ADAPTIVE PATTERN RECOGNITION USING NON-LINEAR ELEMENTS.
Abstract
A pattern recognition device which computes a weighted sum of all products of its binary input variables is investigated. It is proven that this device can uniquely synthesize any real function of its inputs. The size of this device becomes prohibitive as the number of inputs increases and the concept of an incomplete device is introduced and discussed. A convergence proof of an adaptive training procedure for real outputs, based on the equivalence between this device and a multi-threshold linear device in a larger input space, is presented. A set of orthogonal property detectors is defined and shown to be numerically equal to the number of binary inputs. A computer simulation of a 64-input incomplete device shows that its generalizing abilities are more fully utilized with the aid of specific property detectors. A self-organizing technique consisting of periodically discarding low-weighted terms and replacing them with randomly chosen terms is discussed. Experiments conducted using this technique demonstrate that the learning and generalizing performances of the incomplete device are greatly improved if the device is free to change its structure according to a heuristic policy. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 1964
- Accession Number
- AD0623215
Entities
People
- Anthony N. Mucciardi