LoRKD - Low-Rank Knowledge Decomposition for Medical Foundation Models

Explore a novel approach to improving medical foundation models through knowledge decomposition in this 49-minute conference talk from Stanford University's MedAI Group Exchange Sessions. Learn about Low-Rank Knowledge Decomposition (LoRKD), a framework that addresses the domain gap challenges faced by large-scale medical foundation models when applied to specific tasks. Discover how this innovative method deconstructs foundation models into multiple lightweight expert models, each specialized for particular anatomical regions, while simultaneously reducing resource consumption. Examine the technical components including low-rank expert modules that resolve gradient conflicts between heterogeneous data and efficient knowledge separation convolution that achieves knowledge separation within a single forward propagation. Understand how these decomposed models achieve state-of-the-art performance on segmentation and classification tasks while exhibiting superior transferability on downstream tasks, often surpassing original foundation models in task-specific evaluations. Gain insights from Jiangchao Yao, assistant professor at Shanghai Jiao Tong University and research scientist at Shanghai AI Laboratory, whose research focuses on trustworthy machine learning and AI4Science applications, and who has published over 80 journal articles and conference papers in the field.