Network synthesis and generalization properties of artificial neural net using Fahlman and Lebiere´s learning algorithm

In designing neural network systems, it is desirable to use already-trained networks, each performing a specific task, to design a system that performs a global or extended task without destroying the information gained by the previously trained nets. This can be done by synthesizing the trained networks or adding new output layer units in the case of incremental learning by incorporating new hidden units to acquire additional information required to realize the newly defined task. Fahlman and Lebiere´s (FL) learning algorithm is particularly suitable for this purpose. It is shown that network synthesis and incremental learning can be by an FL algorithm and a backpropagation (BP) algorithm. Investigation shows that the synthesized or expanded FL networks have generalization ability superior to BP networks