Reducing Model Complexity and Overcoming Overfitting: Deep Learning Algorithms and Medical Applications
Open Access
Author:
Chen, Jianhong
Graduate Program:
Mathematics
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
October 06, 2020
Committee Members:
Jinchao Xu, Dissertation Advisor/Co-Advisor Wenrui Hao, Committee Member Xiantao Li, Committee Member Ludmil Tomov Zikatanov, Committee Member Xiang Zhang, Outside Member Alexei Novikov, Program Head/Chair Jinchao Xu, Committee Chair/Co-Chair
Keywords:
Deep learning Compressed sensing Medical application
Abstract:
In this dissertation, we first propose the xRDA algorithm with an adaptively weighted $\ell^1$-regularization scheme and momentum for training sparse neural networks.
Then we compare some variants of the cross-validation algorithms and propose a validation set theorem
on the reliability of the validation accuracy as a measure of the performance of deep learning models for classification.
Moreover, we use a fluid-structure interaction model with plethysmography signal only to predict the continuous blood pressure.
Lastly, we apply deep learning models on the pulse wave signal to classify whether people are pregnant or not.