A unified framework for connectionist systems.

Pattern classification using connectionist (i.e., neural network) models is viewed within a statistical framework. A connectionist network's subjective beliefs about its statistical environment are derived. This belief structure is the network's "subjective" probability distribution. Stimulus classification is interpreted as computing the "most probable" response for a given stimulus with respect to the subjective probability distribution. Given the subjective probability distribution, learning algorithms can be analyzed and designed using maximum likelihood estimation techniques, and statistical tests can be developed to evaluate and compare network architectures. The framework is applicable to many connectionist networks including those of Hopfield (1982, 1984), Cohen and Grossberg (1983), Anderson et al. (1977), and Rumelhart et al. (1986b).