Pattern Recognition Research

Machine Learning and Pattern Recognition is treated as the problem of adaptively constructing approximations to the joint probability densities of the N-variables with which members of classes are represented. The adaptive techniques studied construct approximations to the joint probability densities in the form of generalized N-dimensional histograms in which the locations, shapes and sizes of the histogram cells are generated by the known samples of the pattern classes. To economize on the number of cells constructed, a cell growth mechanism was devised to adapt the size and shape of the cells to best represent the probability densities. The accuracy of this method of representation was tested with the aid of a digital computer on large quantities of pattern samples of known probability distribution. The experimental results were compared with those that could be predicted theoretically. Quality criteria to assess the reliability of the decision rendered by a classification device and to influence the mechanism of machine learning were considered.