Pattern recognition models

The main purposes of multivariate analysis are data reduction (unsupervised analysis) and data modeling like regression and/or classification models (supervised analysis). 

In this discussion, we will consider PCA and artificial neural network (ANN), both extensively applied to e-noses and e-tongues results both for discrimination and monitoring 

PCA is an unsupervised technique whose main idea is to explore data analysis and to reduce dimensionality with a minimum loss of information. This is achieved by projecting the data onto fewer dimensions that are chosen to exploit the relationships between the variables. It also measures qualitative associations between variables. To analyze the results, a line of best fit through a system of points in space is obtained. This technique allows the similarities and differences between objects and samples to be better assessed [30]. 

Several different types of ANN are available and the most popular is the backpropagation approach. In this procedure, input patterns presented to the input layer, for example, signals from an array of chemical sensors, generate a flow of activation to the output layer. Errors in the output are then fed back to the input layer to modify the weights of the interconnections. It should be emphasized that backpropagation does not describe a network but represents a learning algorithm. In this way, the network can be trained with known parameters, such as sensor array responses to sets of known chemicals. 

Since a neural network can arrive at different solutions for the same data, if different values of the initial network weights are provided, the network should be trained several times. The goal is to try and find a neural network model for which multiple trainings approach the same final mean squared error (MSE). 

---

The goal of unsupervised techniques is to identify and display natural groupings in the data without imposing any prior class membership. Even when the ultimate goal of the project is to develop a supervised pattern recognition model, we recommend the use of unsupervised techniques to provide an initial view of the data. 

In pattern recognition modeling, such as ADAPT, it is difficult to effectively visualize and manipulate chemical structure. Instead, there has been an effort to translate abstract structure into quantities and/or numerical entities (10), referred to as molecular descriptors. Such descriptors have been classified as presented in Table IV. 

Fluorescence spectra are collected under excitation conditions that are optimized to correlate the emission spectral features with parameters of interest. Principal components analysis (PCA) is further used to extract the desired spectral descriptors from the spectra. The PCA method is used to provide a pattern recognition model that correlates the features of fluorescence spectra with chemical properties, such as polymer molecular weight and the concentration of the formed branched side product, also known as Fries s product, that are in turn related to process conditions. The correlation of variation in these spectral descriptors with variation in the process conditions is obtained by analyzing the PCA scores. The scores are analyzed for their Euclidean distances between different process conditions as a function of catalyst concentration. Reaction variability is similarly assessed by analyzing the variability between groups of scores under identical process conditions. As a result the most appropriate process conditions are those that provide the largest differentiation between materials as a function of catalyst concentration and the smallest variability in materials between replicate polymerization reactions. 

Within the field of chemistry, various applications have already been published. Jansson " and Zupan and Gasteiger published an overview of an MLP (multilayer perceptron), that is trained by back-propagation of errors, and other types of neural networks. However, the basic source in this field continues to be the well-known book of Zupan and Gasteiger." In analytical chemistry, neural networks have been applied to pattern recognition, modeling, and prediction, for example, in multicomponent analysis or process control, to classification, clustering and pattern association. 

In SAR, the descriptors of molecules are used to determine whether their activities are high or not. Using data set of known groupings, a supervised pattern recognition model can be set up. Finally, the activity of an unknown sample can be predicted. 

Increased trust in pattern recognition The active user involvement in the data mining process can lead to a deeper understanding of the data and increases the trust in the resulting patterns. In contrast, "black box" systems often lead to a higher uncertainty, because the user usually does not know, in detail, what happened during the data analysis process. This may lead to a more difficult data interpretation and/or model prediction. 

Often the goal of a data analysis problem requites more than simple classification of samples into known categories. It is very often desirable to have a means to detect oudiers and to derive an estimate of the level of confidence in a classification result. These ate things that go beyond sttictiy nonparametric pattern recognition procedures. Also of interest is the abiUty to empirically model each category so that it is possible to make quantitative correlations and predictions with external continuous properties. As a result, a modeling and classification method called SIMCA has been developed to provide these capabihties (29—31). 

Artificial Neural Networks (ANNs) attempt to emulate their biological counterparts. McCulloch and Pitts (1943) proposed a simple model of a neuron, and Hebb (1949) described a technique which became known as Hebbian learning. Rosenblatt (1961), devised a single layer of neurons, called a Perceptron, that was used for optical pattern recognition. 

Reasoning in Time Modeling, Analysis, and Pattern Recognition of Temporal Process Trends... 

REASONING IN TIME MODELING, ANALYSIS, AND PATTERN RECOGNITION OF TEMPORAL PROCESS TRENDS... 

Many other subjects are important to achieve successful pattern recognition. To name only two, it should be investigated to what extent outliers are present, because these can have a profound influence on the quality of a model and to what extent clusters occur in a class (e.g. using the index of clustering tendency of Section 30.4.1). When clusters occur, we must wonder whether we should not consider two (or more) classes instead of a single class. These problems also affect multivariate calibration (Chapter 36) and we have discussed them to a somewhat greater extent in that chapter. 

J.D.F. Habbema, Some useful extensions of the standard model for probabilistic supervised pattern recognition. Anal. Chim. Acta, 150 (1983) 1-10. 

S. Wold, Pattern recognition by means of disjoint principal components models. Pattern Recogn., 8 (1976) 127-139. 

M.P. Derde and D.L. Massart, UNEQ a disjoint modelling technique for pattern recognition based on normal distribution. Anal. Chim. Acta, 184 (1986) 33-51. 

E. Saaksjarvi, M. Khaligi and P. Minkkinen, Waste water pollution modeling in the southern area of Lake Saimaa, Finland, by the simca pattern recognition method. Chemom. Intell. Lab. Systems, 7(1989) 171-180. 

In the following sections we propose typical methods of unsupervised learning and pattern recognition, the aim of which is to detect patterns in chemical, physicochemical and biological data, rather than to make predictions of biological activity. These inductive methods are useful in generating hypotheses and models which are to be verified (or falsified) by statistical inference. Cluster analysis has... 

While principal components models are used mostly in an unsupervised or exploratory mode, models based on canonical variates are often applied in a supervisory way for the prediction of biological activities from chemical, physicochemical or other biological parameters. In this section we discuss briefly the methods of linear discriminant analysis (LDA) and canonical correlation analysis (CCA). Although there has been an early awareness of these methods in QSAR [7,50], they have not been widely accepted. More recently they have been superseded by the successful introduction of partial least squares analysis (PLS) in QSAR. Nevertheless, the early pattern recognition techniques have prepared the minds for the introduction of modem chemometric approaches. 

Radial basis function networks (RBF) are a variant of three-layer feed forward networks (see Fig 44.18). They contain a pass-through input layer, a hidden layer and an output layer. A different approach for modelling the data is used. The transfer function in the hidden layer of RBF networks is called the kernel or basis function. For a detailed description the reader is referred to references [62,63]. Each node in the hidden unit contains thus such a kernel function. The main difference between the transfer function in MLF and the kernel function in RBF is that the latter (usually a Gaussian function) defines an ellipsoid in the input space. Whereas basically the MLF network divides the input space into regions via hyperplanes (see e.g. Figs. 44.12c and d), RBF networks divide the input space into hyperspheres by means of the kernel function with specified widths and centres. This can be compared with the density or potential methods in pattern recognition (see Section 33.2.5). 

---