Skip to Main Content
When training a neural network it is tempting to experiment with architectures until a low total error is achieved. The danger in doing so is the creation of a network that loses generality by over-learning the training data; lower total error does not necessarily translate into a low total error in validation. The resulting network may keenly detect the samples used to train it, without being able to detect subtle variations in new data. In this paper, a method is presented for choosing the best neural network architecture for a given data set based on observation of its accuracy, precision, and mean square error. The method, based on , relies on k-fold cross validation to evaluate each network architecture k times to improve the reliability of the choice of the optimal architecture. The need for four separate divisions of the data set is demonstrated (testing, training, and validation, as normal, and an comparison set). Instead of measuring simply the total error the resulting discrete measures of accuracy, precision, false positive, and false negative are used. This method is then applied to the problem of locating cryptographic algorithms in compiled object code for two different CPU architectures to demonstrate the suitability of the method.
Date of Conference: 18-23 July 2010