Date
6-2010
Document Type
Dissertation
Degree Name
Ph.D.
Department
Department of Science & Engineering
Institution
Oregon Health & Science University
Abstract
Brain computer interfaces (BCIs) provide a non-muscular avenue for the user to communicatewith others and to control external devices. Over the last two decades BCIs have been developed to assist the severely motor-disabled people, such as traumatic braininjury, stroke, or amyotrophic lateral sclerosis. Electroencephalography (EEG) is one of the most popular noninvasive BCI approaches. The inputs to EEG-based BCIs are event-related potentials (ERPs), which are neural signatures representing the responses to an external stimulus. Traditional BCI systems, which have had some success, make inferences based on trial-averaged ERPs, where each trial consists of one stimulus. In this thesis, (1) we develop a single-trial, EEG-based BCI to increase the throughput of visual image search and (2) we unveil a neural correlate of human visual perception that occurs in rapid visual-recognition tasks. Our first task is to develop a BCI. Our BCI makes inferences from single-trial ERPs; hence, it is more efficient than traditional methods. It uses cross-session training and a novel, hybrid generative/discriminative classifier (which combines a mixed effect model and a support vector machine via a Fisher kernel) to improve ERP detection performance, and it uses dimension reduction and incremental learning to reduce computational complexity. Based on the analysis of our BCI, we conclude that: single-trial ERP detection is possible; cross-session training outperforms the often-used single-session method; our hybrid classifier has a detection performance that is as good or better than some of the well-known classifiers; and dimension reduction and incremental learning substantially reduces computational complexity and they do so without an associated drop in detection performance. Our second task is to characterize a neural correlate of human visual perception. Our approach involves measuring physiological signals and behavioral performance as a function of both the difficulty of the task (measured by the length of time images are available for viewing) and the difficulty of the target (estimated by the minimum viewing time required for a fixed detection rate). We find that the neural responses are highly correlated with both target difficulty and task difficulty. Based on these findings we further surmise that, during visual information processing, the brain dynamically allocates additional cognitive resources under increasingly difficult conditions.
Identifier
doi:10.6083/M48P5XGJ
Division
Division of Biomedical Engineering
School
School of Medicine
Recommended Citation
Huang, Yonghong, "Event-related potentials in electroencephalography characteristics and single-trial detection for rapid object search" (2010). Scholar Archive. 377.
https://digitalcommons.ohsu.edu/etd/377