next up previous
Next: The output layer Up: The Models Previous: The input layer

The middle layer

The middle layer is composed of a set of groups of 30 nodes simulating the MST neurons. All MST nodes in a group are fully connected to the MT nodes in the input layer. As a result, the receptive fields of the MST nodes are much larger than that of the input MT nodes. The number of MT nodes each MST node is connected to is so chosen that it will properly relate the size of the receptive field of the MT neurons and that of the MST neurons. As MST nodes in different groups do not compete with each other during learning, there can be an arbitrary number of independent groups in the middle layer without affecting the learning process.

The MST nodes will be trained to respond selectively to the different optic flow patterns presented to the input layer of the network, such as translational motions of different directions, and rotational, spiral and radial motions of different directions and center of motion (COM) locations. However, it is important to note that these motion patterns do not form a set of separable clusters in the feature space, a favorable situation for competitive learning. Different types of motion patterns may share some input nodes, and the boundaries between them are further blurred due to the existence of various types of noise. In other words, all input patterns of different types and COM locations form a continuum in the feature space. Same as in the MT model, the balancing technique described in Appendix is again needed to ensure the input patterns are relatively evenly responded to by all the middle layer nodes. The continuum may be clustered in randomly different ways into different number of clusters of different sizes. It is possible that two or even three types of motions are classified into one cluster and represented by the same MST node because they share quite some nodes in the input layer. It is this learning mechanism that enables the network model to simulate many important responsive features of the MST neurons, such as the multi-component neurons and their different position-dependent responses.

After the middle layer is trained, each node in a particular group will respond to, and become the winner of, an input motion pattern if its COM is located in a local area. This response selectivity for the COM location is not very sharply tuned, and the precise COM location of a motion pattern cannot be determined from the responses of the nodes in a single group. However, the COM location tuning will be sharpened by adding the output layer of MST nodes.


next up previous
Next: The output layer Up: The Models Previous: The input layer
Ruye Wang
2000-04-25