A neural network model of frontal cortical circuits for learning conditional sequences

The authors propose a neural network model for learning and producing conditional sensorimotor sequences. This model is based on several computational principles related to the properties of the frontal areas of the cerebral cortex, which are critically involved in time processing. Two types of automata are defined: associative units perform sigma-pi transformations of the inputs, while frontal units have a bistable state of activity which allows correlation of two successive events occurring with variable delays. These two types of unit cooperate in a network which is able to learn sequences of sensory or motor events with conditional reinforcements. Activation and learning rules for controlling the bistable state of activity are provided. It is shown how sequences of arbitrary length can be learned according to two recursive learning processes (sub-sequence and bifurcation). These processes allow creation of dynamical representation of sequences in subsets of frontal and associative units