Think Locally, Act Globally: Federated Learning with Local and Global Representations

Federated learning is a method of training models on private data distributed multiple devices. To keep device data private, the global model is trained only communicating parameters and updates which poses scalability challenges large models. To this end, we propose a new federated learning algorithm jointly learns compact local representations on each device and a global across all devices. As a result, the global model can be smaller since it operates on local representations, reducing the number of communicated. Theoretically, we provide a generalization analysis which shows a combination of local and global models reduces both variance in the data well as variance across device distributions. Empirically, we demonstrate local models enable communication-efficient training while retaining. We also evaluate on the task of personalized mood prediction from-world mobile data where privacy is key. Finally, local models handle data from new devices, and learn fair representations that protected attributes such as race, age, and gender.