Scalable Federated Learning

Machine learning is already widely used for a variety of tasks.

Initial approaches were strictly centralized, but have recently evolved into the so-called federated learning, primarily to address privacy concerns. In this model, multiple clients can collaboratively train a model under the guidance of a centralized server.

The clear advantage here is that the data remains with the clients, and they can still influence the global model in their individual training sessions.

One of the first deployed systems is at Google, for character recognition on smart phones, but is too slow when operating at massive scale.

We have identified a specific societal instance (fatal Tesla Model X crash) in which fast, distributed and federated learning could have prevented a loss of life.

Moreover, 21st-century pandemics present dramatic societal problems and require new scalable federated learning techniques.

This project aims to develop a highly scalable, flexible, extensible, distributed federated machine learning (Scalable Federated Learning, for short) approach that can directly benefit public health and wellness. Special attention will be paid to the so-called outliers (high-entropy samples).

We have two main objectives, and our work is structured around them: 1) Creation of a feasible and flexible scalable model for the life sciences and 2) Development of the scalable, flexible, federated machine learning framework.