The relevance of non-parametric reconstructions of cosmological functions
lies in the possibility of analyzing the observational data independently of
any theoretical model. Several techniques exist and, recently, Artificial
Neural Networks have been incorporated to this type of analysis. By using
Artificial Neural Networks we present a new strategy to perform non-parametric
data reconstructions without any preliminary statistical or theoretical
assumptions and even for small observational datasets. In particular, we
reconstruct cosmological observables from cosmic chronometers, $f\sigma_8$
measurements and the distance modulus of the Type Ia supernovae. In addition,
we introduce a first approach to generate synthetic covariance matrices through
a variational autoencoder, for which we employ the covariance matrix of the
Type Ia supernovae compilation. To test the usefulness of our developed
methods, with the neural network models we generated random data points mostly
absent in the original datasets and performed a Bayesian analysis on some
simple dark energy models. Some of our findings point out to slight deviations
from the $\Lambda$CDM standard model, contrary to the expected values coming
from the original datasets.