We have previously shown that maximum-likelihood Hebbian learning rules are able to find interesting structures in visual data; in particular, such rules are able to find filters of video data which are both local in time and local in space. We have also previously applied lateral connections derived from the rectified Gaussian distribution to globally organise learning (Corchado and Fyfe, 2002 International Journal of Computational Intelligence and Applications in press). In this paper, we combine these two ideas. We show that on a standard artificial data set (composed of a mixture of horizontal and vertical bars) the method enables clean separation of each orientation, even when the training set only contains mixtures of the two orientations. We further show that, on real video sequences, the type of filters found depends on the nature of the data set and that the lateral connections enable us to globally organise the filters found so that topographic relationships are maintained across the filters.