Coupling of evolution and learning to optimize a hierarchical object recognition model

Abstract

A key problem in designing artificial neural networks for visual object recognition tasks is the proper choice of the network architecture. Evolutionary optimization methods can help to solve this problem. In this work we compare different evolutionary optimization approaches for a biologically inspired neural vision system: Direct coding versus a biologically more plausible indirect coding using unsupervised local learning. A comparison to state-of-the-art recognition approaches shows the competitiveness of our approach. © Springer-Verlag 2004.