Computing with connections in visual recognition of origami objects

How does one represent Origami objects in a connectionist system?l How does one control the recognition process to take full advantage of the potential for massive parallelism available in the connectionist approach?l Furthermore, how does one make this control process robust enough to deal with imperfect and occluded input in predictable and "pleasing" ways?To answer these questions, we looked to research from two different sources; one was interdisciplinary research in animal vision and the other was, not surprisingly, computer vision.We looked to animal vision (Hubel & Wiesel, 1979; Marr, 1978) since animals are the only existing proof we have that such problems can not only be solved but can be solved exceedingly quickly (relative to current computers or programs).Our findings supplied the original motivation for the principles underlying the entire connectionist approach (Feldman & Ballard, 1981) and also helped in making several design choices within our system: l Small units that have only one unique output per computation.The complexity of the output is limited although the fan-out is very large.