Recognition of topological features of graphs and images in neural networks
Reiner Kree, Alfred Zippelius · Journal of Physics A Mathematical and General · 1988
The authors extend the architecture of the Hopfield network, such that it can recognise transformed versions of a set of learnt prototypes. As an example they construct a network which can generalise over all topologically equivalent representations of graphs or images. The construction is based on two coupled networks: a Hopfield network to store and retrieve patterns and a preprocessor to transform the input data.