Title: Weakly Supervised Deep learning for Human Activity Recognition

Sungtae An

Ph.D. Student in Computer Science

School of Interactive Computing

Georgia Institute of Technology

Date: Thursday, December 9, 2021

Time: 2:00 PM - 4:00 PM EST

Location(Remote via BlueJeans): https://bluejeans.com/190340851/5419

Committee:

Dr. Omer T. Inan (Advisor), School of Electrical and Computer Engineering, Georgia Institute of Technology

Dr. James Rehg (Co-advisor), School of Interactive Computing, Georgia Institute of Technology

Dr. Thomas Ploetz, School of Interactive Computing, Georgia Institute of Technology

Dr. Jon Duke, School of Interactive Computing, Georgia Institute of Technology

Dr. Mindy L. Millard-Stafford, School of Biological Sciences, Georgia Institute of Technology

Dr. Alessio Medda, Aerospace, Transportation & Advanced Systems Laboratory, Georgia Tech Research Institute

Abstract:

Human activity recognition (HAR) using wearable sensors and machine learning algorithms is an emerging capability in domains including but not limited to healthcare and ergometric analysis of populations by providing context to physiological measures from wearable sensors during natural daily living activities.

Despite the success of deep supervised models in recent years, obtaining a fully labeled HAR dataset is often difficult due to the high cost and workforce associated with labeling.

Another critical challenge that has often been underestimated is to learn a robust HAR model with noisy training data.

I address these challenges in this dissertation research work with the following contributions.

First, I present the bilateral domain adaptation problem in HAR for the first time and propose AdaptNet, a semi-supervised deep translation network, which enables information fusion of two different data domains using both unlabeled and labeled data.

Next, I propose a novel framework DynaLAP-VAE, a semi-supervised variational autoencoder with dynamic latent state-space and dynamic prior distribution, which implicitly exploits the information about the environment to enhance HAR in fixed protocols such as military and athletic training with few labeled subject data.

Then, I propose a supervised contrastive learning method with stochastic embeddings to enable the data uncertainty measurement in HAR problems.