When perceiving an object, human beings can construct its internal image from its fragmental features and recognize it based on the image, without observing it all over. In this article, the authors propose a neural network model which constructs internal representation of a given figure and recognizes it based on the representation.
The model consists of a sensory unit, a figure representation field and a recognition layer. The model moves the sensory unit and observes a local feature of the figure. The figure representation field consists of many “feature units”, whose activation represents the figure. These units interact through mutual connections which reflect topological structure of the external world. Through the interaction, the field constructs global representation of the figure from the local features observed by the sensory unit.
The recognition layer makes a recognition output using the field activation. It means that the recognition is based not on sensory information itself but on internal representation modulated by the field dynamics. This “internal-representation-based recognition” bears an aspect of human recognition. By imposing external input reflecting “context” to the field, for instance, “contextdependent recognition” is realized.
In addition, the authors propose an active searching algorithm which helps representation construction to advance smoothly. In the course of construction, competition often arises between feature units, especially around edges or corners of the figure, which have significant meaning for constructing the representation. The model can observe “intentionally” such essential parts by moving the sensory unit to the corresponding position. Such competition-invoked searching is a new algorithm for active perception.
The model behavior is investigated through computer simulation. Some related matters on representation in the brain are discussed.