A Distributed Learning Simulation Platform for Edge Hierarchies

We develop a distributed learning simulation platform that allows users to create multi-level Edge hierarchy for a given application by simulating resource constrained Edge devices and communication links amongst them. The resulting Edge computing hierarchy is used to run a given DNN for the application in a data distributed fashion, rolling up learned parameter values up a hierarchy of parameter servers that merge parameters received from the lower levels. The root of this hierarchy has the latest model, which is then pushed lazily back down the tree to the Edge servers. The platform can be used to study cost vs accuracy analysis of a given application for different Edge hierarchy configurations. We use handwritten digit recognition problem as a case study to show the usefulness of our platform.