Marcílio Carlos Pereira de Souto,  Paulo J. L. Adeodato. 

It is well known that, in a broad sense, recurrent neural networks are equivalent to Turing machines. However, in general, such a computational power has not been achieved by the current learning algorithms. In this paper, the learning capability of the existing algorithms for sequential {RAM-based} neural networks is analysed. These learning algorithms will be proved to have limitations which prevent the networks from attaining their computability.

http://csdl.computer.org/comp/proceedings/sbrn/1998/8629/00/86290020abs.htm

Processando consulta.