An interesting parameter appears in decentralized machine learning systems that is completely missing in a centralized FL system -- number of collaborators per node. This parameter can have a significant impact on the convergence of models in a decentralized setting.
A research question I am interested in understanding --
How does performance (convergence) scale as we increase this number. Is the improvement linear, exponential or logarithmic?
How does 1. change as we increase the number of users. Is there a scaling law here?
How do the above two results compare across different topologies?
In a truly decentralized system, number of collaborators per node will be different for different nodes. Therefore, what would be a good metric for comparison? Average degree of the graphs, centrality, clustering coefficient etc.?
There are too many moving pieces in this problem so a systematic research study will be a pretty impactful paper. The results will nicely corroborate with a strong body of empirical and theoretical work in graph theory and complex systems.
An interesting parameter appears in decentralized machine learning systems that is completely missing in a centralized FL system -- number of collaborators per node. This parameter can have a significant impact on the convergence of models in a decentralized setting.
A research question I am interested in understanding --
There are too many moving pieces in this problem so a systematic research study will be a pretty impactful paper. The results will nicely corroborate with a strong body of empirical and theoretical work in graph theory and complex systems.