Much research has been made no natural networks (NNs) for pattern recognition with promising results. We have proposed a structure reduction method of the NNs, where we adopt slab values. Here, the slab values are the sum of input pixels generated as characteristics of the inputs by masks. The mask covers some parts of the input. The slab values enable us to distinguish the difference between input patterns. Initially, we use random numbers to determine the masks. However, mask optimization has not been elaborated, sufficiently. In this paper, we adopt a genetic algorithm (GA) which is a search procedure based on the mechanism of natural genetics to determine the masks. Adopting the GA to mask determination, we show that useful masks are automatically determined using US dollar bills in the simulation.