Abstract: A hierarchical network as a simple model of the visual system is presented and trained with an algorithm for unsupervised learning of invariances. If trained with moving patterns, the network learns to generate a largely translation invariant representation, which can be used for pattern recognition (what-information). This generalizes also to new patterns. Interestingly, the same learning principle also yields where-information about the location of the patterns independent of pattern identity. Network performance, learned network structure, and performance dependence on various parameters are investigated and show the robustness of the system.