Neural Computation And It's Application In Sensory Evaluation

In recent years, artificial neural networks have not only achieved great progress both in theories and application, but expanded connotation of computing to make neural computing a new science. Now Neural computing and its applications have already come into many disciplines and achieved plentiful fruits in diversified fields, including signal processing, intelligent controlling, pattern recognition, machine vision, nonlinear optimization, automatic target identification, knowledge processing, remote sensing, etc. It has become not only the tastes of scientists but also the interests of governments and forces. The governments and industrial communities of many countries/regions are so keen on neural computing techniques that they have invested a large amount money on corresponding research. Therefore the progress of neural computing will not only promote the development of science and technology but also influence the national powers. In fact, the researches on then relation of brain-thinking-computing are just on the way. Some frontiers in the theories and applications of neural computing will be associated with many challenging issues in the 21th century and make great breakthrough.In this dissertation, 3 problems standing in need of solutions are investigated, which includes improving the comprehensibility of neural networks, combining neural computing with traditional multivariate statistical analysis techniques, and applying neural networks in tobacco sensory evaluation. The main contributions of this dissertation are summarized as follows:Firstly, based on the BP neural net's powerful ability of simulating function, the knowledge in the trained BP nets may be extracted by the method of constructing ladder-samples, visually reflecting it in the form of ladder-image. We can analyze the correlation between inputs and outputs of the nets quantitatively by this way. In contrast with traditional method, this new method unfolds the linear and nonlinear correlation between inputs and outputs in the samples more effectively. When this method is applied in forecasting the sensory-qualities evaluation of tobaccos, thesimulative experiments have testified its validity.Secondly, two kinds of simulated elastic-forces are defined, which are in direct proportion to the distance between pattern vectors and weight vectors. The mechanism of the evolvement of the weight vectors and the topological ordering of Kohonen's SOM may be understandably and visually explained by the proposed elastic-forces.Thirdly, based multivariate statistical analysis, a new method to evaluate performance of neural networks is proposed. Experimental results show that we indeed can select an excellent trained neural networks by this new method.Finally, combining neural computing and multivariate statistical analysis, industrial data set has been cleaned and it's feature been extracted. Furthermore neural networks model has been proposed to help tobacco experts in sensory evaluating. Practical applications have shown that models work quite well.