The Adaptive Weight Using RAM

This article analyses the saturation problem of a RAM neural network, a n-tuple classifier [1] containing 340 12-input neurons applied to the character recognition task, using the British mail data bank [2]. It presents data to evaluate this problem and correlates it to other characteristics of the RAM nets [3] [4]. Therefore, two novel approaches were suggested to reduce the network saturation and improve the recognition level: the Filtered RAM and the Adaptive Weight Using RAM (AWURAM). The first version simply multiplies each input vector by a digital filter during the training and the recall phases. The second approach associates the weight concept to the network in order to distinguish different regions among the trained classes.