Pattern recovery in networks of recursive processing elements with continuous learning

This paper addresses a continuous learning method using associative memories based on recursive processing elements (RPEs). In order to decide if a pattern recovered by the associative RPEs is known or unknown, we are using two discriminators: a network stabilization criterion and a Hamming distance criterion. The network stabilization criterion is based on the disagreement between the current and the next state, and the Hamming distance criterion checks the number of bits flipped between the prompting pattern and the recovered pattern. Experiments for the performance of continuous learning when the prompting patterns are exposed to digital noise and experiments for the evaluation of the capacity of network storage are presented and analyzed.