A new architecture for achieving translational invariant recognition of objects

A multistage network that will reduce the translational uncertainty of a one-dimensional object is presented. To implement this network, novel network structures like multiple-valued outputs, competition between links instead of nodes, and cooperation of signals at the links are used. The number of nodes and links needed to implement the architecture is small. If the input field consists of n cells, then the total number of cells needed is only O(n). The total number of connections needed is O(nlogn). It is shown that size-invariant recognition can also be achieved if the input to the architecture is provided by a scale-sensitive network called a masking field.<
>