Completion Date
Summer 8-19-2022
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Program or Discipline Name
Data Sciences
Abstract
Advancements in deep learning or deep neural networks have made it possible to reach expert-level performance in a variety of applications, even in challenging situations. However, a central challenge in all deep learning, as well as machine learning applications, is dealing with its dependency on the quality of data which can be significantly impacted by biases, confounders, and irrelevant variations in data which leads to spurious relationships and erroneous decisions. The main purpose of this dissertation is to build a robust deep learning model which considers and mitigates these biases. Another challenge with the deep learning model is learning associations present in the data rather than causations. This also leads to bias problems and non-interpretable systems. So, the purpose of this dissertation also includes introducing causality in the deep learning models. Thus, developing a novel deep learning model to learn bias invariant features and learn causal discoveries are promising areas of research with high potential impact.
The overarching goal of this dissertation is to improve the performance, reliability, and generalization ability of deep learning even in the presence of biases and spurious associations in the data. This entails several research directions. First, we introduce a decorrelated framework that addresses the imbalanced and scanner dependencies issues present in the Parkinson’s Disease (PD) dataset. Second, we further define the general formulation of decorrelated deep learning models. This provides the foundation for generic bias mitigation analysis and the design of robust decorrelated deep learning models. This dissertation also focuses on the topic of Granger Causality (GC) introduction in the deep learning model. Thus, the third research direction includes extending the LSTM-based Granger Causality framework to incorporate Graph Neural Network (GNN) and distance correlation which enables improvement in the performance of the deep learning model and provides interpretable GC interactions.
We propose a novel bias mitigation method for deep learning models by leveraging the distance correlation function to decorrelate the features and biases to provide a robust solution. We explore the use of this method in neuroimaging study settings for disease classification. We also derive the generic decorrelation-based bias mitigation framework for different data scenarios and different deep learning architectures. These results show how our approach provides a robust, flexible, scalable, and generic framework that improves the performance of deep learning models while reducing bias effects on model predictions. In addition to this, we define a mathematical framework to introduce the fusion of GC with GNN and distance correlation and showcase their success in learning complex non-linear Granger causal connections. We study the implications of our work in the deep learning field and discuss future work to further leverage this robust decorrelated framework and improve the performance irrespective of the quality of data.
Recommended Citation
Patil, P. (2022). Decorrelated Deep Neural Networks: Learning Bias Invariant & Scanner Independent Features, and Causal relationships Using a novel deep learning methods based on Distance Correlation. Retrieved from https://digitalcommons.harrisburgu.edu/anms_dandt/4