Network neural

Automated data interpretation will usually be done using some statistical or AI technique. Because statistical classifiers are similar in their use to neural networks [Sarle, 1994] we will not discuss them separately. 

Artificial Neural Networks. An Artificial Neural Network (ANN) consists of a network of nodes (processing elements) connected via adjustable weights [Zurada, 1992]. The weights can be adjusted so that a network learns a mapping represented by a set of example input/output pairs. An ANN can in theory reproduce any continuous function 95 —>31 °, where n and m are numbers of input and output nodes. In NDT neural networks are usually used as classifiers... 

Hybrid systems. Depending on the problem to be solved, use can also be made of a combination of techniques leading to a hybrid system. For example, a rule-based system may use neural networks for solving classification subproblems (as is described in [Hopgood, 1993]), or a combination of a rule-based and a CBR system can be used as in the system for URS data interpretation described later in this paper. 

Neural network classifiers. The neural network or other statistical classifiers impose strong requirements on the data and the inspection, however, when these are fulfilled then good fully automatic classification systems can be developed within a short period of time. This is for example the case if the inspection is a part of a manufacturing process, where the inspected pieces and the possible defect mechanisms are well known and the whole NDT inspection is done in repeatable conditions. In such cases it is possible to collect (or manufacture) as set of defect pieces, which can be used to obtain a training set. There are some commercially available tools (like ICEPAK [Chan, et al., 1988]) which can construct classifiers without any a-priori information, based only on the training sets of data. One has, however, always to remember about the limitations of this technique, otherwise serious misclassifications may go unnoticed. 

The recognition ratios achieved by CBR systems developed as part of this project could not be bettered by either neural-network classifiers or rule-based expert system classifiers. In addition, CBR systems should be mote reliable than simple classifiers as they are programmed to recognise unknown data. The knowledge acquisition necessary to build CBR systems is less expensive than for expert systems, because it is simpler to describe the knowledge how to distinguish between certain types of data than to describe the whole data contents. 

The end users of CBR systems should in principle be able to maintain the case-bases themselves and use the systems for varying inspection types (within certain limits). Adaptation of neural-network based systems, though possible by end-users, is difficult to be done reliably. Adaptation of rule-based systems usually has to be done by the rule-base designer. 

Hopgood, F.F., N. Woodcock, N.J. HaUam, and P.D. Picton, Interpreting ultrasonic images using rules, algorithms and neural networks , European Journal of NDT, Vol. 2, No. 4, April 1993, pp. 135-149. 

Sarle, W.S., (1994), Neural Networks and Statistical Models , Proceedings of the Nineteenth Annual SAS Users Group International Conference, April, 1994. 

Neural Network Based Classifier for Ultrasonic Resonance Spectra. 

We have presented a neural network based spectrum classifier (NSC) aimed at ultrasonic resonance spectroscopy. The ultrasonic spectroscopy and the NSC has been evaluated in many industrial applications, such as concrete inspection, testing of aerospace composite structures, ball bearings, and aircraft multi-layer structures. The latter application has been presented in some detail. 

The results obtained with NSC in different applications show that both flaw detection and localization can be performed automatically by the use of a neural network classifier. 

Pattern Recognition of Artificial Neural Network to Waveform Data. 

Freeman, J. A., Skapura, D. M. Neural Networks Algorithms, Applications and Programming Techniques, Computation and Neural systems Series. Addison Wesley Publishing Company, 1991... 

The local dynamics of tire systems considered tluis far has been eitlier steady or oscillatory. However, we may consider reaction-diffusion media where tire local reaction rates give rise to chaotic temporal behaviour of tire sort discussed earlier. Diffusional coupling of such local chaotic elements can lead to new types of spatio-temporal periodic and chaotic states. It is possible to find phase-synchronized states in such systems where tire amplitude varies chaotically from site to site in tire medium whilst a suitably defined phase is synclironized tliroughout tire medium 51. Such phase synclironization may play a role in layered neural networks and perceptive processes in mammals. Somewhat suriDrisingly, even when tire local dynamics is chaotic, tire system may support spiral waves... 

The method that was developed builds on computed values of physicochemical effects and uses neural networks for classification. Therefore, for a deeper understanding of this form of reaction classification, later chapters should be consulted on topics such as methods for the calculation of physicochemical effects (Section 7.1) and artificial neural networks (Section 9.4). 

Previous work in our group had shown the power of self-organizing neural networks for the projection of high-dimensional datasets into two dimensions while preserving clusters present in the high-dimensional space even after projection [27]. In effect, 2D maps of the high-dimensional data are obtained that can show clusters of similar objects. 

Let us start with a classic example. We had a dataset of 31 steroids. The spatial autocorrelation vector (more about autocorrelation vectors can be found in Chapter 8) stood as the set of molecular descriptors. The task was to model the Corticosteroid Ringing Globulin (CBG) affinity of the steroids. A feed-forward multilayer neural network trained with the back-propagation learning rule was employed as the learning method. The dataset itself was available in electronic form. More details can be found in Ref. [2]. 

This format was developed in our group and is used fruitfully in SONNIA, software for producing Kohonen Self Organizing Maps (KSOM) and Coimter-Propaga-tion (CPG) neural networks for chemical application [6]. This file format is ASCII-based, contains the entire information about patterns and usually comes with the extension "dat . 

Now, one may ask, what if we are going to use Feed-Forward Neural Networks with the Back-Propagation learning rule Then, obviously, SVD can be used as a data transformation technique. PCA and SVD are often used as synonyms. Below we shall use PCA in the classical context and SVD in the case when it is applied to the data matrix before training any neural network, i.e., Kohonen s Self-Organizing Maps, or Counter-Propagation Neural Networks. 

The profits from using this approach are dear. Any neural network applied as a mapping device between independent variables and responses requires more computational time and resources than PCR or PLS. Therefore, an increase in the dimensionality of the input (characteristic) vector results in a significant increase in computation time. As our observations have shown, the same is not the case with PLS. Therefore, SVD as a data transformation technique enables one to apply as many molecular descriptors as are at one s disposal, but finally to use latent variables as an input vector of much lower dimensionality for training neural networks. Again, SVD concentrates most of the relevant information (very often about 95 %) in a few initial columns of die scores matrix. 

M. Smith, Neural Networks in Statistical Modelling. Van Nostrand Reinhold, New York, 1993. 

J. Zupan, J. Gasteiger, Neural Networks in Chemistry and Drug Design. Wiley-VCH, Weinheim, 1999. 

SONNIA - KSOM and CPG neural networks. The key features are robustness of training and excellent visualization capabilities. 

Another approach employing the autocorrelation coefficients as descriptors was suggested by Gasteiger et al, [22]. They used the neural networks as a working tool for solving a similarity problem. 

A challenging task in material science as well as in pharmaceutical research is to custom tailor a compound s properties. George S. Hammond stated that the most fundamental and lasting objective of synthesis is not production of new compounds, but production of properties (Norris Award Lecture, 1968). The molecular structure of an organic or inorganic compound determines its properties. Nevertheless, methods for the direct prediction of a compound s properties based on its molecular structure are usually not available (Figure 8-1). Therefore, the establishment of Quantitative Structure-Property Relationships (QSPRs) and Quantitative Structure-Activity Relationships (QSARs) uses an indirect approach in order to tackle this problem. In the first step, numerical descriptors encoding information about the molecular structure are calculated for a set of compounds. Secondly, statistical and artificial neural network models are used to predict the property or activity of interest based on these descriptors or a suitable subset. 

A structure descriptor is a mathematical representation of a molecule resulting from a procedure transforming the structural information encoded within a symbolic representation of a molecule. This mathematical representation has to be invariant to the molecule s size and number of atoms, to allow model building with statistical methods and artificial neural networks. 

Neural networks were trained on the basis of these codes to predict chiralit> -dependent properties in enantioselective reactions [42] and in chiral chromatography [43]. A detailed description of the chirality codes is given in the Tutorial in Section 8,6,... 

Chirality codes are used to represent molecular chirality by a fixed number of de-.scriptors. Thc.se descriptors can then be correlated with molecular properties by way of statistical methods or artificial neural networks, for example. The importance of using descriptors that take different values for opposite enantiomers resides in the fact that observable properties are often different for opposite enantiomers. 

Therefore the 28 analytes and their enantiomers were encoded by the conformation-dependent chirality code (CDCC) and submitted to a Kohoiien neural network (Figure 8-1 3). They were divided into a test set of six compounds that were chosen to cover a variety of skeletons and were not used for the training. That left a training set containing the remaining 50 compounds. 

Figure 8-1J. Training ofa Kohonen neural network with a chirality code, The number of weights in a neuron is the same as the number of elements in the chirality code vector, When a chirality code is presented to the network, the neuron with the most similar weights to the chirality code is excited (this is the ivinning or central neuron) (see Section 9.5,3),...

To understand neural networks, especially Kohonen, counter-propagation and back-propagation networks, and their applications... 

Kohonen networks, also known as self-organizing maps (SOMs), belong to the large group of methods called artificial neural networks. Artificial neural networks (ANNs) are techniques which process information in a way that is motivated by the functionality of biological nervous systems. For a more detailed description see Section 9.5. 

---