The convergent learning of three-layer artificial neural networksfor any binary-to-binary mapping

摘要

In this paper, the learning algorithm called expand-and-truncatenlearning (ETL) is proposed to train three-layer binary neural networksn(BNN) with guaranteed convergence for any binary-to-binary mapping. Thenmost significant contribution of this paper is the development ofnlearning algorithm for three-layer BNN which guarantees the convergence,nautomatically determining a required number of neurons in the hiddennlayer. Furthermore, the learning speed of the proposed ETL algorithm isnmuch faster than that of backpropagation learning algorithm in a binarynfield. Neurons in the proposed BNN employ a hard-limiter activationnfunction, only integer weights and integer thresholds. Therefore, thisnwill greatly facilitate actual hardware implementation of the proposednBNN using currently available digital VLSI technology