Here we will concentrate on modeling the pattern MT cells, as the component MT cells are similar to the motion selective V1 cells which have already been considered in previous section.
Many different models proposed for detecting global (pattern) motion ([16], [17], [22], [23], [42], [43]) are based on either the correlation method or the spatiotemporal method. Most of the models follow the basic idea that motion processing is carried out in two stages. The first stage measures the local motion, presumably in the V1 area, and the second stage integrates the local motion information and computes the global motion, presumably in the MT area. These models also make different assumptions about the types of information available as the input. As discussed previously, many different types of information are processed and extracted in the V1 area, such as direction, speed, spatial and temporal frequency, etc. Models for the motion processing in the MT area can be classified according to the types of information assumed to be available from V1. Here we will concern ourselves only with network models based on speed and direction information as the input.
Two such models are proposed in [24] and [25], both of which use a two-layer network with its input layer units simulating the V1 neurons (or the component MT cells) responsive to local component motions and the output layer units simulating the pattern MT cells responsive to global motions. The model in [24] uses supervised Widrow-Hoff learning, and the one in [25] uses unsupervised Hebb learning, which is obviously more biologically plausible than the supervised network. However, they have no mechanism to account for other MT properties such as the suppression under transparent condition or the different direction selectivities under different perceptual transparency of the grating patterns. In the following, a multi-level network model is presented to account for as many known properties of the MT neurons as possible.