scientific article; zbMATH DE number 2134210
The dissertation thesis is dedicated to the topic of classification by machine learning approaches. In particular, the Learning Vector Quantization method (LVQ) and the Self-Organizing Map (SOM) is considered. Both methods belong to prototype-based vector quantization algorithms [see \textit{T. Kohonen}, Self-organizing maps. Berlin: Springer (1994; Zbl 0827.68092); 2nd ed. (1997; Zbl 0866.68085); 3rd ed. (2001; Zbl 0957.68097)]. While the first one is a supervised learning scheme for classification, the second is unsupervised for clustering. In the work, generalizations and extensions of the LVQ are considered. These are the LVQ with cost function and adaptive (may be non-standard) metrics (for example: scaled Euclidean metric, local kernels,\dots). Further, neighborhood cooperativeness is included into the learning strategy. Metric adaptation is called relevance learning. The resulting algorithm is the Supervised Relevance Neural Gas. Thereby, relevance learning yields a stepwise hierarchy of metric parameters which can be used to generate desicion trees for classification (here BB-trees). The possibilities of the algorithm are studied for several real-world problems ranging from bioinformatics to linguistics. For the unsupervised clustering by SOM new extensions of the SOM for sequences and strucured data are introduced and investigated for several examples of data. SOMs originally were introduced for low-dimensional representation of possibly high-dimensional data by a typically two-dimensional grid in a similarity preserving way, whereby the SOM model as well as the training procedure are inspired by neural information processing in biological brain. The proposed extension is a context-based determination of the best matching prototype. In case of sequences the context could be the information about the preceding best matching prototypes, for example. The influence of the new learning strategy is considered for different lattice topologies and subsequently transferred to another neural network scheme -- the neural gas introduced by T. Martinetz. Neural gas is similar to SOM but with adaptive lattice structure instead of an apriori fixed grid. All new features are illustrated by real-world applications underlining the possibilities and properties. Summarizing, the thesis gives valuable new results in the field of supervised and unsupervised vector quantization by neural maps and learning vector quantization.
This page was built for publication:
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3159374)
