Fig. 7

Possible regions for MLPs (from [76]); given the extreme complexity of regions in most ANN applications, many layers are needed (three are shown here). The challenge of partitioning for complex spaces is illustrated for simple cases of handling targets that distribute in non-convex ways. The middle column poses a case of two classes (A, B) that are not contiguous, and the next column theorizes a case where classes A and B may have more arbitrary shapes, blocking simple classification. Possible class regions are shown and shaded or not