The network structure

The hierarchical structure of the network model is illustrated in Fig. 1 The input layer simulates the V1 neurons which are selective to local velocities. The visual field is covered by an array of $n \times n$ small patches (represented by the 7 by 7 squares in the first layer in Fig. 1a) each containing 32 V1 nodes for 8 directions and 4 speeds (Fig. 1b). The middle layer is composed of $(n-1)/2 \times (n-1)/2$ groups each containing a set of more than 32 component MT nodes, which are all fully connected to all $32 \times 3 \times 3 =288$V1 nodes in a local region in the input layer composed of $3 \times 3$ patches. Two neighboring middle layer groups have an overlap of 3 patches in the input layer (sharing $3 \times 32 = 96$ V1 nodes). The output layer has a set of groups (three in Fig. 1a) each containing more than 32 pattern MT nodes fully connected to all the middle layer nodes. The number of groups in the output layer is arbitrary as they are independent of each other (nodes in different groups do not interact, as there is no lateral connection between two groups). The weights between two consecutive layers will be determined by competitive learning algorithm, as discussed in Appendix.

  

Figure 1: The network model for MT