Showing papers on "Unsupervised learning published in 1976"

Sequential learning employing unfamiliar teacher hypothesis (SLEUTH) with concurrent estimation of both the parameters and teacher characteristics

Abstract: Parametric learning in an unfamiliar environment, i.e., under unknown degree of supervision, forms the topic of this study. In addition to the learning of the distribution parameters, the learning scheme presented here has the novel capability of learning about the unfamiliar teacher, i.e., estimating the teacher characteristics inherent in the environment. This learning, under an unfamiliar teacher hypothesis, is seen to be computationally feasible and the results of simulation reveal the efficacy of the scheme in improving the learning under the unfamiliar teacher through learning about the teacher.

Quasi-Bayes procedures for unsupervised learning

Abstract: Unsupervised Bayes sequential learning procedures for classification and estimation are often useless in practice because of computational constraints. In this paper, a quasi-Bayes approach is motivated, and discussed in detail for some versions of a two-class decision problem. The proposed procedure mimics closely the formal Bayes solution, whilst involving only a minimal amount of computation. Some numerical illustrations are provided, and the approach is compared with a number of other proposed learning procedures.

HINDU: Histogram Inspired Neighborhood Discerning Unsupervised Pattern Recognition System

Abstract: A newly developed system for pattern recognition in unsupervised environments, capable of processing large volume data sets with minimal computational resources and human intervention, which is currently operational at the Marshall Space Flight Center, is presented in this study. Here, in this system, the problem of unsupervised learning is viewed as one of clustering the large volume multidimensional data sets and is approached through· the novel gambit of terrain development in the multidimensional histogram space. The terrain is developed by connecting each histogram cell to all of its higher density neighbors. This process leads to amalgamation of all the cells belonging to each of the clusters. Certain of these cells, being connected to more than one cluster, defiriafuzzy boundaries between the clusters. Discriminant hyperplanes, which not only separate these clusters but also form least square fits to the centroids of the cells defining the fuzzy boundaries, are derived. The design of these discriminant functions is through a new algorithm developed specifically for catering to this problem environment of discriminating between clusters with fuzzy boundaries. The dimensionality curse, an often encountered problem of computational complexity arising from higher dimensionality of data sets, is tackled here by a relatively simple pre-processing technique of ordering and selecting features on the basis of an adhoc histogram information measure sensed for each of the features. The conceptual and computational claims of this HINDU system, presently in operational status on an IBM 360/65, have been verified by simulation tests using remotely sensed multispectral LANDSAT data. *This work was performed under Contract NAS8-31639 for the Data Systems Laboratory Nationa·l Aeronautics and Space Administration, George C. Harshall Space Flight center, Huntsville, Alabama.