Algorithm programmed by a neural network model for coordinate transformation

A neural-network model has been trained for coordinate transformations simulating the action of parietal area 7a. Examination of this network reveals that its planar responses extract the vector coordinates of the angles represented by the inputs, in order to add them together. The hidden unit responses were most planar when the number of units was minimal; the corresponding area of the brain evidently has several times as many hidden units for the task being performed. The cortical area may be using the same algorithm as the neural-network model, so examination of the model clarifies the function of the neurons