Cortical architecture and self-referential control for brain-like processing in artificial neural systems

Abstract

Progress in understanding the way the brain processes information while it is constantly interacting with the sensory environment is hampered by inadequate models and theories. Current models and theories of brain computing are, obviously, still not completely correct when confronted with so-called real-world problems. Sensory recognition and the subsequent selection and optimization of a proper behavior are basically constraint satisfaction problems. Both conventional AI and current formal neural network systems operate with set constraints: the architecture and parameters are defined a priori and then the input data are structured according to these set constraints on the learning process. However, as long as the constraints are set from outside the system (by the programmer, designer), the system has no ability for self-organization. There is the ability for adaptation within these a priori defined limits, but not the ability to include new knowledge into the consistent relational framework of existing knowledge beyond the prespecified constraints. Therefore, self-organization of constraints in complex systems is the key problem for getting self-organization of knowledge representation under real-world conditions. We show that a value system and self-referential control in a modular architecture are crucial prerequisites for both robust recognition of sensory input and the ability to integrate new knowledge into the already acquired knowledge representation. Finally, we outline a philosophy and propose a model approach that is a first step toward implementing those capabilities in artificial neural systems.