Rugged spectroscopic calibration using neural networks

One of the greatest obstacles to wide spread adoption of multivariate spectroscopic assays for process control is the lack of ruggedness with respect to changes in instrument response. In the course of applying backpropagation artificial neural networks to a wide range of spectroscopic calibration applications we have found that neural networks have a tendency to overfit calibration data sets. Networks that overfit calibration data sets give poor predictive performance and exhibit sensitivity to small measurement errors on future samples. In addition, retraining after random initialization to new values does not converge to the same result when the quasi-Newton optimization method is employed. This behavior is due in part to the very large number of parameters being optimized and noise in the optimization surface. We have recently devised a training strategy to develop rugged calibrations that utilizes the tendency of neural networks to converge to different results. A rugged calibration is one which we hypothesize is insensitive to slight changes in instrument response due to wavelength calibration errors, baseline offsets or path-length changes