|
|
Libann: Pattern Classification and Neural Networks |
|
1. Pattern Classification and Neural NetworksPattern classification problems have three distinct stage:
The pre-processing stage typically involves filtering, normalisation and conversion of your data into a form acceptable to the computer. Feature extraction takes each sample and creates a feature vector from it. The elements of a feature vector can be any quantities which you think will be useful to classify the datum. However, you must choose the same feature vector definition for all your samples.
1.1 The Feature VectorFor example, given a bunch of physical data about set of persons, you might be asked classify them into male and female. You believe that, in general, men are taller than women, and so you would use the height as one of the elements of your feature vector. Height alone, cannot give very satisfactory results, and so you would probably want to have the individuals' weight as another element. The combination of height and weight is better, but still not perfect for the problem. You need to choose as an optimum number of features for your vector. In this case you would probably want to include the size of the individuals' hips, chest and other important measurements, assuming that this data is available.You must choose the optimum set of features and your program must arrange extract these into a vector. The order of elements in the vector is not important, except that it must be the same for all samples. 3. Creating a Feature Vector shows you how to set about creating your feature vectors.
1.2 ClassifiersA classifier takes a feature vector and attempts to decide to which class it belongs. Neural Networks are just one type of classifier. Other classifiers work on statistical principles.In the male/female example, a statistical classifier with a feature vector which comprised only the individual's height would, based upon the statistical evidence, decide upon a critical height value. All samples above this height would be classified as `male' and all below as `female'. Clearly, there would be a number of samples which would be misclassified, but this number has been minimised according to statistical principles. If a larger the feature vector is used, then a more complex decision rule (and hopefully a more accurate one) can be formed. A Neural Network is a classifier which has the advantage that it doesn't need to know any explicit rules. This makes it useful in situations where no mathematical model exists for the problem, or where the model is so complex as to be intractable.
1.3 Training a Neural NetworkIn general Neural Networks have two modes. There is a training mode, and there is a recall mode. In the training mode, the network is presented with the feature vectors of a large number of samples. During this mode it `learns' the pattern of the data. In the recall mode, the network is presented with the feature vector of a single sample. The output of the network determines the class to which the sample belongs.
1.3.1 Supervised and Unsupervised LearningDepending upon the type of network, the training mode may be a supervised learning mode or an unsupervised learning mode. Supervised learning means that the network is told the classes to which each sample belongs at training time. Unsupervised learning means that the network is not told the class of any of its training samples. It has the sole responsibility of determining the most appropriate way to classify the samples.
1.4 Performance of Neural NetworksAlthough the training mode can take a long time, a neural network has the advantage that the recall mode is very fast. In certain applications, a long training time is not of penalty (for example the network can be pre-trained in an environment with a lot of computing power) and this makes them an ideal candidate for the neural net approach.
This document was generated by John Darrington on May, 15 2003 using texi2html |