Autonomous Situation Understanding and Self-Referential Learning of Situation Representations in a Brain-Inspired Architecture

Making sense of a scene has been considered a problem of sensory analysis traditionally. Prediction is used to cope with combinatorial explosion of possible alternative interpretations of the sensory signals. However, high complexity and variability of natural scenes limit the use of sensory appearance-based prediction dramatically. Brains of living beings seem to use a different strategy. Evolution discovered the power of storing an episode of successful behavior and re-using this memorized experience in similar situations. Such episodes consisting of intended behavior, its outcome, the spatial context, and relevant objects constitute situation models which control the selective inspection of sensory input required for its smooth execution. We argue that this behavior-based approach enables dynamically composed situation models that make the world more regular than it is indeed, and that can be learned autonomously.