Methods for solving pattern recognition tasks generally assume a sequential model for the pattern recognition process, consisting of pattern environment, sensors to collect data from the environment, feature extraction from the data and association/ storage/classification/clustering using the features. The simplest solution to a pattern recognition problem is to use template matching, where the data of the test pattern are matched point by point with the corresponding data in the reference pattern. Obviously, this can work only for very simple and highly restricted pattern recognition tasks. At the next level of complexity, one can assume a deterministic model for the pattern generation process, and derive the parameters of the model from given data in order to represent the pattern information in the data. Matching test and reference patterns are done at the parametric level. This works well when the model of the gene;ation process is known with reasonable accuracy. One could also assume a stochastic model for the pattern generation process, and derive the parameters of the model from a large set of training patterns. Matching between test and reference patterns can be performed by several statistical methods like likelihood ratio, variance weighted distance, Bayesian classification etc. Other approaches for pattern recognition tasks depend on extracting features from parameters or data. These features may be specific for the task. A pattern is described in terms of features, and pattern matching is done using descriptions of the features.
Another method based on descriptions is called syntactic or structural pattern recognition in which a pattern in expressed in terms of primitives suitable for the classes of pattern under study (Schalkoft 1992). Pattern matching is performed by matching the descriptions of the patterns in terms of the primitives. More recently, methods based on the knowledge of the sources generating the patterns are being explored for pattern recognition tasks. These knowledge-based systems express I knowledge in the form of rules for generating and perceiving patterns.
The main difficulty in each of the pattern recognition techniques alluded to above is that of choosing an appropriate model for the pattern generating process and estimating the parameters of the model in the case of a model-based approach, or extraction of features from data parameters in the case of feature-based methods, or selecting appropriate primitives in the case of syntactic pattern recognition, or deriving rules in the case of a knowledge-based approach. It is all the more difficult when the test patterns are noisy and distorted versions of the patterns used in the training process. The ultimate goal is to impart to a machine the pattern recognition capabilities comparable to those of human beings. This goal is difficult to achieve using most of the conventional methods, because, as mentioned earlier, these methods assume a sequential model for the pattern recognition process. On the other hand, the human pattern recognition process is an integrated process involving the use of biological neural processing even from the stage of sensing the environment. Thus the neural processing takes place directly on the data for feature extraction and pattern matching. Moreover, the large size (in terms of number of neurons and interconnections) of the biological neural network and the inherently different r mechanism of processing are attributed to our abilities of pattern recognition in spite of variability and noise in the data. Moreover, we are able to deal effortlessly with temporal patterns and also with the so-called stability-plasticity dilemma as well.
It is for these reasons attempts are being made to explore new models of computing, inspired by the structure and function of the biological neural network. Such models for computing are based on artificial neural networks,