Interpreting FCDNNs via RG on Exponential Family

Collection of new arXiv submissions (stat.ML / ML) covering theoretical and methodological advances: a proposed interpretability link between fully-connected DNN training and renormalization group (exponential-family inputs); an SVM-based framework for extreme quantile regression with heavy-tailed covariates; impossibility and sufficient-condition results for zero-shot super-resolution in operator learning; ERICA, a pipeline for quantifying replicability of clustering; Riemannian stochastic optimization (SMAVE) for sufficient dimension reduction with improved runtime; a parameter-free method实现