November 14, 2024
12:15 pm – 1:15 pm
Speaker:
Samson Cheung, Electrical and Computer Engineering, University of Kentucky
Where:
327 McVey Hall
(Zoom link: https://uky.zoom.us/j/82467171189)
Title:
Challenges in building AI systems for Smart Health
Abstract:
Artificial intelligence is touted as the next frontier in healthcare, promising to revolutionize medical research and deliver equitable and low-cost care to all. However, there are many significant challenges to apply and develop AI for medical applications. Deep-learning based systems are at the forefront of AI but they are notorious at demanding large amounts of carefully labeled and annotated data. While simple labeling tasks can rely on crowdsourcing, medical data labeling requires expertise that could be rare and costly. In addition, there are usually significant bias and class imbalance issues with medical data. Expanding the knowledge base and diversity of data can be very difficult as privacy regulations limit or even prohibit medical data sharing with external organizations. In this talk, I will present results from a number of joint research projects with students and collaborators at University of Kentucky and University of California, Davis that aim at addressing some of these challenges. Specifically, I will discuss the use of semi-supervised learning and active learning to reduce labeling effort by developing novel weighting schemes to mitigate class imbalance and efficient pruning strategies to identify highly informative samples to label. To preserve privacy in distributed AI applications, we develop novel differentially private generative models for data sanitization and optimal acquisition functions for private active learning. The proposed methods are tested on both standard image benchmarks and target medical datasets used in digital pathology and autism diagnosis.