Abstract
In Federated Learning (FL), a primary challenge to the server-sideaggregation of client models is device heterogeneity in both loss landscapegeometry and computational capacity. This issue can be particularly pronouncedin clinical contexts where variations in data distribution (aggravated by classimbalance), infrastructure requirements, and sample sizes are common. Wepropose AsyncManifold, a novel asynchronous FL framework to address theseissues by taking advantage of underlying solution space geometry at each of thelocal training, delay-correction, and aggregation stages. Our proposal isaccompanied by a convergence proof in a general form and, motivated throughexploratory studies of local behaviour, a proof-of-concept algorithm whichperforms aggregation along non-linear mode connections and hence avoidsbarriers to convergence that techniques based on linear interpolation willencounter.