Spreading activation layers, visual saccades, and invariant representations for neural pattern recognition systems

摘要: Abstract This paper shows how a simple spreading activation network in the form of two-dimensional (2D) diffusion followed by local maximum detection can quickly perform large number early vision tasks, for example, feature extraction, clustering, feature-centroid determination, and boundary gap completion, all on multiple scales. The results process be used to facilitate 2D object learning recognition from silhouettes generating representations bottom-up fixation cues which are invariant translation, orientation, scale. In addition, proposed suggests possible approach longrange apparent motion correspondence multiscale decomposition. theory is described implementation examples presented.