CS@W&M PhD student Viet Duong and his advisor Prof. Huajie Shao Win Best Paper Award at ACM SIGKDD 2024 for Novel AI Interpretability Framework
Computer Science PhD student Viet Duong, along with his advisors and collaborators, have been honored with the Best Paper Award at the prestigious ACM SIGKDD International Conference on Knowledge Discovery and Data Mining in 2024. Notably, this paper was selected at the best of 2,046 submissions. The paper titled “CAT: Interpretable Concept-based Taylor Additive Models” developed a novel interpretable framework capable of explaining DNNs predictions through high-level concepts that humans can easily understand.While deep neural networks (DNNs) have demonstrated remarkable success in various areas, the lack of interpretability impedes their deployment in high-stakes applications, such as autonomous vehicles, finance, and healthcare. Thus, enhancing DNN interpretability has emerged as a pivotal area of research in recent years.
“Human explanations often rely on concept-based reasoning, which semantically groups low-level features into high-level concepts, and then explains decisions using these high-level concepts. For example, in medical diagnostics like diabetes, physicians usually explain their conclusions by referring to high-level factors, such as family history, medical history, dietary patterns, and blood tests.” said Viet Duong Ami, lead author of the paper and PhD Candidate at Computer Science, William & Mary.
This research papers introduced a novel interpretable concept-based Taylor additive model called CAT to explain DNNs predictions. CAT does not require domain experts to annotate concepts and their ground-truth values. Moreover, it can explicitly learn the non-linear relationship between the inputs and outputs using polynomials.
The CS Department celebrates this significant achievement and the team's contributions to advancing the field of machine learning. Their work not only enhances our understanding of DNNs but also facilitate the deployments of DNNs in high-stakes applications.
