Towards Computational Neural Systems Through

摘要: A major consideration when using artificial neural networks (ANNs) is the design procedure. This includes the architectural decisions (eg number of neurons, number of layers, connectivity) and the selection of a learning algorithm. A number of approaches have been reported to automate the design process. In classical ANNs there are examples of constructive algorithms whereby ANN architectures are incrementally created [Fahlman and Lebiere, 1990] together with the opposite approach of pruning networks which are initially too large (see [Reed, 1993] for a survey). Another more fruitful approach has been to combine evolutionary computation techniques and ANNs (see [Yao, 1993]).Although the above design algorithms have produced interesting results, they are characterised amongst other things by producing ANNs with stereotyped architecture. In order to approach the computational complexity of biological …