More about HKUST
Causal-Driven Interpretation and Inspection of Deep Learning
PhD Thesis Proposal Defence Title: "Causal-Driven Interpretation and Inspection of Deep Learning" by Mr. Zhenlan JI Abstract: Deep learning systems have achieved remarkable success across various domains, yet their black-box nature and inherent complexity present significant challenges to interpretability, reliability, and software quality assurance. To address these challenges, this thesis leverages causality, a canonical analysis framework for understanding complex systems, to enhance the interpretability and inspection of deep learning systems. This thesis establishes causality-based methodologies that offer a principled and scalable foundation for the interpretation and analysis of deep learning systems across various aspects and granularity levels, from neuron-level interactions to system-level evaluations. Specifically, this thesis contains three key contributions. First, it introduces Causal Coverage (CC), the first causality-aware test coverage criterion for deep neural networks, which formalizes neuron interactions and quantifies test adequacy based on uncovered causal dependencies. Second, it proposes a causal analysis framework for fairness trade-offs in machine learning pipelines, enabling systematic identification and quantification of causal trade-offs among fairness, robustness, and performance metrics. Finally, it develops a causal evaluation approach for large language model (LLM)-based code generation, constructing causal graphs over prompt and code features to explain and improve LLM outputs through counterfactual analysis. Date: Friday, 4 July 2025 Time: 2:00pm - 4:00pm Venue: Room 5501 Lifts 25/26 Committee Members: Dr. Shuai Wang (Supervisor) Prof. Nevin Zhang (Chairperson) Dr. Dan Xu