Evolutionary feature design for object recognition with hierarchical networks

A major problem in designing neural vision models is the large dimensionality of the search space for defining the needed networks. By using hierarchical vision models inspired by biology we narrow the space of possible architectures. We perform evolutionary optimization of remaining critical network parts e.g. the combination features, which are up to now mostly subject to manually determination. We show that the evolutionary approach leads to an optimized recognition system with respect to speed and performance, which is highly competitive with other state of the art systems.