Amira Soliman

Researcher at RISE SICS

Abstract

The vast amount of data that people produce and share over the Internet has revolutionized many fields in Machine Learning (ML). Moreover, this data tends to originate more and more often from ubiquitous and autonomous sources such as mobile phones, fitness trackers, and IoT devices. Centralized and federated ML solutions represent the predominant way of providing smart applications for users. However, moving data to a central storage creates many privacy concerns. Therefore, ML models need to be trained in a collaborative and decentralized manner, similar to the way the data is originally generated without requiring any central authority for data or model aggregation. In this talk, we will be discussing the challenges of decentralized learning with skewed data distribution. Specifically, when there are no guarantees on the distribution of data owned by participating peers (i.e, local data can be highly skewed and non-IID data).