An algorithm for dynamically adapting neural network topologies

Recently, it has been proposed that biological networks change not only the synaptic strengths of connection but also partially their internal topologies, according to either the received external stimuli or the pre-existent connection layouts. Following this idea, a method is presented to dynamically adapt the topology of neural networks with supervised learning, using only the information of the training set. The method eliminates connections from an initial fully connected network, concurrently with the learning algorithm. Several experimental results obtained with multilayered networks are also reported to demonstrate the capabilities of the proposed method