Files and links (1)
PDFOpen Access (License Unspecified), Open Access
Dissertation
Classification of insight and analytic problem-solving using an interpretable deep-learning model based on single-trial EEG data
Doctor of Philosophy (Ph.D.), Drexel University
Jun 2022
DOI:
https://doi.org/10.17918/00001107
Abstract
When people solve a given problem, there are two types of processing modes that can be employed: Insight and Analysis. Insight is an interesting cognitive phenomenon in which an individual experiences a sudden realization of the solution. Insight often involves restructuring of the conceptual or perceptual set giving rise to generation of creative ideas. Although research effort on insight using neuroimaging techniques has expanded and numerous exciting findings have been reported, most of the works utilize trial-averaging approach which ultimately reflect the brain dynamics of average brain but less of the brain differences within- and between-individuals. The aim of this project was to build an interpretable Convolutional Neural Network (CNN) model that can classify insight and analytic problem solving at single-trial level. To investigate significant insight-analytic differences at both trial-averaging and single-trial approaches, 89-subject EEG datasets from two studies were processed using fully automated pipeline. Using Statistical Parametric Mapping (SPM) as a means of mass-univariate testing, significant insight-analytic differences were observed at beta-gamma frequency band around 1 second pre-solution, replicating temporal and anatomical location from the previous work. Training a machine learning model using activation from this analysis resulted in a chance-level classification accuracy. Using four time-segment ensemble of EEGNet, a CNN-based architecture capable of extracting spectro-spatial features, a whole-group level test classification accuracy of 66.3% was achieved using combined dataset, with true insight rate of 74.5%. Classification at individual-level only resulted in a modest test accuracy of 51.7%, with true insight rate of 68.8%. Extraction of hidden layer weights from spatial filters and subsequent dipole localization showed brain areas associated at respective stages, including internal-oriented attention during preparatory interval, working memory and inhibitory control during problem solving, and motor response during solution realization. The result from individual level testing suggests the possibility of individual variability in self-report and/or corresponding EEG activity during insight. In addition, further work on stabilization of model is necessary for reproducible feature extraction and evaluation.
Metrics
51 File views/ downloads
64 Record Views
Details
- Title
- Classification of insight and analytic problem-solving using an interpretable deep-learning model based on single-trial EEG data
- Creators
- Yongtaek Oh
- Contributors
- John Kounios (Advisor)
- Awarding Institution
- Drexel University
- Degree Awarded
- Doctor of Philosophy (Ph.D.)
- Publisher
- Drexel University; Philadelphia, Pennsylvania
- Number of pages
- xii, 98 pages
- Resource Type
- Dissertation
- Language
- English
- Academic Unit
- Psychological and Brain Sciences (Psychology); College of Arts and Sciences; Drexel University
- Other Identifier
- 991018528111704721