The Study On Adaptive Resonance Theory Principles And Applications

The artificial neural network is a signal processing system which mimicthe biological neural network, and, it mimic the human neural signalprocessing through a large number of interconnected artificial neurons weightchange. The artificial neural network has learning, classification, predictioncapabilities. So far, artificial neural networks for a variety of applicationenvironments have been developed. The artificial neural network is a highlynonlinear technology, the internal processing units (neurons) of which canwork independently, and, it's capable of parallel processing of signals and ableto tolerate a certain degree of input noise.Adaptive Resonance Theory (ART) is a non-supervised learning artificialneural network (Unsupervised learning), its appearance solved the stability ofneural network learning/plasticity dilemma. The ART is proposed byGrossberg in the1970s, it can continue to learn new knowledge without affectthe old ones and design or application of the training process. For the learningframework is unsupervised type, ART is very suitable for pattern recognitionand error detection.In-depth analysis of ART1and ART2are presented in this paper, which include their mathematical basis, the basic rules to follow when learning, inaddition, experimental verification of their performance. On this basis, theprinciples of experimental psychology, developed a fast adaptive resonancenetwork. The network according to the memory strength of grouping adaptiveresonance theory network clustering, pattern recognition process is dividedinto several sequential sub-process. A reasonable set of network parameterswill ensure accurate identification under the premise of the new network asmuch as possible to avoid the adaptive resonance network inherent in thetraversal matching operation. Fast adaptive resonance theory network retainsall the advantages of ART, and joined a pre-processing aspects of ART1withthe ability to handle analog signals. The experiments show that the fairly newnetwork, pattern recognition performance with the original network, while thecomputational efficiency has been significantly improved.