Pruning deep neural networks for encoding and decoding the human connectome

• Main Author: Tang, Sean
• Other Authors: Jagath C Rajapakse
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computer applications::Life and medical sciences
• Online Access: https://hdl.handle.net/10356/149048
• Institution: Nanyang Technological University

The main focus of this project is to identify biomarkers of neurodegenerative disorders such as Alzheimer’s Disease (AD) and Parkinson’s Disease (PD) in functional Magnetic Resonance Imaging (fMRI) scans. Deep learning models can be used to encode the human functional connectome and classify between healthy subjects and patients with diseases, followed by a decoding process to identify salient features used in the classification. However, fMRI datasets have much more features than data samples, causing models to overfit easily. Existing solutions involving pruning the neural network range from recursive feature elimination which is too slow to a one-shot pruning approach which prunes too harshly. Thus, this project will explore the viability of improved pruning methodologies to attain an improved, sparser architecture. This project also goes beyond existing work on pruning multi-layer perceptron (MLP) to propose pruning approach for convolutional neural network (CNN), which can take in dynamic functional connectivity (dFC) matrices, as well as graph convolutional network (GCN), which is a better fit for encoding functional connectomes. The pruning algorithms proposed can also generalise to non-neuroimaging datasets, which is demonstrated by applying them to datasets like MNIST, CIFAR-10 and the CORA dataset, suggesting applications beyond the initial scope defined by this project.