Revision with unchanged content. The perception of dynamic spatio-temporal patterns is a fundamental part of visual cognition. Psychophysical experiments have shown that a spatiotemporal memory for early vision is evident and that the processing is realized in different representational layers of abstraction. The properties of this memory structure are reflected in a multi-layered neural network model which was developed in this work. Major architectural features of the model are recurrent derivatives of Kohonen's self-organizing maps. The model has the advantage of a self-teaching learning algorithm and stores temporal information by local feedback in each computational layer. Prediction capability was integrated by adding neural associative memories to each processing layer.