Title :
What Elman networks cannot do
Author :
Haselsteiner, Ernst
Author_Institution :
Dept. of Med. Inf., Tech. Univ. Graz, Austria
Abstract :
Finds out properties of Elman networks which are inherent to the architecture. Using methods from control theory a representation for the network is developed, which shows all the capabilities of the network in a very compact form. With this new kind of representation the impact of the different weights is analyzed in detail. A very simple learning task, which seems unsolvable with a given Elman network is analyzed and the developed representation is used to prove that the task cannot be learned. The results of the simple learning task are generalized to networks of any size and for certain learning tasks a lower boundary for the minimum number of hidden units is given
Keywords :
learning (artificial intelligence); neural net architecture; recurrent neural nets; Elman networks; hidden units; lower boundary; simple learning task; Backpropagation; Computer networks; Constraint theory; Control theory; Delay effects; Neurons;
Conference_Titel :
Neural Networks Proceedings, 1998. IEEE World Congress on Computational Intelligence. The 1998 IEEE International Joint Conference on
Conference_Location :
Anchorage, AK
Print_ISBN :
0-7803-4859-1
DOI :
10.1109/IJCNN.1998.685952
