Neural networks for artificial intelligence?

In recent years there has been a lot of research into artificial neural networks, which offer a number of potential advantages over conventional artificial intelligence methods. Neural networks can easily be trained, they fail `gracefully´ and they are more amenable to implementation in VLSI. On the other hand, they suffer from a number of limitations which must be overcome if they are ever to be of widespread use: their capacity for generalisation is often poor; training time increases rapidly with the size of the network and it is often difficult to understand the resulting encoding of data. The authors suggest some possible directions for future research to overcome these problems, and present the results of some experiments which show that training time may be reduced by structuring the training of such networks