Electrical and Computer Engineering
Xiaoxiao Li is an Assistant Professor in the Electrical and Computer Engineering Department at the University of British Columbia (UBC). Before joining UBC, Dr. Li was a Postdoc Research Fellow in the Computer Science Department at Princeton University. Dr. Li obtained her PhD degree from Yale University in 2020. Dr. Li received her bachelor’s degree from Zhejiang University in 2015. In the recent few years，Dr. Li has over 30 papers published in leading machine learning conferences and journals, including NeurIPS, ICML, ICLR, MICCAI, IPMI, BMVC, AAAI, and Medical Image Analysis. Her work has been recognized with the OHBM Merit Abstract Award, the MLMI Best Paper Award, and the DART Best Paper Award. In addition, she has received travel awards from NeurIPS/ICML/MICCAI/IPMI. Dr. Li has also organized a number of workshops on the topic of machine learning and healthcare. She is the Associate Editor of Frontiers in NeuroImaging and a reviewer for a number of international conferences and journals.
Dr. Li’s research interests range across the interdisciplinary fields of deep learning and biomedical data analysis, aiming to improve the trustworthiness of AI systems for healthcare.
“Federated Learning on Distributed and Imperfect Medical Data.”
Small data, noisy data, heterogeneous data, data labeling, and data privacy are the challenges in medical data analysis. We aim to address these problems in the federated learning framework, in which different institutions collaboratively train deep learning models without data sharing. In this project, we have developed multiple methods to address various learning scenarios. First, we propose FedBN, an effective strategy that uses local batch normalization to alleviate the feature shift before averaging models, to address an important problem of FL that different scanners and sensors in medical imaging result in non-iid data distribution. Second, we propose FedUL, which adds a surrogate task transition layer to train the classification model on unlabeled data based on empirical risk minimization. Third, graphs provide a natural way to represent the population data for disease analysis. We propose FedGDP for disease prediction on distributed population graphs to leverage the similarity of subjects. We expect to leverage the federated learning framework and techniques for imperfect data learning to accelerate AI deployment in medical applications.