An Unsupervised Learning Problem Using Limited Storage Capacity

In unsupervised learning pattern recognition problems, the need arises for updating conditional density functions of uncertain parameters using probability density function mixtures. In general, the form of the density mixtures is not reproducing, invoking the need for unlimited system storage requirements. One suboptimal method for achieving limited storage is to restrict the uncertain parameters in question to come from finite sets of values. An alternate method is proposed for a class of problems and its performance is shown to converge to that of the optimum unlimited storage system. A generalization of the procedure is also discussed.