Preparng the right data diet for training neural networks

Neural networks are a good way to interrelate nonlinear variables in a robust manner. The simplex method for optimization is not nearly as effectual, and neither are the various statistical methods for classifying and associating data and predicting results. The reason is that neural networks are put through a training phase, during which they can automatically fine-tune themselves as often as proves necessary to get the desired performance. Of course, the old adage "garbage in...garbage out" applies as much to neural networks as it does to all other data-processing applications. If the training data set (the collection of input data and its associated correct output data) is not thoughtfully chosen, the resulting network is unlikely to hold up well in an industrial environment. So it is hardly surprising that massaging the set of training data consumes some 80 percent of the engineering time spent getting a real-world neural network up and running-that is, getting it to converge under a broad enough range of conditions to be deployed with confidence in a production situation. If that data preparation is done systematically, much time can be saved and a more robust end-product can be obtained. A nine-step process is given that experience (the author's) has shown can enhance the probability of obtaining a learning convergence robust enough for industrial use.