Active Activity Recognition with Context-Aware Annotator Selection
Loading...
Links to Files
Permanent Link
Author/Creator
Author/Creator ORCID
Date
2019-01-01
Type of Work
Department
Information Systems
Program
Information Systems
Citation of Original Publication
Rights
Distribution Rights granted to UMBC by the author.
Access limited to the UMBC community. Item may possibly be obtained via Interlibrary Loan thorugh a local library, pending author/copyright holder's permission.
This item is likely protected under Title 17 of the U.S. Copyright Law. Unless on a Creative Commons license, for uses protected by Copyright Law, contact the copyright holder or the author.
Access limited to the UMBC community. Item may possibly be obtained via Interlibrary Loan thorugh a local library, pending author/copyright holder's permission.
This item is likely protected under Title 17 of the U.S. Copyright Law. Unless on a Creative Commons license, for uses protected by Copyright Law, contact the copyright holder or the author.
Abstract
The proliferation of wearable and pervasive devices have revolutionized many application domains ranging from health care, sports, entertainment etc. By exploiting the sensor rich wearable and Internet-of-Things (IoT) devices in smart environments, assessment and inference of Activities of Daily Living (ADL) have been a major research proposition over the past decade. A multitude of activity recognition models based on supervised, semi-supervised and unsupervised approaches with significant performance improvements have been posited by the researchers. However, collecting robust label information is fundamental for interpreting the underlying activity data distributions distinctively, and training of the supervised and semi-supervised learning models adequately. Moreover, in practical setting, the limited availability of ground truth information and the variabilities in activities make the activity recognition models impractical for real-world deployment. This ground truth annotation is objectively manual and tedious as it needs considerable amount of human interventions. With the advent of Active Learning with multiple annotators, the burden can be somewhat mitigated by actively acquiring labels of most informative data instances. However, multiple annotators with varying degrees of expertise poses new set of challenges in terms of quality of the label received and availability of the annotator. In this thesis, we investigate how this obligation of collecting ground truth information can be mitigated by acquiring labels of most informative data instances using Active Learning in activity recognition domain. We propose several active learning enabled activity recognition models which help collect activity labels from human annotators online and reduce the training time warranted while achieving reasonably similar accuracy compared to traditional supervised models. We also propose an active learning combined deep model which updates its network parameters based on the optimization of a joint loss function. In addition, it is both difficult and annoying for an user to provide his own activity information continuously while employing active learning. Introducing multiple annotators can alleviate this adversity, but which annotator can provide reliable label is a fundamental research question. We propose to model the interactions between the users and annotators as relationships spanning across spatial and temporal space of activity domain. We introduce a novel approach to quantify the strength of relations using the patial and temporal information of interactions, type of the relationships, and activities. Our proposed model leverages model-free deep reinforcement learning in a partially observable environment setting to capture the action-reward interaction among multiple annotators. Our experiments in real-world settings exhibit that our active deep model converges to optimal accuracy with fewer labeled instances and achieves $\approx 8\%$ improvement in accuracy in fewer iterations.