Geometric deep learning in neuroimaging and human reward behavior

Azcona, Emanuel Aquiles

doi:https://doi.org/10.21985/n2-7w07-j654

Work

Geometric deep learning in neuroimaging and human reward behavior

Public

Download PDF

Download All Files (.zip)

In recent years, machine learning on graphs (or networks) has gone from a niche topic with only a few active researchers worldwide, to a heavily invested field with novel use cases for dealing with relationships and/or interactions within complex systems in the natural and social sciences. Traditionally, choosing the right type of model for leveraging the inductive biases of the task at-hand is a crucial step in machine learning scenarios (mostly supervised) because "there ain't no such thing as a free lunch," as the saying goes. Convolutional neural networks (CNNs) in particular, have been incredibly effective in numerous image processing problems and are becoming the de facto choice for considering data with grid-like topology. In a graph setting, where a grid-like structure is not always a guarantee, it is useful to leverage relational inductive biases within deep learning architectures in order to build systems that can learn, reason, and generalize from graph data. Graph-structured data is ubiquitous and all around us; often real-world entities are characterized by their connection(s) to other things. Recent advances in research on graph representation learning, particularly geometric deep learning (GDL), has led to a plethora of techniques for deep graph embeddings, generalizations of CNNs to graph data, and a reframing of neural message-passing approaches to graphs inspired by belief propagation. By maintaining the notion of representation or feature learning, and learning by local gradient-descent type methods, advances in GDL have led to new state-of-the-art results in several domains, including social network analysis, 3D surface manifold modeling, mapping/way-finding, molecular modeling, question answering, and recommender systems. The goal of this body of work was to provide a technical synthesis of GDL, through some methodological foundations and a demonstration of various benefits of GDL in structural neuroimaging and neuropsychological indicators, specifically human reward behavior. We begin with a discussion of GDL, specifically through GNN formulation, which has become amongst one of the fastest-growing paradigms for deep learning on graphs. Then we provide novel use cases in the analysis of human brain shape using 3D mesh surface manifolds to improve upon the state-of-the-art in machine learning for the classification of Alzheimer's disease and generating 3D brain models that are based on phenotypic priors. This thesis concludes with a new advancement in modeling human reward behavior as heterogeneous graphs (i.e., varying node/edge types), specifically using a portfolio of neurocognitive features to describe human preference towards a stimulus set, which are captured using a non-operant picture rating task across multiple distinct cohorts of human participants. Although comparatively nascent to other graph-based methods in the biomedical arena, the success of deep graph embeddings provided through GDL continues to showcase the advantages of graph representation learning in neuroimaging and computational cognitive science.

Creator