ICML 2019 Expo Talk

June 9, 2019

Expo 2019 Schedule »

Flea: An open source framework for federated learning

Sponsor: doc.ai

Organizers:
Neeraj Kashyap (doc.ai)

Presenters:
Neeraj Kashyap (doc.ai)

Abstract:

Federated learning allows data scientists to build models from updates computed at sites of data collection and storage. This introduces a different modality as compared to the traditional model-building workflow, wherein data is collected centrally before being used to train models. The new modality is especially powerful for cases where models have to be trained on data which is privacy-sensitive, as is the case in healthcare or finance. There are currently very few examples of federated learning being used at scale, the most prominent one being Google's training of the Google Keyboard on Pixel devices. In this talk, we introduce Flea, a production-ready, open-source federated learning framework. We describe the components that comprise Flea -- the mobile clients as well as cloud-based clients, and the Flea circus backend -- and how they can be integrated into Machine Learning workflows.