Figure 2. Decentralized stochastic neighbor embedding (dSNE) results for quality control metric of the ABIDE datasets ( Saha et al., 2017).

We randomly split the data into ten local and one reference dataset. The centralized results show ten different clusters. For three random splits of decentralized computation, we also obtain ten different clusters, and the number of clusters in the embedding is stable regardless of how the data are split among sites.