Date of Award
Doctor of Philosophy (PhD)
Frederic D. McKenzie
Alan T. Pope
This dissertation studies methods that improve engagement assessment for pilots. The major work addresses two challenging problems involved in the assessment: individual variation among pilots and the lack of labeled data for training assessment models.
Task engagement is usually assessed by analyzing physiological measurements collected from subjects who are performing a task. However, physiological measurements such as Electroencephalography (EEG) vary from subject to subject. An assessment model trained for one subject may not be applicable to other subjects. We proposed a dynamic classifier selection algorithm for model individualization and compared it to other two methods: base line normalization and similarity-based model replacement. Experimental results showed that baseline normalization and dynamic classifier selection can significantly improve cross-subject engagement assessment.
For complex tasks such as piloting an air plane, labeling engagement levels for pilots is challenging. Without enough labeled data, it is very difficult for traditional methods to train valid models for effective engagement assessment. This dissertation proposed to utilize deep learning models to address this challenge. Deep learning models are capable of learning valuable feature hierarchies by taking advantage of both labeled and unlabeled data. Our results showed that deep models are better tools for engagement assessment when label information is scarce.
To further verify the power of deep learning techniques for scarce labeled data, we applied the deep learning algorithm to another small size data set, the ADNI data set. The ADNI data set is a public data set containing MRI and PET scans of Alzheimer's Disease (AD) patients for AD diagnosis. We developed a robust deep learning system incorporating dropout and stability selection techniques to identify the different progression stages of AD patients. The experimental results showed that deep learning is very effective in AD diagnosis.
In addition, we studied several imbalance learning techniques that are useful when data is highly unbalanced, i.e., when majority classes have many more training samples than minority classes. Conventional machine learning techniques usually tend to classify all data samples into majority classes and to perform poorly for minority classes. Unbalanced learning techniques can balance data sets before training and can improve learning performance.
"Improving Engagement Assessment by Model Individualization and Deep Learning"
(2015). Doctor of Philosophy (PhD), Dissertation, Electrical/Computer Engineering, Old Dominion University, DOI: 10.25777/1yt3-3873