Off-campus UMass Amherst users: To download campus access dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.

Non-UMass Amherst users: Please talk to your librarian about requesting this dissertation through interlibrary loan.

Dissertations that have an embargo placed on them will not be available to anyone until the embargo expires.

Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy (PhD)

Degree Program

Computer Science

Year Degree Awarded

2018

Month Degree Awarded

September

First Advisor

Benjamin Marlin

Subject Categories

Artificial Intelligence and Robotics

Abstract

Wearable wireless sensors have the potential for transformative impact on the fields of health and behavioral science. Recent advances in wearable sensor technology have made it possible to simultaneously collect multiple streams of physiological and context data from individuals in natural environments; however, extracting reliable high-level inferences from these raw data streams remains a key data analysis challenge. In this dissertation, we address three challenges that arise when trying to perform activity detection from wearable sensor streams. First, we address the challenge of learning from small amounts of noisy data by proposing a class of conditional random field models for activity detection. We apply this model class to three different activity detection problems, improving performance in all three when compared with standard independent and structured models. Second, we address the challenge of inferring activities from long input sequences by evaluating strategies for pruning the inference dynamic programs used in structured prediction models. We apply these strategies to the proposed structured activity detection models resulting in inference speedups ranging from 66x to 257x with little to no decrease in predictive performance. Finally, we address the challenge of learning from imprecise annotations by proposing a weak supervision framework for learning discrete-time detection models from imprecise continuous-time observations. We apply this framework to both independent and structured models and demonstrate improved performance over weak supervision baselines.

Share

COinS