46 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Towards Understanding Transfer Learning with Applications to Medical Imaging
Cornell University & Google Brain
Recording
Abstract
Maithra Raghu surveys what pretraining transfers when a network is fine-tuned on another task. She compares evidence from natural and medical images, where differences in the data and tasks challenge standard intuitions about feature reuse. The talk examines the role of model size, the performance of lightweight architectures, and improvements that can occur independently of the transferred features.