The back propagation training algorithm, used to train non-linear feed forward multi-layer artificial neural networks, is capable of estimating the error present in the data presented to a network. While of no use during the training of a network, such information can be useful after training to permit the input data to be itself adjusted to better fit the internal model of a trained neural network. After this has been done, the difference between the modified and original data can be useful. This paper discusses how such data adjusting may be done, demonstrates the results for two simple data sets and suggests some uses that may be made of such differences.
Journal of Computational and Theoretical Nanoscience,
Vol. 7, no. 5 (May 2010), pp. 862-867