Characterizing Learning Dynamics under Relative Reparameterization of Singular Models
This research paper introduces a novel technique called relative reparameterization to analyze the learning dynamics of singular statistical models. Singular models, common in machine learning, often exhibit slower convergence due to attractor behaviors. The proposed method aims to extract regular sub-models from these singular ones, theoretically and numerically analyzing convergence rates for gradient descent on Gaussian Mixture Models and Neural Networks. The study distinguishes between algorithmic and information-geometric factors influencing convergence by examining second-order methods and the Fisher Information Matrix. AI
IMPACT Introduces a theoretical framework for improving the analysis of learning dynamics in complex statistical models.