Kevin Cherkauer, Jude Shavlik
The choice of an input representation for a neural network can have a profound impact on its accuracy in classifying novel instances. However, neural networks are typically computationally expensive to train, making it difficult to test large numbers of alternative representations. This paper introduces fast quality measures for neural network representations, allowing one to quickly and ac(cid:173) curately estimate which of a collection of possible representations for a problem is the best. We show that our measures for ranking representations are more accurate than a previously published mea(cid:173) sure, based on experiments with three difficult, real-world pattern recognition problems.