Design and implementation of modular neural networks based on the ALOPEX algorithm [face recognition application]

The hypothesis of the current work is that a modular neural network provides a better recognition of templates containing missing features, as opposed to a conventional multi-layer neural network. The authors designed and implemented an algorithm which allows one to create neural networks containing multiple modules, as well as interneuronal lateral inhibition. The training stage of the network in this project is based on the ALOPEX algorithm. The hierarchy of the network provides independence of its functions on the number of modules in the network, number of layers within each module, number of nodes within each layer, and finally, number of connections of each node. Three sets of tests were performed on a modular network, and on a control network, using the same optimization ALOPEX algorithm. In the tests the authors used five, six, and seven templates respectively. The results show that the modular approach gives a more accurate recognition of templates containing missing features, which proves the authors´ initial hypothesis. One possible application of this approach is in face recognition, when certain parts of a face image (like nose or eyes) are not available for some images