Figure 2. Example of the non-linear and the linear generative model.

Suppose all hidden units are zero except of units and . In this case the patch is generated using the basis functions and . If the two basis functions have the form as displayed on the right-hand-side, the non-linear and the linear model generate the patches on the left-hand-side. Given a pixel , the non-linear model chooses the basis function with the maximal absolute value () to set the value of the patch, . For the example pixel (red box), is chosen but for other pixels may be chosen. Note that the -superposition models the exclusiveness of occlusion without considering object or edge depths. The linear model always sums the basis function values, . After the generation of the noiseless patches both models assume the addition of Gaussian noise for the generation of patches (see Eqns. 2 and 4 but not shown in the figure). The color scale is defined as in Fig. 1.