In-network aggregation has been proposed to accelerate gradient aggregation in the distributed training of neural networks. My current research focus is on loss-tolerant transmission protocol for federated learning.
Some background can be found in the following papers.