Artificial neural networks for pattern recognition

Abstract

The use of artificial neural networks (ANNs) for recognising patterns in biological data is explained. The architecture and training of back propagation (multilayer perceptron), radial basis function (RBF) and learning vector quantization ANNs are described, as examples of ANNs which employ supervised learning and which are appropriate for biological identification. The Kohonen self-organising map (SOM) and ART (adaptive resonance theory) are presented as valuable classification techniques. The major considerations for implementing ANNs are discussed, including software, data pre-processing and coding, optimisation, testing trained networks, and coping with missing data. General issues such as scaling up, detecting novel data patterns, modifying networks, determining significance of different parameters, and pruning networks are also dealt with. Application of the supervised training approach is illustrated with a case study on identification and quantification of marine phytoplankton from flow cytometry data, using RBF ANN. The unsupervised approach is illustrated in a case study on classification/recognition of groups of phytoplankton, using Kohonen SOMs.