency magnitude masking for speech separation

    Luo, Y., Mesgarani, N., 2019. Conv-tasnet: Surpassing ideal time–frequency magnitude masking for speech separation. IEEE/ACM Trans. Audio, Speech and Lang. Proc. 27, 1256–1266

    2019

  23. [30]

    Protecting sensory data against sensitive inferences, in: Proceedings of the 1st Workshop on Privacy by Design in Distributed Systems, ACM

    Malekzadeh, M., Clegg, R.G., Cavallaro, A., Haddadi, H., 2018. Protecting sensory data against sensitive inferences, in: Proceedings of the 1st Workshop on Privacy by Design in Distributed Systems, ACM. p. 2

    2018

  24. [31]

    Learning deep physiological models of affect

    Martinez, H.P., Bengio, Y., Yannakakis, G.N., 2013. Learning deep physiological models of affect. IEEE Computational intelligence magazine 8, 20–33

    2013

  25. [32]

    Incremental real-time personalization in human activity recognition using domain adaptive batch normalization

    Mazankiewicz, A., Böhm, K., Bergés, M., 2020. Incremental real-time personalization in human activity recognition using domain adaptive batch normalization. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 4, 1–20

    2020

  26. [33]

    Communication-efficient learning of deep networks from decentralized data, in: Artificial Intelligence and Statistics, PMLR

    McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A., 2017. Communication-efficient learning of deep networks from decentralized data, in: Artificial Intelligence and Statistics, PMLR. pp. 1273–1282

    2017

  27. [34]

    Distributed representations of words and phrases and their compositionality, in: Advances in neural information processing systems, pp

    Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J., 2013. Distributed representations of words and phrases and their compositionality, in: Advances in neural information processing systems, pp. 3111–3119

    2013

  28. [35]

    Pruning convolutional neural networks for resource efficient inference

    Molchanov, P., Tyree, S., Karras, T., Aila, T., Kautz, J., 2016. Pruning convolutional neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440

    2016 arXiv

  29. [36]

    Improvements to context based self-supervised learning, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Nathan Mundhenk, T., Ho, D., Chen, B.Y., 2018. Improvements to context based self-supervised learning, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 9339–9348

    2018

  30. [37]

    Unsupervised learning of visual representations by solving jigsaw puzzles, in: European Conference on Computer Vision, Springer

    Noroozi, M., Favaro, P., 2016. Unsupervised learning of visual representations by solving jigsaw puzzles, in: European Conference on Computer Vision, Springer. pp. 69–84. Sense and Learn: Self-Supervision for Omnipresent Sensors 27

    2016

  31. [38]

    Representation learning with contrastive predictive coding

    Oord, A.v.d., Li, Y., Vinyals, O., 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748

    2018 arXiv

  32. [39]

    Toward machine emotional intelligence: Analysis of affective physiological state

    Picard, R.W., Vyzas, E., Healey, J., 2001. Toward machine emotional intelligence: Analysis of affective physiological state. IEEE Transactions on Pattern Analysis & Machine Intelligence , 1175–1191

    2001

  33. [40]

    Feature learning for activity recognition in ubiquitous computing, in: Twenty-second international joint conference on artificial intelligence

    Plötz, T., Hammerla, N.Y., Olivier, P.L., 2011. Feature learning for activity recognition in ubiquitous computing, in: Twenty-second international joint conference on artificial intelligence

    2011

  34. [41]

    Multimodal deep learning for activity and context recognition

    Radu, V., Tong, C., Bhattacharya, S., Lane, N.D., Mascolo, C., Marina, M.K., Kawsar, F., 2018. Multimodal deep learning for activity and context recognition. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 1, 1–27

    2018

  35. [42]

    Faster r-cnn: Towards real-time object detection with region proposal networks, in: Advances in neural information processing systems, pp

    Ren, S., He, K., Girshick, R., Sun, J., 2015. Faster r-cnn: Towards real-time object detection with region proposal networks, in: Advances in neural information processing systems, pp. 91–99

    2015

  36. [43]

    Transition-aware human activity recognition using smartphones

    Reyes-Ortiz, J.L., Oneto, L., Samà, A., Parra, X., Anguita, D., 2016. Transition-aware human activity recognition using smartphones. Neurocomputing 171, 754–767

    2016

  37. [44]

    Learning classification with unlabeled data, in: Advances in neural information processing systems, pp

    de Sa, V.R., 1994. Learning classification with unlabeled data, in: Advances in neural information processing systems, pp. 112–119

    1994

  38. [46]

    Model adaptation and personalization for physiological stress detection, in: 2018 IEEE 5th International Conference on Data Science and Advanced Analytics (DSAA), IEEE

    Saeed, A., Ozcelebi, T., Lukkien, J., van Erp, J., Trajanovski, S., 2018. Model adaptation and personalization for physiological stress detection, in: 2018 IEEE 5th International Conference on Data Science and Advanced Analytics (DSAA), IEEE. pp. 209–216

    2018

  39. [47]

    Personalized driver stress detection with multi-task neural networks using physiological signals

    Saeed, A., Trajanovski, S., 2017. Personalized driver stress detection with multi-task neural networks using physiological signals. arXiv preprint arXiv:1711.06116

    2017 arXiv

  40. [48]

    Self-supervised ecg representation learning for emotion recognition

    Sarkar, P., Etemad, A., 2020. Self-supervised ecg representation learning for emotion recognition. IEEE Transactions on Affective Computing

    2020

  41. [49]

    Making the world differentiable: On using self-supervised fully recurrent neural networks for dynamic reinforcement learning and planning in non-stationary environments

    Schmidhuber, J., 1990. Making the world differentiable: On using self-supervised fully recurrent neural networks for dynamic reinforcement learning and planning in non-stationary environments

    1990

  42. [50]

    Time-contrastive networks: Self-supervised learning from video

    Sermanet, P., Lynch, C., Chebotar, Y., Hsu, J., Jang, E., Schaal, S., Levine, S., . Time-contrastive networks: Self-supervised learning from video. Proceedings of International Conference in Robotics and Automation (ICRA)

  43. [51]

    Mastering the game of go without human knowledge

    Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., et al., 2017. Mastering the game of go without human knowledge. Nature 550, 354

    2017

  44. [52]

    Stisen, A., Blunck, H., Bhattacharya, S., Prentow, T.S., Kjærgaard, M.B., Dey, A., Sonne, T., Jensen, M.M., 2015. Smart devices are different: Assessing and mitigatingmobile sensing heterogeneities for activity recognition, in: Proceedings of the 13th ACM Conference on Embedde...

    2015

  45. [53]

    Deepsleepnet: A model for automatic sleep stage scoring based on raw single-channel eeg

    Supratak, A., Dong, H., Wu, C., Guo, Y., 2017. Deepsleepnet: A model for automatic sleep stage scoring based on raw single-channel eeg. IEEE Transactions on Neural Systems and Rehabilitation Engineering 25, 1998–2008

    2017

  46. [54]

    Self-supervised audio representation learning for mobile devices

    Tagliasacchi, M., Gfeller, B., de Chaumont Quitry, F., Roblek, D., 2019. Self-supervised audio representation learning for mobile devices. arXiv preprint arXiv:1905.11796

    2019 arXiv

  47. [55]

    Improving generalization and stability of generative adversarial networks

    Thanh-Tung, H., Tran, T., Venkatesh, S., 2019. Improving generalization and stability of generative adversarial networks. arXiv preprint arXiv:1902.03984

    2019 arXiv

  48. [56]

    Contrastive multiview coding

    Tian, Y., Krishnan, D., Isola, P., 2019. Contrastive multiview coding. arXiv preprint arXiv:1906.05849

    2019 arXiv

  49. [57]

    The mobifall dataset: Fall detection and classification with a smartphone

    Vavoulas, G., Pediaditis, M., Chatzaki, C., Spanakis, E.G., Tsiknakis, M., 2014. The mobifall dataset: Fall detection and classification with a smartphone. IJMSTR 2, 44–56

    2014

  50. [58]

    Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion

    Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., Manzagol, P.A., Bottou, L., 2010. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research 11

    2010

  51. [59]

    Sensegan: Enabling deep learning for internet of things with a semi-supervised framework

    Yao, S., Zhao, Y., Shao, H., Zhang, C., Zhang, A., Hu, S., Liu, D., Liu, S., Su, L., Abdelzaher, T., 2018. Sensegan: Enabling deep learning for internet of things with a semi-supervised framework. Proceedings of the ACM on interactive, mobile, wearable and ubiquitous technolog...

    2018

  52. [60]

    A survey on behavior recognition using wifi channel state information

    Yousefi, S., Narui, H., Dayal, S., Ermon, S., Valaee, S., 2017. A survey on behavior recognition using wifi channel state information. IEEE Communications Magazine 55, 98–104

    2017

  53. [61]

    Wavesplit: End-to-end speech separation by speaker clustering

    Zeghidour, N., Grangier, D., 2020. Wavesplit: End-to-end speech separation by speaker clustering. arXiv preprint arXiv:2002.08933

    2020 arXiv

  54. [62]

    S4l: Self-supervised semi-supervised learning

    Zhai, X., Oliver, A., Kolesnikov, A., Beyer, L., 2019. S4l: Self-supervised semi-supervised learning. arXiv preprint arXiv:1905.03670

    2019 arXiv

  55. [63]

    mixup: Beyond empirical risk minimization

    Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D., 2017. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 . 28 Aaqib Saeed, Victor Ungureanu, and Beat Gfeller

    2017 arXiv

  56. [64]

    Colorful image colorization, in: European conference on computer vision, Springer

    Zhang, R., Isola, P., Efros, A.A., 2016a. Colorful image colorization, in: European conference on computer vision, Springer. pp. 649–666

  57. [65]

    Tracking persons-of-interest via adaptive discriminative features, in: European conference on computer vision, Springer

    Zhang, S., Gong, Y., Huang, J.B., Lim, J., Wang, J., Ahuja, N., Yang, M.H., 2016b. Tracking persons-of-interest via adaptive discriminative features, in: European conference on computer vision, Springer. pp. 415–433. Sense and Learn: Self-Supervision for Omnipresent Sensors 29...

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.