Sensory-instrumental correlations by combining data analysis and neural network techniques

Abstract The scope of the paper is to investigate the potential of multi-layer neural networks for modelling the correlations between instrumental and sensory characteristics of food products. Multi-layer neural networks are artificial intelligence models that have shown an increased power over data analysis techniques when dealing with non-linear prediction problems. Techniques such as principal components analysis or canonical analysis are traditionally used for modelling the correlations between instrumental and sensory variables. The authors suggest combining these techniques in order to extract non-linear regularities and to improve predictions with noisy or complete data. This method has already been tested on various applications. The paper will focus on the prediction of beverage flavours from their composition. The data bases used for constructing the models consist of 81 beverage samples for which sensory responses have been scored by trained panelists and data on chemical composition have been collected. Although linear predictions are sometimes difficult to improve upon, the authors show that neural networks can give a very significant performance increase compared to classical multi-variate analysis.