A computational model of proprioceptive maps

The authors developed a computational model of proprioceptive somatosensory cortex. A model arm controlled by three pairs of muscle groups moves in a 3D space. The muscle length end tension determined by randomly generated neuronal input to the muscles was presented to the network as a proprioceptive input. The network, consisting of two layers of units, with arm layer of 12 length and tension inputs and SI layer of 20/spl times/20 laterally connected units, distributed activation competitively. Trained with a variant of Hebb-type learning, the network developed feature maps of muscle length and tension, and a spatial map of hand position by encoding the interrelationships between muscle lengths. Thus, the network learned to capture the mechanical constraints of the model arm. These results can be viewed as testable predictions for future experimental studies.