| Introducing Neural Networks |
|
The Neural Networks module displays classification results from Self-Organizing Maps (SOM), which are a special class of neural network. A self-organizing map is a network of neurons, arranged in the form of a two-dimensional lattice. The size of a lattice can either be calculated automatically or defined by the user. During a classification neurons become selectively activated to various input spectra as a result of a competitive learning process. The objective of classification analysis using SOM is to find classes of spectra on the map that exhibit common or similar properties. If one or more spectra activate the same neuron, we can assume the spectra belong to a common class. In this case the spectra should exhibit certain similarities. In addition, spectra that activate neighboring neurons, and those neurons that have low Euclidian distance between each other (shown by border line thickness), can also be considered as related. In neural networks each mass spectrum must always activate a neuron and this spectrum is shown on the particular neuron. Neurons are displayed as rectangles on the screen. Spectra are represented as symbols or numbers and are placed onto neurons. Since the spectra are located in discrete objects the interpretation of SOM is relatively easy, as, in contrast to PCA and Fuzzy Clustering, you do not have to deal with diffuse clusters. However, it may happen that larger numbers of neurons are activated by a single spectra and this advantage is lost. Spectra that activate the same neuron belong, in terms of classification, to the same pattern. Therefore, all spectra drawn inside a neuron box are equal for classification purposes and their graphical positions within a neuron are irrelevant. Note: Neural Networks topics:
See also: |
