Shift invariant neural net for machine vision

Abstract
A multi-layer network is described which is able to recognise simple shapes in a shift, size, and rotation invariant manner. The use of layers of units to smooth and then to shift the image eliminates the need for the very large numbers of cells which are often proposed in shift invariant networks. The network was trained using back-propagation and is not intended to be plausible as a model of biological vision at the level of cell and connection detail. Some interesting parallels with human vision are noted in the emergent behaviour of the network.

This publication has 3 references indexed in Scilit: