SEMINT1 - Selbstorganisation aktivierbarer dynamischer Raum- und Formrepräsentation durch SEnso Motorische INTeraktion


The aim of this project is to further develop and concretise an alternative approach to visual scene interpretation and to prove its applicability to problems of controlling mobile, autonomously operating robots.

The focus is on biologically inspired or motivatable mechanisms with which an internal interpretation of space and form can be developed from the sensorimotor interaction with the environment.

In this context, the recognition process is viewed as an internal simulation of a set of one's own actions and prediction of their consequences in the given situation. The actions characterise the sensory situations in which the system finds itself. At the same time, from the many descriptive actions, those can be selected for execution that have a positive effect according to the system's objective function.

A prerequisite for the prediction of action consequences is that the system has an internal model of the relationships between different data streams. According to our model, information from sensory sources (e.g. visual, tactile) and information about the internal system state (e.g. movements executed at the moment) are combined in the associative cortex after several stages of complex unimodal processing. There, the frequent simultaneous or fixed temporal relation activation of neurons is detected. Such fixed relations point to regularities of the physical world that are incorporated into the " world model ".

Hypothesising can be understood as the application of detected coincidences to real data or other hypotheses. Its mechanism is the coupling or feedback between different or the same data streams via detected coincidences in the form of adaptive weightings.