Ways to Store Words in Attractor Neural Networks

Patricio Perez; Dino Salinas

Abstract

We present here a comparative study on the abilities of four different neural nets to store words. We test them on the retrieval of three letter words when two letters of a stored word are presented. It is shown that the best results are obtained by coupling nets of the Hopfield type, as many as letters in the word. The synaptic coupling between neurons of different subnets is of order n–1, where n is the number of letters.

Más información

Título de la Revista: ICONIP: International Conference On Neural Information Processing
Volumen: 3
Número: 2
Fecha de publicación: 1994
Página de inicio: 1361
Página final: 1363
Idioma: Inglés