Exact learning and default-rule governed behaviour

We have modeled “exact” and “regularized” learning in artificial neural networks (ANNs), which can be trained to reproduce the Markovian state transition matrix of a time sequence. We consider that a “quasi-regular” mapping corresponds to a sequence in which transition rules of widely different orders coexist. To train the network a cost function is minimized that counts the number of times that each rule is violated in a sufficiently long string. “Generalization” is checked comparing the sequences generated during training with the target one. We find that for all realistic situations the ANN rapidly convergences to a “default rule”. The default rule governed behaviour appears within the present model as a consequence of the special training protocol and the structure of the synaptic phase space.