Interpretation and improvement of an artificial neural network MIR calibration

This work investigates whether useful information can be extracted from a minute analysis of the parameters of a trained artificial neural network (NN). Understanding the data processing that are performed is very useful for the optimisation and validation of neural network multivariate models, particularly if it compensates for their “black box” drawbacks. We focused on the development of calibration models to predict the degree of hydrolysis of bovine hemoglobin from on line mid-infrared (MIR) spectra recorded in a batch reactor. The situation is challenging since the trained network architecture has to model a clustered data set where relationships among the data clearly deviate from the ideal linear situation. Through the analysis of the transfer functions, activation and outputs of the hidden nodes, we show (1) how the nonlinear aspect of the data is processed and (2) how the strong clustering of the data set is considered. The data representations in the hidden layer present meaningful abstraction levels for the analysis of the learning performed. Beyond these results, the major improvement is to help decide the choice of architecture that will be provided through understanding the role played by each unit. The size of the hidden layer that seems the most suitable for generalisation is chosen despite that the root mean squared (RMS) prediction error is not the lowest possible.