Optimal Rates for Generalization of Gradient Descent Methods with Deep Neural Networks

Collection of recent ML/statistics preprints (arXiv) covering theoretical and applied advances: two complementary papers establish minimax-optimal generalization rates for deep ReLU/smooth-activation networks trained by gradient descent/SGD and relate overparameterized DNN dynamics to kernel methods; stability and concentration results that extend algorithmic stability guarantees to finite L_p moments; empirical transfer-operator methods and finite-sample change detection for noisy dynamical systems; analysis of in-context learning via low-dimensional subspace models; a semiparametric Deep-SIn