초록 ▼

A type of neural network called a self-organizing map (SOM) is useful in pattern classification. The ability of the SOM to map the density of the input distribution is improved with two techniques. In the first technique, the SOM is improved by monitoring the frequency for which each node is the win

A type of neural network called a self-organizing map (SOM) is useful in pattern classification. The ability of the SOM to map the density of the input distribution is improved with two techniques. In the first technique, the SOM is improved by monitoring the frequency for which each node is the winning node, and splitting frequently winning nodes into two nodes, while eliminating infrequently winning nodes. Topological order is preserved by inserting a link between the preceding and following nodes so that such preceding and following nodes are now adjacent in the output index space. In the second technique, the SOM is trained by applying a weight correction to each node based on the frequencies of that node and its neighbors. If any of the adjacent nodes have a frequency greater than the frequency of the present node, then the weight vector of the present node is adjusted towards the highest-frequency neighboring node. The topological order of the nodes is preserved because the weight vector is moved along a line of connection from the present node to the highest-frequency adjacent node. This second technique is suitable for mapping to an index space of any dimension, while the first technique is practical only for a one-dimensional output space.

대표청구항 ▼

[ What is claimed is:] [9.] In a neural network for the classification of a plurality of patterns, wherein each of said plurality of patterns is represented by a respective multi-dimensional vector in a multi-dimensional space to form a training set of multi-dimensional input vectors corresponding t

[ What is claimed is:] [9.] In a neural network for the classification of a plurality of patterns, wherein each of said plurality of patterns is represented by a respective multi-dimensional vector in a multi-dimensional space to form a training set of multi-dimensional input vectors corresponding to said plurality of input patterns, said neural network having a plurality of nodes in an ordered array corresponding to a topological ordering of said plurality of input patterns, a method for improving the ordering of said plurality of multi-dimensional input vectors, said method comprising:determining the frequency that each node in said multi-dimensional space has been previously determined as the closest node to some portion of said plurality of input vectors forming said training set of multi-dimensional input vectors corresponding to said plurality of input patterns, to form a determined frequency for each of said plurality of nodes; andcomparing the determined frequency of a given node to the determined frequency of an adjacent node; andchanging the position of said given node to form a moved node based on said determined frequency of said given node and said determined frequency of said adjacent node.