Lunch seminar: Privacy-aware machine learning method: Federated Learning

Presenter: Hongyi Zhang, PhD student, Chalmers

Federated Learning is an improved Machine Learning approach which enables edge devices to collaboratively learn a shared machine learning model. The data exchanged between the edge and central server are no longer user data but the updated model weights. Furthermore, as this model training method is distributed, there are chances to make the model become smarter and adapt to their environment. (For example, devices in Malmö and Stockholm can have slightly different models based on their unique environment.) Due to the rapid increase of data collected on the edge, it will become more expensive to gather data into a centralized server or cloud. Machine learning with distributed edge training will be the main approach for future AI engineering. This is also why Federated Learning is essential to industries.

Advantages of Federated Learning:

• Privacy-Preserving
• No need to send training data to a cloud or single server
• Fast model deployment and evolution
• Low latency and efficient bandwidth utilization

There is a classic deployment from Google, where they deployed Federated Learning in Gboard (A Mobile App) to collaboratively train models for improving keyboard query and predicting emoji.

In the software center, we made close collaboration with industries. With Volvo Cars, we validated Federated Learning in various industrial cases, such as steering wheel angle prediction, objective recognition, etc. We have proved the efficiency of FL and further improved the technique by introducing asynchronous protocols for heterogeneous hardware settings in real-world cases. With Sahlgrenska University Hospital, we are trying to help them estimate their resource allocation in a distributed manner and investigate the methods to continuously enhance the data privacy and security when applying machine learning in the medical health care field.

As the next step of our research, we would like to investigate the possibility of Federated Learning with various industrial scenarios. We will try to adding autonomously improving mechanism into participated nodes to enhance the model quality and node autonomy. We are eager to collaborate with companies to help them accelerate and improve system autonomy of their machine learning procedure.

__________________________________________________

Microsoft Teams meeting

Join on your computer or mobile app

Click here to join the meeting