Learning from noisy and difficult data.

My PhD research focuses on developing lightweight methods that learn from differences between training samples. LiLAW adapts the weight given to easy, moderate, and hard examples; LiNC estimates how much to trust each label and corrects likely errors; and LiBaGS selects synthetic samples that fill useful gaps near a model's decision boundary. Related work examines fairness and proper use of synthetic data.

I work across image, tabular, time-series, text, and multimodal data. Projects with SickKids and UHN KITE have included pediatric cancer surveillance and automated pain assessment in older adults with severe dementia. I have also worked on surgical intervention prediction in pediatric renal ultrasounds and understanding bias in dermatological AI systems. Other current work explores large language models, medical vision-language models, agentic healthcare systems, socially grounded AI safety, AI literacy, and hospital AI governance.