As we have already mentioned, the main advantage of the indirect encoding employed by the HyperNEAT algorithm is the ability to encode the topology of the large-scale ANN. In this section, we will describe an experiment that can be used to test the capacity of the HyperNEAT method to train a large-scale ANN. Visual pattern recognition tasks typically require large ANNs as detectors due to the high dimensionality of the input data (the image height multiplied by the image width). In this chapter, we consider a variation of this family of computer science problems called visual discrimination tasks.
The task of visual discrimination is to distinguish a large object from a small object in a two-dimensional visual space, regardless of their positions in the visual space and their positions relative to each other. The visual discrimination task...