Computational properties of the cerebral cortex to learn sensorimotor programs

We develop a neural network model of the neuronal circuits of the cerebral cortex that can learn different functions known to be learned by different parts of the cerebral cortex: visual integration for invariant pattern recognition, performed by temporal areas; visual to motor transformation for 3D arm reaching movements, performed by parietal and motor areas; temporal integration and storage of sensory-motor programs, performed by frontal areas. We are gradually imposing constraints to this network model to make it consistent with the maximal number of experimental data on on the different part of the cerebral cortex by using a minimal number of computational principles: the network must learn these different tasks in a natural way, using available sensory, motor and internal signals, and perform after learning the appropriate sensorimotor transformations in relation with the occurrence of external events; the processing units in the network should communicate via connections similar to those between and within the corresponding cortical areas, with similar sensory and motor inputs and outputs; during and after learning of these tasks, processing units should display tuning properties similar to those revealed by neuronal recording experiments in the cerebral cortex.