Previous | Session 139 | Next | Author Index | Block Schedule
M. Ho (NASA Summer High School Apprenticeship Research Program), D.M. McIntosh, A.N. Srivastava (NASA Ames)
Artificial neural networks (ANNs) serve to process, learn, and predict information using layers of interconnected computational units. The quality of the network’s performance on such cases depends on its generalization ability, or the ability to recognize trends from the training data and employ what it has learned to make predictions on new test data.
Nonetheless, ANNs often perform poorly when applied to new cases dissimilar to those they have encountered, a flaw possibly attributed to data anomalies that adversely affect the training process. Therefore, it is important to develop methods to improve a neural network's generalization ability, since the quality of future predictions on a comprehensive set of all possible data is the ultimate determinant of a network's proficiency.
In this study we use the Sloan Digital Sky Survey data to ascertain whether the network’s performance error was distributed throughout the data or concentrated in a single area. We isolate inputs and distinguish combinations of independent inputs that most significantly contribute to the overall error. We also investigate the consequences of training a network with a data set containing outliers and the effects of eliminating outliers on the error value generated by the testing process. Moreover, we examine the relationship between the size of a Voronoi cell and the magnitude of the disparity between the correct output (known distances) and predicted output of outliers. Given that a substantial correlation exists, we can utilize Voronoi tessellation to “warn” a network to detect outliers based on the high volume and low distribution of their Voronoi cells and to train carefully on such points such that the true generalization of the data is maintained.
Previous | Session 139 | Next
Bulletin of the American Astronomical Society, 37 #4
© 2005. The American Astronomical Soceity.