Date of Award
Master of Science
Electrical and Computer Engr
Categorical perception (CP) of speech is a complex process reflecting individuals’ ability to perceive sound and is measured using response time (RT). The cognitive processes involved in mapping neural activities to behavioral response are stochastic and further compounded by individuality and variations. This thesis presents a data-driven approach and develops parameter optimized models to understand the relationship between cognitive events and behavioral response (e.g., RT). We introduce convolutional neural networks (CNN) to learn the representation from EEG recordings. In addition, we develop parameter optimized and interpretable models in decoding CP using two representations: 1) spatial-spectral topomaps and 2) evoked response potentials (ERP). We adopt state-of-the-art class discriminative visualization (GradCAM) tools to gain insights (as oppose to the’black box’ models) and building interpretable models. In addition, we develop a diverse set of models to account for the stochasticity and individual variations. We adopted weighted saliency scores of all models to quantify the learned representations’ effectiveness and utility in decoding CP manifested through behavioral response. Empirical analysis reveals that the γ band and early (∼ 0 - 200ms) and late (∼ 300 - 500ms) right hemisphere IFG engagement is critical in determining individuals’ RT. Our observations are consistent with prior findings, further validating the efficacy of our data-driven approach and optimized interpretable models.
dissertation or thesis originally submitted to the local University of Memphis Electronic Theses & dissertation (ETD) Repository.
Moinuddin, Kazi Ashraf, "Decoding Perception of Speech from Behavioral Responses using Spatio-Temporal CNNs" (2020). Electronic Theses and Dissertations. 2147.