Deep autoassociative networks

Autoassociative networks are powerful and versatile information processing systems with some inconvenient limitations. Two of these limitations are the small number of patterns that can be clearly distinguished and the impoverished ability of the net to form clearly defined neighborhoods around target patterns. This paper introduces a refinement of Autoassociative networks, called 'deep' autoassociative networks. These novel networks can distinguish between large numbers of patterns, and have clearly defined basins of attraction around target patterns. The typical performance of deep autoassociative networks is significantly superior to the typical performance of standard autoassociative networks on the same tasks. Deep autoassociative networks pay for this increased level of performance by having an increased number of weights.