Coordinate-free self-organising feature maps

The successful application of a new strategy for classifying images of biological macromolecules, and resolving their rotational orientations, was recently introduced by R. Marabini and J.M. Carazo [Pattern recognition and classification of images of biological macromolocules using artificial neural networks, Biophys. J. 66 (1994) 1801–1814]. Their work was based on Kohonenʹs self-organising feature map (SOFM) defined on a plane, and has been extended here by allowing an SOFM to operate independently of topology. An SOFM has been constructed which follows instructions according to the current values of a variable, which alone drive the self-organising process. The instructions that the SOFM follows are only available internally to the map and so the behaviour of the SOFM must be supervised by providing suggestions as to what the state of its components should be. The method is shown to be useful in identification and clustering of recurring motifs, of resolving metastable states in which the process can occasionally become trapped, and in discarding data unsuitable for further analysis.