Associative Memory with Hidden Units by a Hybrid Learning

Abstract

In the conventional neural associative memories, auto-correlation (Hebbian) learning has often been used. This paper proposes an associative memory model consisting of visible units and hidden units. The connections among the visible units are determined by auto-correlation learning, and the connections between the visible units and hidden units are determined by the error back-propagation learning. These two kinds of learning constitiute a hybrid unsupervised learning. Experiments show that our associative memory based on the hybrid learning has larger basins of attraction and less spurious memories than the existing models.