The DOI matches its Crossref record; 26 of the 31 references it cites were found in a registry or an index, and 5 are not checked yet; 2 of the 10 sources its claims cite were checked for support.
How Much Unlabeled Data is Really Needed for Effective Self-Supervised Human Activity Recognition?
Open accessAbstract
The prospect of learning effective representations from unlabeled data alone has led to a boost in developing self-supervised learning (SSL) methods for sensor-based Human Activity Recognition (HAR). Typically, (large-scale) unlabeled data are used for pre-training, with the learned weights being used as feature extractors for recognizing activities. While prior works have focused on the impact of increased data scale on performance, instead, we aim to discover the pre-training data efficiency of self-supervised methods. We empirically determine the minimal quantities of unlabeled data required for obtaining comparable performance to using all available data. We investigate three established SSL methods for HAR on three target datasets. Out of these three methods, we discover that Contrastive Predictive Coding (CPC) is the most efficient in terms of pre-training data requirements: just 15 minutes of sensor data across participants is sufficient to obtain competitive activity recognition performance. Further, around 5 minutes of source data is enough when there are sufficient amounts of target application data available. These findings can serve as starting point for more efficient data collection practices.
Verified 2026-10-05 · 5 references not checked yet
Overview
CoChat Summary
TL;DR
This paper empirically determines the “Critical Mass” of unlabeled pre-training data needed for self-supervised human activity recognition, defined as the minimum data yielding an F1 score within 5% of using the full training split. Across three SSL methods and three wrist-based datasets, Contrastive Predictive Coding (CPC) is the most data-efficient, needing roughly 15 minutes for Bottom-Up evaluation and 5 minutes for Transferred evaluation.
Key findings
- CPC had the lowest Critical Mass across target datasets: approximately 9, 11, and 15 minutes of data for HHAR, PAMAP2, and RealWorld, respectively, gave comparable performance within 5% F1 to using the entire training data.
- In Transferred evaluation, external data amounts of a maximum of 5 minutes were sufficient for CPC to achieve comparable performance on target datasets.
- Autoencoder and SimCLR Critical Masses were “relatively much higher” than CPC; the authors attribute this to Autoencoders relying on reconstructing input windows and SimCLR relying on large batch sizes for useful negatives.
- CPC near its Critical Mass performed highest relative to both self-supervised and supervised baselines, and was comparable to supervised methods with larger labeled data; Supervised-AE outperformed pre-training with CPC overall, but CPC was more data-efficient especially on HHAR and PAMAP2.
Methods
- The study evaluated three established SSL methods for sensor-based HAR: Autoencoders, Contrastive Predictive Coding (CPC), and SimCLR, on three public wearable datasets: HHAR, PAMAP2, and RealWorld. Only wrist and forearm IMU data were used; all sensor readings were downsampled to 50 Hz.
- Critical Mass was calculated by pre-training on the full training split to obtain maximum macro F1 score P, then progressively reducing pre-training data by a factor of 0.5 until F1 dropped below P − 5%; the data amount at that point was defined as Critical Mass. Two settings were used: Bottom-Up evaluation, pre-training and fine-tuning on target data, and Transferred evaluation, pre-training on external source data and fine-tuning on target data.
- Participant-based splits were used: 20% of subjects as test, 20% of remaining as validation, rest for training, with 5-fold cross-validation. Sliding windows were fixed-size 50 with overlap 25 timesteps. Pre-training ran for 75 epochs, then weights were frozen and a single fully-connected layer was trained for 150 epochs. Results report mean test macro F1 across 5 folds and 5 random runs.
Limitations
- The paper states that the specific 5% tolerance may be changed for other scenarios, that precise Critical Mass values were interpolated based on the selected threshold, and that sliding-window overlap can affect final Critical Mass values; they assume the worst case of no overlap, so actual Critical Masses may be even lower than reported.
- The paper states that no hyperparameter tuning was performed, as they aimed to replicate real-world scenarios.
- Evident from the text: the study covers only three SSL methods, three datasets, wrist/forearm IMU sensing, and mostly locomotion-style activities; Transferred evaluation results are shown only for CPC.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
2 Valid · 0 Not valid · 3 Not checked · 21 Own results and interpretations
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“Therefore, self-supervision clearly has the capability of being data efficient, i.e., pre-training only on small amounts of data, with similar downstream performance.”
§ How Much Unlabeled Data is Really Needed for Effective Self-Supervised Human Activity Recognition? > 2 BACKGROUND
What the sources say
- Not checked yetcites [2]
- Not checked yetcites [11]
- Supportedcites [19] source
Section 5.2.1 ('Quantity of the Unlabeled Data Required') frames its purpose as evaluating 'the data efficiency of the self-supervised methods' and reports that 'a majority of the self-supervised methods including Multi-task self-supervision, CPC, Autoencoder, SimCLR, SimSiam, and BYOL do not require the entire training set (135 participants) for best performance, and can be trained for increased efficiency with fewer participants without much loss in performance', and that 'excellent performance can be obtained while using 10% of the training windows, and in some cases using as few as 1%'. The Fig. 3 caption states that 'utilizing fewer users results in comparable if not better performance to using the entirety of the training set'. The claim's hedged wording ('has the capability of') matches the source's scope, though the source notes exceptions (masked reconstruction / large transformer encoders need more data, and one subject alone sometimes gives poorer performance). (LLM, conf=0.88)
boostedstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Specifically, access to as little as approx.9 mins, 11 mins, and 15 mins of data, respectively, result in comparable performance (within 5% F1 score) to utilizing the entire available training data.”
§ How Much Unlabeled Data is Really Needed for Effective Self-Supervised Human Activity Recognition? > 4.1 Bottom-Up Evaluation
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Specifically, access to as little as approx.9 mins, 11 mins, and 15 mins of data, respectively, result in comparable performance (within 5% F1 score) to utilizing the entire available training data.”
§ How Much Unlabeled Data is Really Needed for Effective Self-Supervised Human Activity Recognition? > 4.1 Bottom-Up Evaluation
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Comparing the performances of self-supervised methods with their supervised counterparts reveals a crucial observation: the performance of CPC near its critical mass is the highest relative to both self-supervised and supervised baselines.”
§ How Much Unlabeled Data is Really Needed for Effective Self-Supervised Human Activity Recognition? > 4.1 Bottom-Up Evaluation
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“However, we also note that CPC is typically more data efficient than Supervised-AE, especially with datasets like HHAR and PAMAP2.”
§ How Much Unlabeled Data is Really Needed for Effective Self-Supervised Human Activity Recognition? > 4.1 Bottom-Up Evaluation
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
12 more claims
Create a free account to see all 26 claims and to verify the 8 references still open.
- Probably not a substantial claim. 6
- It may state two separate claims as one. 6
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The cited source says what the claim attributes to it.
In the citing paper
“The pre-training data efficiency of selfsupervision was evaluated in [43], which found that even a few minutes of unlabeled data with sufficient augmentation can rival using entire dataset for pre-training.”
From Past, Present, and Future of Sensor-based Human Activity Recognition Using Wearables: A Surveying Tutorial on a Still Challenging Task (2025) · cites this paper as [43]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ How Much Unlabeled Data is Really Needed for Effective Self-Supervised Human Activity Recognition? > 3.1 Critical Mass
“We define Critical Mass as the minimum amount of unlabeled data used for pre-training, which results in a recognition F1 score within 5% of using the entire dataset.”
§ How Much Unlabeled Data is Really Needed for Effective Self-Supervised Human Activity Recognition? > 3.1 Critical Mass
“Then, we progressively reduce the quantity of pre-training data by a factor of 0.5, until the resulting F1-score 𝑃 ′ drops below 𝑃 -5%.The amount of data that led to 𝑃 ′ is then defined as 'Critical Mass'.”
Citations
Cites
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
Momentum Contrast for Unsupervised Visual Representation Learning
He, K., Fan, H., Wu, Y., Xie, S., & Girshick, R. (2020) · 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr42600.2020.00975 · Cited by 13,347
Representation Learning with Contrastive Predictive Coding
Oord, A. V. D., Li, Y., & Vinyals, O. (2018) · arXiv · doi:10.48550/arxiv.1807.03748 · Cited by 6,242
Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning
Grill, J.-B., Strub, F., Altché, F., Tallec, C., Richemond, P. H., Buchatskaya, E., Doersch, C., Pires, B. A., Guo, Z. D., Azar, M. G., et al. (2020) · arXiv · doi:10.48550/arxiv.2006.07733 · Cited by 3,846
Exploring Simple Siamese Representation Learning
Chen, X., & He, K. (2021) · 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr46437.2021.01549 · Cited by 3,725
Introducing a New Benchmarked Dataset for Activity Monitoring
Reiss, A., & Stricker, D. (2012) · 2012 16th International Symposium on Wearable Computers · doi:10.1109/iswc.2012.13 · Cited by 1,456 · cited 5×
“In case of PAMAP2 and RealWorld datasets, data is collected through sensors at multiple on-body locations (ankle, chest, arm, etc.), yet we only use the IMU data collected at the wrist and the forearm, respectively.”
InfluentialUm, T. T., Pfister, F. M. J., Pichler, D., Endo, S., Lang, M., Hirche, S., Fietzek, U., & Kulić, D. (2017) · Proceedings of the 19th ACM International Conference on Multimodal Interaction · doi:10.1145/3136755.3136817 · Cited by 729
Stisen, A., Blunck, H., Bhattacharya, S., Prentow, T. S., Kjærgaard, M. B., Dey, A., Sonne, T., & Jensen, M. M. (2015) · Proceedings of the 13th ACM Conference on Embedded Networked Sensor Systems · doi:10.1145/2809695.2809718 · Cited by 714
Revisiting Self-Supervised Visual Representation Learning
Kolesnikov, A., Zhai, X., & Beyer, L. (2019) · 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2019.00202 · Cited by 675
Um, T. T. (n.d.) · mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) · Cited by 602
Research works printed but not linked: 9
- Nicolas Carion , Francisco Massa , Gabriel Synnaeve , Nicolas Usunier , Alexander Kirillov , and Sergey Zagoruyko . 2020 . End-to-end object detection with transformers. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020 , Proceedings, Part I 16 . Springer, 213–229. Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. 2020. End-to-end object detection with transformers. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16. Springer, 213–229.
- Ting Chen , Simon Kornblith , Mohammad Norouzi , and Geoffrey Hinton . 2020 . A simple framework for contrastive learning of visual representations . In International conference on machine learning. PMLR, 1597–1607 . Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning. PMLR, 1597–1607.
- Max Eichler , Gözde Gül Şahin , and Iryna Gurevych . 2019 . LINSPECTOR WEB: A multilingual probing suite for word representations. arXiv preprint arXiv:1907.11438 (2019). Max Eichler, Gözde Gül Şahin, and Iryna Gurevych. 2019. LINSPECTOR WEB: A multilingual probing suite for word representations. arXiv preprint arXiv:1907.11438 (2019). · arXiv
- 10.5555/3104322.3104425 · doi:10.5555/3104322.3104425
- Nitish Srivastava , Geoffrey Hinton , Alex Krizhevsky , Ilya Sutskever , and Ruslan Salakhutdinov . 2014. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15, 1 ( 2014 ), 1929–1958. Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15, 1 (2014), 1929–1958.
- Chi Ian Tang , Ignacio Perez-Pozuelo , Dimitris Spathis , Soren Brage , Nick Wareham , and Cecilia Mascolo . 2021. SelfHAR: Improving Human Activity Recognition through Self-training with Unlabeled Data. arXiv preprint arXiv:2102.06073 ( 2021 ). Chi Ian Tang, Ignacio Perez-Pozuelo, Dimitris Spathis, Soren Brage, Nick Wareham, and Cecilia Mascolo. 2021. SelfHAR: Improving Human Activity Recognition through Self-training with Unlabeled Data. arXiv preprint arXiv:2102.06073 (2021). · arXiv
- Aaron Van den Oord , Yazhe Li , and Oriol Vinyals . 2018. Representation learning with contrastive predictive coding. arXiv e-prints ( 2018 ), arXiv–1807. Aaron Van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv e-prints (2018), arXiv–1807.
- Ivan Vulić , Edoardo Maria Ponti , Robert Litschko , Goran Glavaš , and Anna Korhonen . 2020. Probing pretrained language models for lexical semantics. arXiv preprint arXiv:2010.05731 ( 2020 ). Ivan Vulić, Edoardo Maria Ponti, Robert Litschko, Goran Glavaš, and Anna Korhonen. 2020. Probing pretrained language models for lexical semantics. arXiv preprint arXiv:2010.05731 (2020). · arXiv
- Jure Zbontar , Li Jing , Ishan Misra , Yann LeCun , and Stéphane Deny . 2021 . Barlow twins: Self-supervised learning via redundancy reduction . In International Conference on Machine Learning. PMLR, 12310–12320 . Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. 2021. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning. PMLR, 12310–12320.
Cited by
Transfer Learning in Sensor-Based Human Activity Recognition: A Survey
Dhekane, S. G., & Ploetz, T. (2025) · ACM Computing Surveys · doi:10.1145/3717608 · Cited by 42
Self-citationHaresamudram, H., Tang, C. I., Suh, S., Lukowicz, P., & Plötz, T. (2025) · Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies · doi:10.1145/3729467 · Cited by 33
“The pre-training data efficiency of self-supervision was evaluated in [41], which found that even a few minutes of unlabeled data with sufficient augmentation can rival using entire dataset for pre-training.”
Self-citationEvaluating Large Language Models as Virtual Annotators for Time-Series Physical Sensing Data
Hota, A., Chatterjee, S., & Chakraborty, S. (2024) · ACM Transactions on Intelligent Systems and Technology · doi:10.1145/3696461 · Cited by 39
Haresamudram, H., Essa, I., & Plötz, T. (2024) · Sensors · doi:10.3390/s24041238 · Cited by 14
Self-citationProgress and Thinking on Self-Supervised Learning Methods in Computer Vision: A Review
Chen, Z., Hu, B., Chen, Z., & Zhang, J. (2024) · IEEE Sensors Journal · doi:10.1109/jsen.2024.3443885 · Cited by 11
Metadata
- Authors
- Dhekane, S. G., Haresamudram, H., Thukral, M., & Plötz, T.
- Affiliations
- Georgia Institute of Technology (US)
- Venue
- Proceedings of the 2023 International Symposium on Wearable Computers
- Publisher
- ACM
- Published
- 2023-10-08
- Pages
- 66-70
- OpenAlex
- W4387321389
- Cited by
- 9
- References
- 21
- Citations by year
- 2026: 1 · 2025: 4 · 2024: 4
- FWCI
- 0.96
- Type
- conference paper
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- Locations
- International Symposium on Wearable Computers (publishedVersion, cc-by)
- Funders
- National Science Foundation (IIS-2112633)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Physical Sciences › Computer Science › Signal Processing › Time Series Analysis and Forecasting
- Keywords
- self-supervised learning, human activity recognition, unlabeled data, contrastive predictive coding, sensor data, data collection practices, feature extraction
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: How Much Unlabeled Data is Really Needed for Effective Self-Supervised Human Activity Recognition? https://app.cochat.ai/papers/10.1145/3594738.3611366
- The paper
- Dhekane, S. G., Haresamudram, H., Thukral, M., & Plötz, T. (2023). How Much Unlabeled Data is Really Needed for Effective Self-Supervised Human Activity Recognition? Proceedings of the 2023 International Symposium on Wearable Computers. https://doi.org/10.1145/3594738.3611366
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.