Aspects of network training and validation on noisy data: Part 1. Training aspects

Multi-layered feed-forward (MLF) neural networks are commonly trained by the generalized Delta learning rule. The training method has shown to be robust and easy to implement for various chemical problems in analytical chemistry. However, the slow and unpredictable learning behavior puts considerable limitations to the interpretation and validation of neural network models. In this two-part paper, both training and validation aspects are addressed. The first part of this paper focuses to the generalized Delta learning rule and some important aspects of network training. It is shown that the use of quasi-Newton training on unfolded MLF networks, accelerates the training time to an order of magnitude. The learning behavior of MLF-networks trained by the generalised Delta rule and the quasi-Newton learning rule are compared by means of simulated and real data from analytical chemistry. Some conclusions about the general applicability of the discussed methods are drawn.