Non-parametric unsupervised learning with applications to image classification

摘要

A non-parametric, unsupervised learning technique is described. The technique makes use of a relation matrix to classify binary pattern vectors presented in random sequence. As each vector is classified, the elements of the matrix are adjusted in such a way as to reinforce the latest class assignment. A preliminary analysis shows that this process produces decision surfaces of a reasonable form. Extensive experiments with both simulated and real-world data confirm that the method performs very well in many circumstances.