kth.sePublications
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Confidence-Calibrated Human Activity Recognition
KTH, School of Electrical Engineering and Computer Science (EECS), Computer Science, Software and Computer systems, SCS. (Gale)ORCID iD: 0000-0001-6780-7755
KTH, School of Electrical Engineering and Computer Science (EECS), Computer Science, Software and Computer systems, SCS. (Gale)ORCID iD: 0000-0003-4516-7317
KTH, School of Electrical Engineering and Computer Science (EECS), Computer Science, Software and Computer systems, SCS. (Gale)ORCID iD: 0000-0002-1899-917X
2021 (English)In: Sensors, E-ISSN 1424-8220, Vol. 21, no 19, p. 6566-6566Article in journal (Refereed) Published
Abstract [en]

Wearable sensors are widely used in activity recognition (AR) tasks with broad applicability in health and well-being, sports, geriatric care, etc. Deep learning (DL) has been at the forefront of progress in activity classification with wearable sensors. However, most state-of-the-art DL models used for AR are trained to discriminate different activity classes at high accuracy, not considering the confidence calibration of predictive output of those models. This results in probabilistic estimates that might not capture the true likelihood and is thus unreliable. In practice, it tends to produce overconfident estimates. In this paper, the problem is addressed by proposing deep time ensembles, a novel ensembling method capable of producing calibrated confidence estimates from neural network architectures. In particular, the method trains an ensemble of network models with temporal sequences extracted by varying the window size over the input time series and averaging the predictive output. The method is evaluated on four different benchmark HAR datasets and three different neural network architectures. Across all the datasets and architectures, our method shows an improvement in calibration by reducing the expected calibration error (ECE)by at least 40%, thereby providing superior likelihood estimates. In addition to providing reliable predictions our method also outperforms the state-of-the-art classification results in the WISDM, UCI HAR, and PAMAP2 datasets and performs as good as the state-of-the-art in the Skoda dataset. 

Place, publisher, year, edition, pages
Basel: MDPI, 2021. Vol. 21, no 19, p. 6566-6566
Keywords [en]
Electrical and Electronic Engineering, Biochemistry, Instrumentation, Atomic and Molecular Physics, and Optics, Analytical Chemistry
National Category
Computer Sciences
Research subject
Computer Science
Identifiers
URN: urn:nbn:se:kth:diva-303329DOI: 10.3390/s21196566ISI: 000778245800002PubMedID: 34640886Scopus ID: 2-s2.0-85116672506OAI: oai:DiVA.org:kth-303329DiVA, id: diva2:1602304
Note

QC 20220422

Available from: 2021-10-12 Created: 2021-10-12 Last updated: 2024-02-07Bibliographically approved
In thesis
1. Towards Trustworthy Machine Learning For Human Activity Recognition
Open this publication in new window or tab >>Towards Trustworthy Machine Learning For Human Activity Recognition
2024 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

Human Activity Recognition presents a multifaceted challenge, encompassing the complexity of human activities, the diversity of sensors used, and the imperative to safeguard user data privacy. Recent advancements in machine learning, deep learning, and sensor technology have opened up new possibilities for human activity recognition. Wearable sensor-based human activity recognition involves collecting time-series data from various sensors, capturing intricate aspects of human activities. The focus of the above activity recognition problem is classifying human activities from the time-series data. Hence, this time-series classification problem demands efficient utilization of temporal properties. Moreover, while accurate prediction is crucial in human activity recognition, the reliability of predictions often goes unnoticed. Ensuring that predictions are reliable involves addressing two issues: calibrating miscalibrated predictions that fail to accurately represent the true likelihood of the data and addressing the challenges around uncertain predictions. Modern deep learning models, used extensively in human activity recognition, often struggle with the above issues. In addition to reliability concerns, machine learning algorithms employed in Human Activity Recognition are also plagued by privacy issues stemming from the utilization of sensitive activity data during model training. While existing techniques such as federated learning can provide some degree of privacy protection in these scenarios, they tend to adhere to a uniform concept of privacy and lack quantifiable privacy metrics that can be effectively conveyed to users and customized to cater to their individual privacy preferences. Hence, in the thesis, we identify the challenges around the effective use of temporal data, reliability, and privacy issues of machine learning models used for wearable sensor-based human activity recognition. To tackle these challenges, we put forth novel solutions, striving to enhance the overall performance and trustworthiness of machine learning models employed in human activity recognition.

Firstly, to improve classification performance, we propose a new temporal ensembling framework that uses data temporality effectively. The framework accommodates various window sizes for time-series data and trains an ensemble of deep-learning models based on that. It enhances classification accuracy and preserves temporal information.

Secondly, we address reliability through calibration and uncertainty estimation. The aforementioned temporal ensembling framework is used for calibration and uncertainty estimation. It provides well-calibrated predictions for human activity recognition and detects out-of-distribution activities, an important task of uncertainty estimation. Furthermore, we apply these methods to real-world scenarios, enhancing the reliability of human activity recognition models.

Thirdly, to address the privacy concern, we introduce a differentially private framework for time-series human activity recognition, quantifying privacy. Additionally, we develop a collaborative federated learning framework, allowing users to define their privacy preferences, advancing privacy preservation in human activity recognition.

These contributions address major challenges and promote improved classification, reliability, and privacy preservation in human activity recognition. It helps us to move towards trustworthy machine learning in human activity recognition, facilitating their usage in realistic and practical scenarios.

Place, publisher, year, edition, pages
Stockholm: KTH Royal Institute of Technology, 2024. p. xii, 56
Series
TRITA-EECS-AVL ; 2024:12
National Category
Computer Sciences
Research subject
Computer Science; Information and Communication Technology
Identifiers
urn:nbn:se:kth:diva-343130 (URN)978-91-8040-826-4 (ISBN)
Public defence
2024-03-06, https://kth-se.zoom.us/j/63687967257, Sal C, Kistagången 16, Kista, Stockholm, 13:00 (English)
Opponent
Supervisors
Funder
EU, Horizon 2020, 813162
Note

QC 20240207

Available from: 2024-02-07 Created: 2024-02-07 Last updated: 2024-02-29Bibliographically approved

Open Access in DiVA

ConfcalHAR(29461 kB)122 downloads
File information
File name FULLTEXT01.pdfFile size 29461 kBChecksum SHA-512
50f6834f445c7981e6cc96878ef5effc287d0fcbf1229a196996d9e0ee3482be85c8569671a9d8329311b3dc94501fb0ebc79d1ed960e8e61e5295c3614ca7cb
Type fulltextMimetype application/pdf

Other links

Publisher's full textPubMedScopus

Authority records

Roy, DebadityaGirdzijauskas, SarunasSocolovschi, Serghei

Search in DiVA

By author/editor
Roy, DebadityaGirdzijauskas, SarunasSocolovschi, Serghei
By organisation
Software and Computer systems, SCS
In the same journal
Sensors
Computer Sciences

Search outside of DiVA

GoogleGoogle Scholar
Total: 123 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

doi
pubmed
urn-nbn

Altmetric score

doi
pubmed
urn-nbn
Total: 198 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf