Self-organization of predictive representations

We propose an approach for the development of dynamic representations which are predictive for future sensory inputs. The prediction error allows one to restructure both internal and input connectivity such that, from the initially unstable dynamics of a random network, a reliable behavior is obtained after learning. In particular, we consider the self-organization of connectivities similar to synfire chains (for linear sequences of inputs) or effectively two-dimensional neural layers (for data from an autonomous robot in a maze)