Abstract

Generative Artificial Intelligence models are algorithms to stochastically create new data instances that resemble a given training dataset by learning the underlying patterns and distributions. Most models include neural network components, for which the standard architectures are not flexible enough to accurately reproduce the tail behavior in data. The goal of this talk is to show through theoretical calculations and simulations how neural networks with potentially several layers transform the distribution of extreme events of the input variables (typically independent and identically distributed) towards the output variables. Particular focus will be given to heavy-tailed distributions, such as regularly varying and more general subexponential distributions.