Federated learning is a new class of machine learning models pioneered by the likes of Google and Apple, which allows models to be trained in the cloud, but without requiring customer or user data to be sent or transmitted, in order to preserve their privacy, which grows ever more important in the age of GDPR, etc. Unlike the traditional machine learning modelling approach that collects raw data to train on a centralised model (‘taking the data to the algorithm’), federated learning keeps data stored locally and instead, sends the model to the users (‘taking the algorithm to the data’). Coupled with an added layer of unique encryption, this technique means data can be stored and kept privately, but still allow models to be built upon them.

In this session, the IFoA Federated Learning Working Party will demonstrate how this might apply to modelling insurance claims. Specifically how it might be possible for several different insurance companies to collectively build a combined, shared, claims frequency neural network (similar to the CMI for mortality), whilst keeping their data and predictions completely hidden from each other.

Speakers: 

Malgorzata Smietanka, UCL

Dylan Liew, Data Science Actuary, BUPA