|
Shams,
Ladan EMERGENCE OF RECURRING VISUAL PATTERNS AS NEW REPRESENTATIONS The visual system exploits various sources of information in order to segment an object from the background and recognize it. These sources include the data-driven or bottom up cues of color, motion, stereo, etc. as well as model-driven or top-down information about the object models stored in the memory. How are these stored object models formed? Is presence of segmentation cues such as color, motion, stereo, texture, Gestalt relations, etc. necessary to segment novel objects from the background? The findings of recent psychophysical study (Brady 1998) show that such cues are not necessary so long as the novel object is viewed against varying backgrounds (i.e., appears in different scenes). That is, after viewing a small number of visual patterns in which a novel subpattern (object) recurs, the recurring subpattern, which was not segmentable from the background at the beginning, emerges as a familiar pattern and becomes segmentable! These results point to an active unsupervised learning mechanism which is tuned to the extraction of recurring patterns from the sensory input by means of matching various scenes with each other. We present a model of how such learning may take place. Our model takes a small number of scenes (of the type used in Brady's experiments) as input, and based on a V1 complex cell type representation, and using the labeled graph matching and cross correlation methods, outputs the recurring patterns as the emergent new representations. This model does not use any a priori knowledge about the structure, location or size of the recurring patterns. We had previously shown how the same algorithm can learn object shape primitives from exposure to a small number of composite objects.
|