More about HKUST
Enabling Explainable AI with Transformer Models: Opportunities and Limitations in Visual and Textual Concept Generation
The Hong Kong University of Science and Technology Department of Computer Science and Engineering MPhil Thesis Defence Title: "Enabling Explainable AI with Transformer Models: Opportunities and Limitations in Visual and Textual Concept Generation" By Mr. Ao SUN Abstract: Recent advances in neural networks have driven remarkable progress in computer vision (CV), yet their opaque decision-making has fueled growing interest in Explainable AI (XAI). Conventional explanation methods often depend on linear segmentation and manual annotations, limiting understandability and scalability. Meanwhile, breakthroughs in transformer-based Large Language Models (LLMs) and Vision Language Models (VLMs) offer new opportunities to produce automated, high-quality concept-explanations. In this thesis, we investigate how such models can serve as enablers of XAI to enhance the generation of visual and textual concepts, and we further assess their limitations in fulfilling this role effectively. For visual concepts, we propose the Explain Any Concept (EAC) framework, which leverages the Segment Anything Model (SAM) to faithfully identify human-understandable image regions that influence a target model's predictions. For textual concepts, we introduce the Hierarchical-Concept Bottleneck Model (Hi-CBM), which leverages LLMs to automatically generate conceptual annotations that are both richly informative and well-structured. The richness of these concepts allows any CV model, when processing an image, to perform more faithful reasoning before making a final decision, while their structured organization filters redundant information and mitigates the information-leakage issues of traditional CBMs. To assess the limitations of transformer-based models in generating reliable explanations, we propose the Fast and Slow Effect (FSE) framework. FSE assesses a single model's ability to generate effective concepts by comparing classification performance in two modes: fast mode, simulating a black-box making direct predictions without rationales, and slow mode, simulating an interpretable expert that reasons over conceptual evidence. On specialized datasets, slow mode underperforms fast mode by about 30%, whereas on general-purpose datasets, it outperforms fast mode by roughly 10%. Overall, this thesis lays a strong foundation for transformer models to serve as enablers of XAI, and highlights their understandability to specialized tasks as a critical frontier for future research. Date: Wednesday, 8 October 2025 Time: 2:00pm - 4:00pm Venue: Room 5501 Lifts 25-26 Chairman: Dr. Dan XU Committee Members: Dr. Shuai WANG (Supervisor) Prof. Long QUAN