Martin Heracles, Alexander Gepperth, Jannik Fritsch, Christian Goerick, "Integrating representations for learning higher-order correlations in a brain-inspired cognitive framework", 5th HRI Global Workshop, 2008.

An important biological mechanism for learning in the human brain is Hebbian synaptic plasticity. However, Hebbian learning is restricted to direct synaptic connections between two neural populations, hence cannot directly account for learning that involves multiple neural populations. In order to overcome this limitation, we propose to combine Hebbian learning with self-organizing maps. More specifically, multiple lower-level representations are mapped onto selforganizing maps in an unsupervised process, thus forming higher-order representations of combinations of the lower-level representations. This way, learning that involves multiple neural populations can be reduced to learning that involves only two (higher-order) neural maps, which makes Hebbian learning applicable again. In order to demonstrate the validity of our approach, we consider a simple case of learning that involves three lower-level representations, namely, learning the interrelationship between the 2D size of objects in monocular camera images, their object class as obtained by an object classifier, and their 3D distance to the camera. We present a technical system instance of the proposed system architecture that is able to learn this interrelationship at run-time and, after learning has finished, to predict the 3D distance of objects from their 2D size and their object class. Experiments using both simulated and real world image streams confirm the quality of these predictions and support our hypothesis that the proposed combination of Hebbian learning with self-organizing maps can serve as a generic mechanism for system-wide unsupervised learning.