Abstract of Dissertation AND PhD Defense - Minh Vu

 Abstract of Dissertation Presented to the Graduate School

of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctorate of Philosophy

EXPLAINING THE BOX: A GLIMPSE INSIDE DECISIONS OF MODERN AI
By
Minh Nhat Vu
September 2023

Chair: My T. Thai

Major: Computer and Information Science and Engineering

Recent years have observed a swift adoption of modern Artificial Intelligence (AI) models in real-world applications, especially those in critical and sensitive domains. It has become increasingly important to explain the predictions generated by those complex models. This dissertation is to present heuristic and theoretical findings centering around that emerging problem:

·         Developing a metric and algorithm to evaluate local explanations. While many explanation methods for deep learning models have been introduced, evaluating them remains challenging. In response, we propose the c-Eval metric and its corresponding framework to quantify the correctness of local explanation. Given a prediction of a deep neural network and its explanation, c-Eval is the minimum-distortion perturbation that successfully alters the prediction while keeping the explanatory features unchanged.

·         Improving explanation methods with topological perturbations. We introduce a novel perturbation scheme so that more faithful and robust explanations of perturbation-based explanation methods can be obtained. In particular, the study focuses on the impact of perturbing directions on the data topology. We show that perturbing along the orthogonal directions of the input manifold better preserves the data topology, both in the worst-case analysis of the discrete Gromov-Hausdorff distance and in the average-case analysis via persistent homology.

·         Providing more in-depth explanations for deep neural networks. To obtain more in-depth explanations for neural networks, we propose NeuCEPT, a method to identify critical neurons that are important to the model’s local predictions and learn their underlying mechanisms. We also provide a theoretical analysis to efficiently solve for critical neurons while keeping the precision under control.

·         Explaining predictions of modern neural networks on graph data. Although many explanation methods have been developed for deep models operating on grid-like data, e.g. time series, text and images, the counterparts for graph data are lacking. In response, we introduce PGM-Explainer, a Probabilistic Graphical Model (PGM) explainer for Graph Neural Networks (GNNs). We theoretically show that the resulting explanation guarantees to include all statistical information regarding the target of the explanation.

·         Investigating the limitation of explaining information generated by modern explainers. To circumvent the lack of internal information on the explained models, black-box explainers rely on the responses of the model on some perturbations of input data. We theoretically point out that this lack of internal access limits perturbation-based methods from uncovering certain crucial information about the predictions generated by Temporal Graph Neural Networks (TGNNs), i.e., a class of modern AI models.

The results of the works in this manuscript have a strong impact on a wide range of AI-related applications, especially those that require high levels of trust, safety, and security.

ANNOUNCEMENT OF Ph.D. Defense 

					Date Posted
FROM:  Computer and Information Science and Engineering Department					September 21, 2023
TO:  the Members of the Supervisory Committee for:
First Name	Last Name			M.I.
Minh	Vu			Nhat
SUPERVISORY COMMITTTEE MEMBERS
My Thai		Chair
		Co-Chair
Kejun Huang		Member
Kevin Butler		Member
		Member
		Member
Sandip Ray		External Member
1.  Name of Examination	Ph.D. Defense
2.  Degree Sought	Doctor of Philosophy
3.  Area of Specialization
4.  Thesis/Dissertation Title
5. Examination: Date, Time, Place	October 25, 2023		9:00 a.m.		Room CSE 404 ZOOM-https://ufl.zoom.us/j/3567244093