This paper aims at providing insights onto fundamental properties of two o f the earlier models of matrix memories: non-linear, or Willshaw's matrix model, and the linear matrix memory. These are traditional models of Arti cial Neural Networks (ANNs) on which little has been done in recent y ears on themes such as retrieval and storage properties, which are studied in this paper. Equations for predicting network connectivity and storage capacity as a function of the activity levels at the input

more »

... and ouput elds and network dimension are presented for Willshaw's model. For the linear model, an equation for relating input and output network errors is presented. It is shown that error correction for low storage load happens when l 2 2 < l , w h e r e l is the input noise. The theoretical results of this paper t very well with computer simulations. Antônio de P adua Braga would like t o a c knowledge the support from CNPq. On the information storage of associative m a t r i x memories Abstract This paper aims at providing insights onto fundamental properties of two o f the earlier models of matrix memories: non-linear, or Willshaw's matrix model, and the linear matrix memory. These are traditional models of Arti cial Neural Networks (ANNs) on which little has been done in recent y ears on themes such as retrieval and storage properties, which are studied in this paper. Equations for predicting network connectivity and storage capacity as a function of the activity levels at the input and ouput elds and network dimension are presented for Willshaw's model. For the linear model, an equation for relating input and output network errors is presented. It is shown that error correction for low storage load happens when l 2 2 < l , w h e r e l is the input noise. The theoretical results of this paper t very well with computer simulations.