Oral Candidacy - Robert Bixler
|Start:||11/17/2016 at 3:00PM|
|End:||11/17/2016 at 5:00PM|
|Location:||315 Stinson Remick|
Faculty and students are welcome to attend the presentation portion of the defense.
November 17, 2016 3:00 pm 315 Stinson Remick
Adviser: Dr. Sidney D’Mello
Dr. Adam Czjaka Dr. Ron Metoyer Dr. Dong Wang
Detecting Mind Wandering across Domains using Temporal Models of Eye Gaze
Mind wandering is a ubiquitous phenomenon where attention shifts from task-related thoughts to internal task-unrelated thoughts. It is negatively correlated with performance on tasks requiring conscious control such as reading, viewing dynamic scenes such as a film, and learning. One solution to this problem is to develop systems capable of responding when mind wandering occurs by reorienting attention to the task at hand. However, this requires a reliable, scalable, and unobtrusive method to automatically detect mind wandering. This proposal provides an overview of completed work related to the development of several automatic gaze based detectors of mind wandering, and a proposed extension to the completed work.
Completed work demonstrates the efficacy of using domain-independent global gaze features and domain-specific local gaze features for training supervised machine learning classifiers in order to detect mind wandering during reading and dynamic scene viewing. Gaze data and self-reports of mind wandering were collected across three separate studies while participants completed either computerized reading or dynamic scene viewing tasks. The different tasks and thought sampling methods used to obtain mind wandering self-reports posed unique challenges for building mind wandering detectors and solutions were developed to address them. Supervised machine learning classifiers were trained to detect mind wandering using features computed from gaze data in a participant-independent fashion.
Proposed work will extend the completed work by using recurrent neural networks to model temporal (domain-independent) gaze patterns and viewing content (domain-specific) in order to detect mind wandering. Current gaze based mind wandering detectors have not taken temporal gaze patterns into account, yet these patterns could aid in the detection of mind wandering. Using recurrent neural networks to model these temporal patterns is motivated by the desire to improve upon the accuracy of current detectors and to facilitate cross-domain detection. Novel methods to encode domain-independent gaze data and domain-specific viewing content will be developed. Recurrent neural networks built using each kind of data will be evaluated and compared to current state-of-the-art gaze based mind wandering detectors. The aim of the proposed work is to develop a cross-domain mind wandering detector with accuracy better than or comparable to current detectors.