Demonstrates viable peer-to-peer learning without a central aggregation server.
Gossip learning removes the server. Nodes hold models, pass them to random peers, and merge what arrives. There is nothing in the middle to trust or to fail.
The paper compares it against federated learning head to head, over the same tasks and the same churn and message-loss conditions.
Gossip learning is competitive. The result matters because the server in federated learning is usually justified on performance grounds, and here that justification does not hold.