Developing Neural Networks through Growth

In the paper a new method to develop neural networks is proposed. The method is based on the idea of growth. Usually neural networks have been designed to be a large number of neurons with regular interconnections and adaptation to a specific problem has been implemented through sophisticated learning algorithms. In the proposed approach the adaptation to a specific problem has been implemented partly during the development of the structure of neural network. The growth starts from an initial value and the development of neural network is controlled through a set of production rules. The mechanism has a strong similarity to the nature growing process. The used control mechanism is based on Lindenmayer systems, that have been used successfully for modeling life of plants and bacteria. The growth control of neural networks is only one type of adaptation in a larger project to simulate more generally adaptive behavior. 1 Introduction Artificial neural networks can be divided to general t..