Since the nodes in the output layer are trained by the output from the middle layer, what they learn to respond is the internal patterns formed by the middle-layer nodes, when they are in turn responding to the motion patterns presented to the input layer. After the training of the middle layer is complete, an input pattern will always stimulate the same set of nodes of the middle layer, one winner from each group. These nodes form a binary internal pattern composed of 1's representing the winning nodes and 0's representing the others. The winning nodes may each represent a different area of COM locations, (as each group partitions the visual field randomly differently from others), but these areas must have an intersection subarea covering the COM of the input pattern to which these nodes all responded. Thus a one-to-one correspondence relationship can be established between each subarea in the visual field and a unique internal pattern. Trained on these internal patterns, the MST nodes in the output layer learn to recognize them. An input motion pattern will stimulate an output node only if the COM of the input is located in the subarea corresponding to the internal pattern to which the output node responds. Compared to the middle-layer nodes, the selectivities of the output nodes are more sharply tuned because they only respond to few motion patterns whose COMs are located in a very small subarea.
We can further see that the more groups there are in the middle layer, the more randomly different ways the visual field will be partitioned into areas, the smaller their intersections become, and the more precisely the COM location can be detected by the output nodes. If there are enough groups in the middle layer, it is possible that some intersection subareas become so small that they may each cover only a single COM location. Moreover, for the same reason discussed above, we see that the tuning of the response selectivity for COM locations can be further sharpened by adding even more layers.