FEATURE DETECTION NETWORKS IN PATTERN RECOGNITION,
Abstract
In some pattern recognition problems a large number of patterns may be decomposed into a small set of subpatterns which can reconstitute each of the patterns. The use of such features can result in an economical recognition network. In such cases the pattern set may have an inherent hierarchical structure which can be incorporated in a layered logical network. An algorithm is presented which uses a training set of patterns to determine this structure. The subpatterns, termed features, are generated sequentially through an adaptive process of weight alteration in a neural network as each pattern is iteratively presented. A measure of 'relatedness' of a set of points is utilized to decide which subset of points associated with these sets of weights represents useful information and should be selected as a feature. Experimental results indicate the potential of the algorithm in organizing a recognition network to correspond to the information structure of the pattern set. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 1969
- Accession Number
- AD0689429
Entities
People
- Edward M. Riseman
Organizations
- Cornell University