Characterizing Learning Dynamics under Relative Reparameterization of Singular Models

arXiv:2206. 08598v2 Announce Type: replace-cross Abstract: A common way to analyze learning of statistical models is to consider operations in the models parameter space, however this becomes challenging when there is no one-to-one mapping between the parameter space and the underlying statistical model space.