Position independent recognition with a neural network

A network has been developed which is able to recognise simple shapes in a shift, size, and rotation-independent manner. The use of a relatively coarse network of real valued units limits the number of connections, while a novel `shifting layer´ eliminates the need for mapping units. The net exhibits a number of human-like limitations in its performance, but is not intended to be a `realistic´ model of a biological vision system. The net could be used as a basis for a number of machine vision tasks, but would need to be implemented in hardware to be useful in practice