The proposed senior design project centers on two tasks. The first is prototyping of a federated learning framework that deploys in-network data aggregation. This would involve working with a publicly available network prototyping platform developed by the National Science Foundation (NSF) to enable experiments that explore limitations of networking infrastructure. Such experiments would help users learn how to process and exchange data amongst the network nodes, ultimately enabling efficient implementations of the FL algorithm. A selected machine learning task that involves training of medium to large models could be used as a vehicle to explore the performance/convergence of the described FL scheme. The second task involves designing an algorithm that adaptively decides how to route the data and at which nodes to perform aggregation so as to optimize the convergence rate of federated learning. At a high level, this would boil down to determining the best (i.e., having the lowest overall delay) in-tree rooted at the client based on the delays between the network nodes participating in the FL task.

Team Members:

Ayush Bhattacharya
Brian Vo
Carlos Mari
Ritvik Mahendra
Shakthi Prabhakar
Sanjay Gorur