Coping with Data Heterogeneity in Federated Learning
Published:
Federated Learning (or, more generally, decentralized machine learning) is becoming increasingly important due to the geographically distributed nature of data. Data generated by IoT devices, video cameras, and mobile phones are often stored directly on the device or on nearby edge devices. It is infeasible to transfer all of this data to a centralized location for model training. Sending data over wide-area network (WAN) links is ofent slow and expensive. Futhermore, legal requirements and privacy regulations are becoming more prevalent, which restrict where data can be transferred and stored.