A basic ingredient to understand large scale fluctuations in our present day Universe is the initial conditions. Using N-body simulations, we compare the clustering that arises from Gaussian and non-Gaussian initial conditions. The latter is motivated by a global texture model which has initial $J$-order correlations $\xibar_J$ close to the strongly non-Gaussian regime $\xibar_J \simeq \xibar_2^{J/2}$. The final amplitudes $S_J \equiv \xibar_J/\xibar_2^{J-1}$ in the non-Gaussian (texture) model evolves slowly towards the (Gaussian) gravitational predictions but, even at $\sigma_8=1$, are still significantly larger, showing a characteristic minimum with a sharp increase in $S_J$ with increasing scales. This minimum, which is between $10$ and $15 \Mpc$, depending on the normalization, separates the regime where gravity starts dominating the evolution from the one in which the initial conditions are the dominant effect. In comparing this results with galaxy clustering observations, one has to take into account {\it biasing}, i.e. how galaxy fluctuations trace matter fluctuations. Although biasing could change the amplitudes, we show that the possible distortions to the shape of $S_J$ are typically small. In contrast to the non-Gaussian (texture) predictions, we find no significant minimum or rise in $S_J$ obtained from the APM Galaxy Survey.

We report progress in the development of automatic star/galaxy classifier for processing images generated by large galaxy surveys like APM. Our classification method is based on neural networks using the Kohonen Self-Organizing Map approach. Our method is novel, since it does not need supervised learning, i.e. human factor, in training. The analysis presented here concentrates on separating point sources (stars) from extended ones. Using simple numerical experiments we compare our method of image classification to the more traditional (PSF-fitting) approach of DAOFIND.