The results from the test data sets above indicate that the DBNN classifier works as well as the widely used SExtractor software. All instances of DBNN failure correspond to objects that do not have a counterpart in the training set or objects that are difficult (but not impossible) to classify visually. In both test sets, there is no instance where an obvious misclassification has occurred.
As mentioned before, in addition to the consistency, the increase in speed of the training process is very significant here. Our training procedure for DBNN on the 402 objects in the training dataset took 0.23 s on an Intel Pentium III processor running at a clock speed of 700 MHz. Such short training times are invaluable when one has to optimally deal with large datasets that are collected and processed over a significantly wide span of time, demanding repeated retraining of the classifier to account for variations in observing conditions and use of newer and better parameters for classification. Data from large surveys fall into this category. Also, unlike the back propagation neural network (BPNN), since DBNN is based on Bayesian probability estimates, it is immune to diverse training vectors that often appear in the training set due to noise in the observation. This means that the performance of the network is likely to be significantly better than the BPNN beyond the completeness limit.
In this paper we have illustrated the power of the technique by applying it to the star galaxy classification problem. The technique can easily be applied to all classification problems that currently employ BPNN. For example, by using the large number of photometric and spectroscopic parameters measured (for millions of objects) by surveys such as the the Sloan Digital Sky Survey, it will be possible to apply the DBNN technique to identify interesting samples for study in the vast, largely unexplored parameter space. We are in the process of enhancing DBNN to solve problems that involve regression.
The source code, documentation and the training and test set images described in this paper may be downloaded from the URL: http://www.iucaa.ernet.in/~nspp/dbnn.html
Acknowledgements
The authors would like to thank Ashish Mahabal for discussions and a careful reading of the manuscript. The first author would like to express his sincere thanks to Inter University Center for Astronomy and Astrophysics and the computer staff there for providing him all the required facilities for the successful completion of this project. We also thank the anonymous referee whose detailed comments considerably improved this paper.
This work made use of images and data products provided by the NOAO Deep Wide-Field Survey (Jannuzi & Dey 1999), which is supported by the National Optical Astronomy Observatory (NOAO). NOAO is operated by AURA, Inc., under a cooperative agreement with the National Science Foundation.
Copyright ESO 2002