Multilayered Feedforward Neural Network (MLFNN) Architecture as Bidirectional Associative Memory (BAM) for Pattern Storage and Recall

Abstract

Between two popular ANN architectures – feedforward and feedback, also known as recurrent – feedback architectures have been extensively used for memorization and recall task due to their feedback connections. Due to the inherent simpler dynamics of FNNs, these structures have been explored in the present work for association task. Variation of standard BP algorithm, two-phase BP algorithm has been proposed for training MLFNNs to behave as associative memory. The results thus collected show that with the proposed algorithm, MLFNN start behaving as associative memory and the recall capability for corrupted versions of the stored patterns is at par with BAM but with lesser time.