Abstract

The question of shift and size invariance in the primate visual system is discussed. After a short review of the relevant neurobiology and psychophysics, a more detailed analysis of computational models is given. The two main types of networks considered are the dynamic routing circuit model and invariant feature networks, such as the neocognitron. Some specific open questions in context of these models are raised and possible solutions discussed. 1 Introduction The ease with which we recognize common objects from di#erent distances and perspectives and under di#erent illuminations gives the impression that invariant object recognition is a trivial task. However, an apparently small change in the stimulus can cause a dramatic change in the retinal activity pattern. Assume, for example, you are looking at a zebra and change your gaze by just one width of the zebra stripes. Many responses of retinal sensors will be inverted (change from low to high or vice versa) causing a dramatic...

Citations