SPEAKER:  Dr. Minh Tang

ABSTRACT:

Two-sample hypothesis testing for random graphs arises naturally in   neuroscience, social networks, and machine learning. The talk discusses   the nonparametric problems of whether two finite-dimensional   random dot product graphs have generating latent positions that are
independently drawn from the same distribution, or distributions   that are related via scaling or projection. A consistent test procedure wherein the graphs are first embedded into  Euclidean space via spectral decomposition of the adjacency matrices followed by a kernel-based distance measure between the resultant embeddings is then presented. The talk concludes with a discussion of how the proposed test procedure might be applied to the general problem of identifying and classifying local structure in big data graphs, e.g., the identification of repeated processing modules in the neocortex as suggested by the cortical column conjecture, and the challenges that it entails.