ARTIFICIAL ATTENTION AS A BASIS FOR COGNITIVE ROBOTS
Helge Ritter
Bielefeld University, Germany
The synthesis of artificial attention mechanisms becomes an important part of the realisation of technical systems that need to rely on intelligent behavior, such as robots or complex software systems. At the same time, such work may provide hints for (and – vice versa – benefit from) a better understanding of natural attention mechanisms.
We discuss three different levels of attention mechanisms that appear important to characterize elements of artificial attention systems: at the lowest level, the goal is to focus visual perception on salient constituents of a scene. The next higher level is concerned with focusing attention on significant relationships among detected constituents. The highest level then is concerned with attending to significant actions. We discuss how at each level the necessary "elements of artificial attention" may be implemented in artificial neural systems and report on a demonstrator that integrates some of these elements into an interactive robot system comprising a binocular robot head and a multifingered hand that attends to a human instructor to become directed by gesture and speech.
HYPERBOLIC SELF-ORGANIZING MAPS:
MATCHING DATA DISPLAYS TO HUMAN ATTENTION
Helge Ritter
Bielefeld University, Germany
Over the years, Self-organizing maps have become a successful tool to generate "insightful" two-dimensional visual maps of data of very diverse kinds. However, human attention can link a focused item with items from a "conceptual neighborhood" that is much richer than a two-dimensional euclidean surround. We argue that hyperbolic space with its exponential growth of neighborhood volume (as a function of distance) can provide a much better approximation to this structure and thus should offer a better substrate for creating visual "concept maps" of data of various kinds. To put this into practical use, we show how self-organizing maps can be created on regular discretizations of the hyperbolic plane and compare their properties with those of conventional self-organizing maps. Finally, we consider the application domain of text document mining as a major realistic application scenario. In this case, our approach leads to a highly cross-disciplinary combination of three basic principles from three different fields: (i) the vector space representation of text documents in text mining, (ii) a brain model for the formation of topographicfeature maps, and (iii) the geometry of hyperbolic space. We show how the resulting Hyperbolic Self-Organizing Maps (HSOMs) can create conceptually ordered document maps that combine conceptual clustering, good visualization and ease of browsing in a very appealing way and that remain highly competitive with more specialized state-of-the art document classification algorithms that lack the simultaneous benefits of visualization and browsing.