AI RESEARCH
Beyond ReLU: Bifurcation, Oversmoothing, and Topological Priors
arXiv CS.LG
•
ArXi:2602.15634v2 Announce Type: replace Graph Neural Networks (GNNs) learn node representations through iterative network-based message-passing. While powerful, deep GNNs suffer from oversmoothing, where node features converge to a homogeneous, non-informative state.