The presentation explores an advanced methodology based on Probabilistic Learning on Manifolds (PLoM) for managing and updating statistical surrogates in stochastic systems with uncertainties ranging from nano to macro scales. The core challenge is constructing an updated statistical surrogate model from a target dataset for a vector-valued random quantity of interest given a vector-valued random control variable, complicated by the small size of both training and target data sets, with the latter being incomplete. This involves leveraging small training datasets and incorporating target data constraints through PLoM using the Kullback-Leibler divergence minimization principle. We consider two cases for the random mapping that relates the random quantity of interest to the random control variable: random manifolds due to uncertainties and random manifolds due to random latent variables. Updated statistical surrogates are constructed using learned joint probability measures derived from the small training and target datasets, using either conditional statistics or Polynomial Chaos Expansion, for which the input is the random control variable. Applications presented will span various scales. At the nano scale, we explore nonlinear dynamics for atomistic collisions with random control and latent variables. Concurrent multiscale modeling with nonlinear random materials described by a random field without scale separation is examined at the micro-meso-macro scale. At the macro scale, we focus on nonlinear stochastic dynamics of a nozzle with stochastic excitation and random material properties.