Fausto Milletari, NVIDIA
gtc-dc 2019
We’ll discuss how federated learning addresses issues surrounding the creation of a large healthcare dataset. While training datasets are crucial to building a deep learning model capable of generalization and effective performance, several issues prevent the creation of a large healthcare dataset. These include privacy issues, data annotation costs, intellectual property, different standardization techniques and file formats, and government policies that limit the capability to share data across institutions or companies. Our technique learns and aggregates models across different institutions without sharing any training data, creating a more powerful model. We’ll present the core concepts and challenges of the application of federated learning methods to healthcare, and discuss experimental results on challenging federated learning tasks.