SelfPAB: large-scale pre-training on accelerometer data for human activity recognition. Logacjov, A., Herland, S., Ustad, A., & Bach, K. Applied Intelligence, 54(6):4545–4563, March, 2024. ![SelfPAB: large-scale pre-training on accelerometer data for human activity recognition [link]](https://bibbase.org/img/filetypes/link.svg "link")
Paper doi abstract bibtex Annotating accelerometer-based physical activity data remains a challenging task, limiting the creation of robust supervised machine learning models due to the scarcity of large, labeled, free-living human activity recognition (HAR) datasets. Researchers are exploring self-supervised learning (SSL) as an alternative to relying solely on labeled data approaches. However, there has been limited exploration of the impact of large-scale, unlabeled datasets for SSL pre-training on downstream HAR performance, particularly utilizing more than one accelerometer. To address this gap, a transformer encoder network is pre-trained on various amounts of unlabeled, dual-accelerometer data from the HUNT4 dataset: 10, 100, 1k, 10k, and 100k hours. The objective is to reconstruct masked segments of signal spectrograms. This pre-trained model, termed SelfPAB, serves as a feature extractor for downstream supervised HAR training across five datasets (HARTH, HAR70+, PAMAP2, Opportunity, and RealWorld). SelfPAB outperforms purely supervised baselines and other SSL methods, demonstrating notable enhancements, especially for activities with limited training data. Results show that more pre-training data improves downstream HAR performance, with the 100k-hour model exhibiting the highest performance. It surpasses purely supervised baselines by absolute F1-score improvements of 7.1% (HARTH), 14% (HAR70+), and an average of 11.26% across the PAMAP2, Opportunity, and RealWorld datasets. Compared to related SSL methods, SelfPAB displays absolute F1-score enhancements of 10.4% (HARTH), 18.8% (HAR70+), and 16% (average across PAMAP2, Opportunity, RealWorld).
@article{logacjov_selfpab_2024,
title = {{SelfPAB}: large-scale pre-training on accelerometer data for human activity recognition},
volume = {54},
issn = {1573-7497},
shorttitle = {{SelfPAB}},
url = {https://doi.org/10.1007/s10489-024-05322-3},
doi = {10.1007/s10489-024-05322-3},
abstract = {Annotating accelerometer-based physical activity data remains a challenging task, limiting the creation of robust supervised machine learning models due to the scarcity of large, labeled, free-living human activity recognition (HAR) datasets. Researchers are exploring self-supervised learning (SSL) as an alternative to relying solely on labeled data approaches. However, there has been limited exploration of the impact of large-scale, unlabeled datasets for SSL pre-training on downstream HAR performance, particularly utilizing more than one accelerometer. To address this gap, a transformer encoder network is pre-trained on various amounts of unlabeled, dual-accelerometer data from the HUNT4 dataset: 10, 100, 1k, 10k, and 100k hours. The objective is to reconstruct masked segments of signal spectrograms. This pre-trained model, termed SelfPAB, serves as a feature extractor for downstream supervised HAR training across five datasets (HARTH, HAR70+, PAMAP2, Opportunity, and RealWorld). SelfPAB outperforms purely supervised baselines and other SSL methods, demonstrating notable enhancements, especially for activities with limited training data. Results show that more pre-training data improves downstream HAR performance, with the 100k-hour model exhibiting the highest performance. It surpasses purely supervised baselines by absolute F1-score improvements of 7.1\% (HARTH), 14\% (HAR70+), and an average of 11.26\% across the PAMAP2, Opportunity, and RealWorld datasets. Compared to related SSL methods, SelfPAB displays absolute F1-score enhancements of 10.4\% (HARTH), 18.8\% (HAR70+), and 16\% (average across PAMAP2, Opportunity, RealWorld).},
language = {en},
number = {6},
urldate = {2024-09-24},
journal = {Applied Intelligence},
author = {Logacjov, Aleksej and Herland, Sverre and Ustad, Astrid and Bach, Kerstin},
month = mar,
year = {2024},
keywords = {Accelerometer, Artificial Intelligence, Human activity recognition, Machine learning, Physical activity behavior, Self-supervised learning, Transformer},
pages = {4545--4563},
}
Downloads: 0
{"_id":"9geukZsc6ckH8M9KR","bibbaseid":"logacjov-herland-ustad-bach-selfpablargescalepretrainingonaccelerometerdataforhumanactivityrecognition-2024","author_short":["Logacjov, A.","Herland, S.","Ustad, A.","Bach, K."],"bibdata":{"bibtype":"article","type":"article","title":"SelfPAB: large-scale pre-training on accelerometer data for human activity recognition","volume":"54","issn":"1573-7497","shorttitle":"SelfPAB","url":"https://doi.org/10.1007/s10489-024-05322-3","doi":"10.1007/s10489-024-05322-3","abstract":"Annotating accelerometer-based physical activity data remains a challenging task, limiting the creation of robust supervised machine learning models due to the scarcity of large, labeled, free-living human activity recognition (HAR) datasets. Researchers are exploring self-supervised learning (SSL) as an alternative to relying solely on labeled data approaches. However, there has been limited exploration of the impact of large-scale, unlabeled datasets for SSL pre-training on downstream HAR performance, particularly utilizing more than one accelerometer. To address this gap, a transformer encoder network is pre-trained on various amounts of unlabeled, dual-accelerometer data from the HUNT4 dataset: 10, 100, 1k, 10k, and 100k hours. The objective is to reconstruct masked segments of signal spectrograms. This pre-trained model, termed SelfPAB, serves as a feature extractor for downstream supervised HAR training across five datasets (HARTH, HAR70+, PAMAP2, Opportunity, and RealWorld). SelfPAB outperforms purely supervised baselines and other SSL methods, demonstrating notable enhancements, especially for activities with limited training data. Results show that more pre-training data improves downstream HAR performance, with the 100k-hour model exhibiting the highest performance. It surpasses purely supervised baselines by absolute F1-score improvements of 7.1% (HARTH), 14% (HAR70+), and an average of 11.26% across the PAMAP2, Opportunity, and RealWorld datasets. Compared to related SSL methods, SelfPAB displays absolute F1-score enhancements of 10.4% (HARTH), 18.8% (HAR70+), and 16% (average across PAMAP2, Opportunity, RealWorld).","language":"en","number":"6","urldate":"2024-09-24","journal":"Applied Intelligence","author":[{"propositions":[],"lastnames":["Logacjov"],"firstnames":["Aleksej"],"suffixes":[]},{"propositions":[],"lastnames":["Herland"],"firstnames":["Sverre"],"suffixes":[]},{"propositions":[],"lastnames":["Ustad"],"firstnames":["Astrid"],"suffixes":[]},{"propositions":[],"lastnames":["Bach"],"firstnames":["Kerstin"],"suffixes":[]}],"month":"March","year":"2024","keywords":"Accelerometer, Artificial Intelligence, Human activity recognition, Machine learning, Physical activity behavior, Self-supervised learning, Transformer","pages":"4545–4563","bibtex":"@article{logacjov_selfpab_2024,\n\ttitle = {{SelfPAB}: large-scale pre-training on accelerometer data for human activity recognition},\n\tvolume = {54},\n\tissn = {1573-7497},\n\tshorttitle = {{SelfPAB}},\n\turl = {https://doi.org/10.1007/s10489-024-05322-3},\n\tdoi = {10.1007/s10489-024-05322-3},\n\tabstract = {Annotating accelerometer-based physical activity data remains a challenging task, limiting the creation of robust supervised machine learning models due to the scarcity of large, labeled, free-living human activity recognition (HAR) datasets. Researchers are exploring self-supervised learning (SSL) as an alternative to relying solely on labeled data approaches. However, there has been limited exploration of the impact of large-scale, unlabeled datasets for SSL pre-training on downstream HAR performance, particularly utilizing more than one accelerometer. To address this gap, a transformer encoder network is pre-trained on various amounts of unlabeled, dual-accelerometer data from the HUNT4 dataset: 10, 100, 1k, 10k, and 100k hours. The objective is to reconstruct masked segments of signal spectrograms. This pre-trained model, termed SelfPAB, serves as a feature extractor for downstream supervised HAR training across five datasets (HARTH, HAR70+, PAMAP2, Opportunity, and RealWorld). SelfPAB outperforms purely supervised baselines and other SSL methods, demonstrating notable enhancements, especially for activities with limited training data. Results show that more pre-training data improves downstream HAR performance, with the 100k-hour model exhibiting the highest performance. It surpasses purely supervised baselines by absolute F1-score improvements of 7.1\\% (HARTH), 14\\% (HAR70+), and an average of 11.26\\% across the PAMAP2, Opportunity, and RealWorld datasets. Compared to related SSL methods, SelfPAB displays absolute F1-score enhancements of 10.4\\% (HARTH), 18.8\\% (HAR70+), and 16\\% (average across PAMAP2, Opportunity, RealWorld).},\n\tlanguage = {en},\n\tnumber = {6},\n\turldate = {2024-09-24},\n\tjournal = {Applied Intelligence},\n\tauthor = {Logacjov, Aleksej and Herland, Sverre and Ustad, Astrid and Bach, Kerstin},\n\tmonth = mar,\n\tyear = {2024},\n\tkeywords = {Accelerometer, Artificial Intelligence, Human activity recognition, Machine learning, Physical activity behavior, Self-supervised learning, Transformer},\n\tpages = {4545--4563},\n}\n\n\n\n\n\n\n\n","author_short":["Logacjov, A.","Herland, S.","Ustad, A.","Bach, K."],"key":"logacjov_selfpab_2024","id":"logacjov_selfpab_2024","bibbaseid":"logacjov-herland-ustad-bach-selfpablargescalepretrainingonaccelerometerdataforhumanactivityrecognition-2024","role":"author","urls":{"Paper":"https://doi.org/10.1007/s10489-024-05322-3"},"keyword":["Accelerometer","Artificial Intelligence","Human activity recognition","Machine learning","Physical activity behavior","Self-supervised learning","Transformer"],"metadata":{"authorlinks":{}},"html":""},"bibtype":"article","biburl":"https://bibbase.org/zotero/warren.pettine","dataSources":["nqfZGC6dcujCxizZq","SzG9GQMomjwWhJo7s","zYjFbyFmZWKpRCD4j"],"keywords":["accelerometer","artificial intelligence","human activity recognition","machine learning","physical activity behavior","self-supervised learning","transformer"],"search_terms":["selfpab","large","scale","pre","training","accelerometer","data","human","activity","recognition","logacjov","herland","ustad","bach"],"title":"SelfPAB: large-scale pre-training on accelerometer data for human activity recognition","year":2024}