FIART: Adaptive resonance model of feature integration, proto-objects formation and coherence theory of visual attention

Abstract

Abstract The Adaptive Resonance Theory (ART1, ART2, etc.) provides neural networks with means to model the parallel accumulation of features, followed by a serial search for matching feature-patterns. This process reminds the psychology of visual attention, in particular, feature integration and coherence theories. This paper presents the ART-based neural architecture (FIART), inspired by them. The FIART consists of multifeature-calculation layer, ART1-modules and inter-feature associations. One of the main problems in application of ART to processing of natural images is necessity to adapt ART to the diversity of visual features. This paper presents the bioplausible approach to resolve this problem. The FIART is a multifeature ART-architecture, based on separate ART i x -modules, each processing one visual feature (i) in one location (x) in input image. The ART-modules in FIART are connected by mutual-excitation associative connections A i , j , allowing to initiate the winner-takes-all competition between groupings of feature-patterns corresponding to “proto-objects” in the Coherence Theory of visual attention. The FIART neural network was verified on a set of simple images of textured objects (zebras) and showed its ability to cluster images of the same object and to form its internal representation, corresponding to the memorized image of the object in the brain. FIART was also compared to human behavioral studies of attention, textural features integration and proto-objects formation. The FIART is an expansion of ART1 to multifeature processing, as well as its application to modeling the feature integration theory of visual attention.