Conference Proceeding Article
We investigate the problem of making human activityrecognition (AR) scalable–i.e., allowing AR classifiers trainedin one context to be readily adapted to a different contextualdomain. This is important because AR technologies can achievehigh accuracy if the classifiers are trained for a specific individualor device, but show significant degradation when the sameclassifier is applied context–e.g., to a different device located ata different on-body position. To allow such adaptation withoutrequiring the onerous step of collecting large volumes of labeledtraining data in the target domain, we proposed a transductivetransfer learning model that is specifically tuned to the propertiesof convolutional neural networks (CNNs). Our model, calledHDCNN, assumes that the relative distribution of weights in thedifferent CNN layers will remain invariant, as long as the set ofactivities being monitored does not change. Evaluation on realworlddata shows that HDCNN is able to achieve high accuracyeven without any labeled training data in the target domain,and offers even higher accuracy (significantly outperformingcompetitive shallow and deep classifiers) when even a modestamount of labeled training data is available.
Data Storage Systems | OS and Networks | Programming Languages and Compilers
Data Science and Engineering
2018 IEEE International Conference on Pervasive Computing and Communications, Athens, Greece, March 19-23
City or Country
KHAN, Md Abdullah Hafiz; ROY, Nirmalya; and MISRA, Archan.
Scaling human activity recognition via deep learning-based domain adaptation. (2018). 2018 IEEE International Conference on Pervasive Computing and Communications, Athens, Greece, March 19-23. 1-9. Research Collection School Of Information Systems.
Available at: http://ink.library.smu.edu.sg/sis_research/3977
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.