Explore a comprehensive conference talk on rethinking architecture design for tackling data heterogeneity in federated learning of medical AI. Delve into the challenges of federated learning across heterogeneous devices and discover how replacing convolutional networks with Transformers can reduce catastrophic forgetting, accelerate convergence, and improve global model performance. Learn about the speaker's research, experimental results, and the potential impact on future explorations in robust architectures for federated learning. Gain insights into the intersection of AI and medicine through this MedAI Group Exchange Session, featuring an interactive discussion and Q&A following the presentation.
Rethinking Architecture Design for Data Heterogeneity in FL - Liangqiong Qu
Stanford University via YouTube
Overview
Syllabus
Introduction
Background
Summary
Deep Learning Methods
Data Heterogeneity
Phenology
Recent Efforts
Motivation
Experiments
Simulations
Experimental Results
Conclusion
Drawbacks
Questions
Pretrained models
Takeaway
Why it works better
Taught by
Stanford MedAI
