The implementation of an adaptive visual system founded on the detection of spatio-temporal invariances is described. It is a layered system inspired by the hierarchical processing in the mammalian ventral visual pathway, and models retinal, early cortical and infero-temporal components. A representation of scenes in terms of slowly varying spatio-temporal signatures is discovered through maximising a measure of temporal predictability. This supports categorisation of the environment by a set of view cells (view-trained units or VTUs) that demonstrate substantial invariance to transformations of viewpoint and scale.