A Probabilistic Computational Framework for Neural Network Models
Abstract
Information retrieval in a 'connectionist' or neural network is viewed as computing the most probable value of the information to be retrieved with respect to a probability density function, P. With a minimal number of assumptions, the 'energy' function that a neural network minimizes during information retrieval is shown to uniquely specify P. Inspection of the form of P indicates the class of probabilistic environments that can be learned. Learning algorithms can be analyzed and designed by using maximum likelihood estimation techniques to estimate the parameters of P. The large class of nonlinear auto-associative networks analyzed by Cohen and Grossberg (1983), nonlinear associative multi-layer back-propagation networks (Rumelhart, Hinton, & Williams, 1986), and certain classes of nonlinear multi-stage networks are analyzed within the proposed computational framework. Keywords: Artificial intelligence, Connectionism, Non-linear associator.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 29, 1987
- Accession Number
- ADA218969
Entities
People
- Richard M. Golden
Organizations
- Carnegie Mellon University