Attentive object recognition in the selective tuning network

Abstract:
A hierarchical winner-take-all network derived from the selective tuning model (Tsotsos et al., 1995) allows position-invariant object recognition via selective attention. A notion of cooperation between features is introduced into the selective tuning winner-take-all framework to consider the spatial relationship between features when selecting winners.
The model is demonstrated with example recognition networks in which objects are represented as hierarchical conjunctions of parts and features culminating in the activity of on unit. It is shown that the attentional beam that follows from the selection of such an object unit tightly encompasses the object in the image. It is also shown that top-down inhibitory bias signals allow the selection of objects having particular low-level features. The robustness of the behaviour is investigated with a large number of images containing one or two automatically-generated figures.