Previous abstract Next abstract

Session 88 - Structure and Kinematics of Galaxies.
Oral session, Wednesday, January 17
La Condesa, Hilton

[88.05] The Application of Artificial Neural Networks to Astronomical Classification

A. Naim (JHU)

Galaxies are fundamental to the understanding of the structure and evolution of the universe. They contain stars, gas and dust, and serve as an astrophysical laboratory in which physical processes can be examined. In the context of the large scale structure of the universe galaxies can be viewed as test particles. They are bright and therefore visible at very large distances, and also numerous and so can be used to provide reliable statistics. In previous decades the major obstacle to studying the large scale structure of the universe was the relatively sparse data samples, because obtaining large quantities of galaxian images and spectra requires a lot of observing time, and the accumulation of significant data bases was therefore a slow process. This obstacle is in the process of being removed today, with the advent of large-scale surveys (e.g., the APM galaxy survey, the Sloan Digital Sky Survey and the 2 degree Field survey).

The new challenge with which the astronomical community is faced is the management and analysis of the forthcoming extragalactic data bases. On top of the obvious need for better hardware to give large storage volumes and quick access, one needs to devise automated tools for data analysis. The sheer volume of the data renders manual analysis impractical. It would be best if one could somehow transfer the knowledge and expertise accumulated over years of painstaking manual analysis to a machine.

This thesis is part of an effort to achieve this goal. I borrowed techniques that have proved useful in other fields (e.g., engineering) and applied them to astronomical datasets. The major tool I used was Artificial Neural Networks (ANNs), which was originally conceived as a simplified computational model for the brain. The scope of methods and algorithms referred to as ANNs is quite wide. In particular, a distinction is made between Supervised Learning algorithms and Unsupervised methods. The former put the emphasis on ``teaching'' a machine to do the work of a human expert, usually by showing examples for which the true answer is supplied by the expert. These methods are mostly intended to save time and effort in the less complex stages of the analysis. The unsupervised approach is aimed at learning new things from the data, and is most useful when the data cannot easily be plotted in two or three dimensions. This approach uses only an organising criterion in order to arrange the data, but there is no teacher, nor is there a true answer.

The main problem to which I applied ANNs is the morphological classification of galaxies from their digitised images. On the one hand galaxian morphologies appear to be well correlated with the physics of their interiors, while on the other hand they are the most readily available type of extragalactic information to date. Knowing the morphologies of galaxies can help in defining target lists for observations, in the analysis of the morphology-density relation and in calibrating distances by using morphology-specific relations (e.g., the Faber-Jackson relation for ellipticals and the Tully-Fisher relation for spirals). The entire process - data reduction, feature extraction and classification - was automated and, as demonstrated towards the end of this thesis, is readily applicable to very large numbers of galaxies.

The most important implication of this work is the generality of the methods used. This is only a demonstration of the powerful tools now available for astronomical data analysis. In coming years researchers will have to make use of such methods in order to solve more and more problems in Astronomy.

I would like to acknowledge the financial support I received in the form of an Overseas Research Studentship and the Isaac Newton Studentship.

Program listing for Wednesday