More about HKUST
Reliable and Generalizable AI through Prediction-Critical Information Identification and Preservation
PhD Thesis Proposal Defence
Title: "Reliable and Generalizable AI through Prediction-Critical Information
Identification and Preservation"
by
Miss Luyu QIU
Abstract:
Reliable and generalizable artificial intelligence requires models to use
the right information for prediction. This proposal studies this requirement
through the lens of prediction-critical information: the input components,
model responses, or structural patterns that substantially influence a
model's prediction. Although model explanation and graph learning appear to
be different tasks, both depend on understanding what matters for prediction.
In model explanation, the goal is to identify which input regions drive a
trained model's decision. In graph learning, the goal is to encode and
preserve graph structures, such as paths, cycles, motifs, and functional
subgraphs, that determine graph-level predictions.
The first part of this proposal investigates reliable identification of
prediction-critical information in model explanation. Perturbation-based
explainable AI methods, such as RISE, OCCLUSION, and LIME, estimate feature
importance by masking, removing, or modifying parts of an input and
observing prediction changes. However, the perturbed samples used to infer
importance may fall outside the training distribution. Classifier outputs on
such samples can be confident but unreliable, causing saliency maps to
reflect perturbation artifacts rather than the model's true decision logic.
Chapter 3 addresses this problem by introducing an
out-of-distribution-resistant framework that estimates the inlierness of
perturbed samples and incorporates this reliability signal into explanation
generation and evaluation. In this way, the work improves the reliability of
identifying prediction-critical image regions under distributional
uncertainty.
The second part studies preservation and learning of prediction-critical
information in graph-level prediction through my accepted GEGNN work. Graph
labels often depend on task-relevant structural patterns, including paths,
cycles, motifs, cliques, and chemically meaningful substructures. Standard
GNNs may fail to capture such patterns, while naive graph augmentation may
destroy them. Chapter 4 addresses this challenge by proposing a generalizable
and expressive graph neural network that combines a $k$-path rooted subgraph
encoder, adaptive graph augmentation, and consistency-aware training. The
encoder improves the ability to represent prediction-critical structural
patterns, adaptive augmentation preferentially perturbs less important edges
while preserving core structures, and consistency-aware training encourages
stable predictions across structurally meaningful graph views.
Together, the two completed works support a unified thesis: reliable AI
requires identifying, validating, and preserving prediction-critical
information. Chapter 3 focuses on identifying critical input regions while
validating the reliability of perturbed samples. Chapter 4 focuses on
encoding and preserving critical graph structures while perturbing less
important components to improve generalization. This proposal develops
methods that make AI systems more reliable not by treating all input
components equally, but by understanding which parts of the data matter for
prediction and how these parts should be protected under uncertainty.
Date: Friday, 3 July 2026
Time: 2:00pm - 4:00pm
Venue: Room 3494
Lifts 25/26
Committee Members: Prof. Lei Chen (Supervisor)
Prof. Qian Zhang (Chairperson)
Dr. Xiaomin Ouyang