An RBF Network that Learns Some Aspects
of Perceptual Organization
Thomas M. Breuel
IDIAP, C.P. 609, 1920 Martigny, Switzerland
Introduction. The grandmother cell" has a long history in vision. Sometimes it is being proposed as a serious model for recognition problems,4 and, more often, it is given as a straw man for how recognition could not possibly work.9 In its usual form, the grandmother cell is described detecting the presence of features in particular locations of an image and firing when enough evidence for a grandmother (or whatever other kind of object the cell is specific for) has been found.
Recently, there has been renewed interest in the grandmother cell because it can be shown that a small collection of such cells, each specific for some 2D view of a 3D object, can collectively approximate well the visual appearance of a 3D object viewed from different angles.10,1 Those theoretical results together with psychophysical support5 suggest that a grandmother cell approach, when worked out properly, might account for some aspects of visual object recognition.
Unfortunately, when applied to actual images, such approaches to recognition are empirically prone to false positive detections (misdetections); this phenomenon can be understood in a statistical framework.2
The basic reason is that both because of model variation and because of the presence of occlusions in images, matches are never perfect, and a less than maximal activations of the input features to a grandmother cell must still be considered evidence for the presence of an object. But spurious features can conspire to activate enough of the inputs to a grandmother