Internet guide to the galaxies navigates on neural networks

February 18, 2000

Tony Durham delves deep into

astronomical literature on the web with the help of an ingenious method of library classification based on the brain

Astronomy is about space. So perhaps it is quite natural for astronomers to use maps to find their way around their subject's literature.

On a Strasbourg University website there is a map of 8,039 articles from the journal Astronomy and Astrophysics, clustered according to their subject matter.

At top left is a body of literature about galaxies, and within it, clusters of papers on such topics as galaxy jets and Seyfert galaxies. Lower down, papers on neutron stars form a dense cluster.

Nearby is a cluster of papers on novae. In the same region of the map, other stellar topics can be found. It all looks carefully planned.

A map of 13,000 papers from the Astrophysical Journal has its own quite distinct topography. Its hottest spot is a cluster of papers on gamma ray bursts.

But these clusters are not created artificially by a librarian classifying papers according to a pre-existing scheme. They emerge spontaneously from a computing process in which simulated neurons learn to respond to specific patterns of keywords.

Self-organising maps, as they are called, were devised by the Finnish physicist Teuvo Kohonen, who applied them to a number of problems, including the recognition of Finnish and Japanese speech sounds. He thought that the cerebral cortex, which is a large, deeply folded sheet of tissue, might employ two-dimensional maps when it tackles similar tasks.

The maps of the astronomical literature were created by Philippe Poincot, who recently received his doctorate from Strasbourg's Universite Louis Pasteur. Besides mapping two leading astronomical journals, he has made a map of 2,468 catalogues that between them cover galaxies, quasars, star clusters, radio and X-ray sources, many kinds of stars, asteroids and other celestial objects.

Web visitors can delve deeper into whatever subject interests them by clicking on the appropriate area of any of the maps. They can read abstracts of journal articles, interrogate catalogues or zoom in on a subject area and see a more detailed local map.

During the training process, each neuron progressively adjusts the weightings it gives to several hundred keywords. It becomes an expert recogniser of documents belonging to a tightly defined topic area.

Neurons are dotted over the map in a grid pattern and as training proceeds, each neuron shares a little of its learning with its neighbours.

The result of this cooperation is that neighbouring neurons learn to respond to similar patterns of keywords.

If there is a neuron with an interest in close binary systems, and another nearby which has developed an interest in neutron stars, then there is probably a connection between the two subjects. In this case the likely explanation is that close binary systems, where one partner is a neutron star, have been intensely studied by X-ray astronomers.

Poincot's supervisor was Fionn Murtagh of Queen's University, Belfast, who holds an adjunct post at Strasbourg. With partners in France, Germany, Greece and the United States, Professor Murtagh is about to begin work on an European Union 5th Framework project on navigation in information space. Information maps may be coming to your subject soon.

Please login or register to read this article

Register to continue

Get a month's unlimited access to THE content online. Just register and complete your career summary.

Registration is free and only takes a moment. Once registered you can read a total of 3 articles each month, plus:

  • Sign up for the editor's highlights
  • Receive World University Rankings news first
  • Get job alerts, shortlist jobs and save job searches
  • Participate in reader discussions and post comments

Have your say

Log in or register to post comments