Generation of Neural Network Models of the Brain by Evolutionary Computation

Brain science has been ceaselessly progressing by propulsive forces of several research methodologies. Obviously one of the most important approaches is modeling. Especially neural network models have been giving explanations of information processing in the neural system, which can consistently unify a large number of experimental findings on the regions of electrophysiology, fMRI, psychophysics, and so on. One of the important factors of the conventional modeling approach is that the models had constructed by human modeling researchers, and the principles of the models had been inspired due to the instincts and experiences of the modelers. Especially there has been implicit agreement that model should be represented concisely. On the other hand, the scale of the actual brain is very huge, and the network structures and connection pattern between neurons are quite complex. Accordingly, it might be happen that the models constructed by human researchers fail to account the actual principles of information processing in the brain. The modeling approach should overcome the restriction of modeler’s inspiration in order to successfully catch the real nature of the brain.

This study adopts a new method to construct the neural network models utilizing the framework of evolutionary computation. The neural networks are generated automatically by computer, instead of human modelers (i.e. the authors). We use genetic algorithm (GA) in order to determine the network structure, with some constrains as for the global architecture revealed by physiology and anatomy. In our framework, each gene on the GA computation corresponds to a neural network, and a lot of neural networks coexist in a search space at the same time. Fitness function is defined as the behavioral similarity of each neural network and actual characteristics of the brain. We adopt this method to generate a neural network model of the visual system. We currently focus on the visual functions of figure/ground separation, and pattern recognition.

There are some preceding studies about the hybrid method of neural network and GA. However, the purposes of almost all of them were to offer multipurpose adaptive method applicable for any engineering scene. In contrast, we focus on the method of automatic generation of models for the purpose of reducing arbitrariness of the human researchers’ inspiration.