An algorithm for self-structuring neural net classifiers

A new algorithm for self-structuring neural net classifiers is presented. It is called EMANN for Evolving Modular Architecture for Neural Networks. The basic idea is to increase the biological likelihood by extensively using internal local variables instead of external global variables to evolve the structure. We do believe that such alternative can be profitable improving the accuracy of the resulting classifier while maintaining the neural architecture to a minimal size. We introduce the “connection strength” of a neuron as the key internal local variable used to increment the structure and we show how this variable reflects the neuron behavior. Some heuristics we follow for the network building are also presented. Finally experimental results for two classification tasks are presented