Uncertain Reasoning with RAM Neural Network

J. Austin

Synopsis

An overview of the development of the RAM based ADAM associative memory system is presented from its origins in the N tuple method of pattern recognition. Its relationship with the WISARD system and biological neural networks is discussed. The use of ADAM tor reasoning about uncertain data is presented showing how it may be used to efficiently track multiple uncertain hypotheses using a parallel distributed mechanism.

Key words:

ADAM, neural networks, uncertain reasoning.

1. Introduction

This paper introduces a method for reasoning with data using associative memories based on digital neural networks. It describes how networks may be used to infer from uncertain data at high speed Although the implementation is not biologically plausible the method by which reasoning is achieved may have some cognitive significance. There are many approaches to the problem of automated reasoning, most are typified by that used in Expert Systems where inference is carried out by an inference engine. One method operates by comparing pre-conditions of rules with a knowledge base and then altering that knowledge base by the action of the rule.

Some of the major problems faced by inference engines have been:

  1. Fast and efficient matching of rules.
  2. Dealing with uncertainties.
What we wish to show here is how associative memories, based upon neural networks, are able to easily overcome these problems. A great deal of interest has been given to the use of neural networks for applications which have previously involved conventionalexpert systems. However, one of the frequently quoted limitations of their application has been the speed at which new rules may be learned by the systems. If they are to be used in practical applications where rule update happens regularly, conventional multilayer networks (MLN) are not suitable due to their long training times.

Our approach is based upon the use of digital neural networks. These networks, in the form of an advanced distributed associative memory (ADAM) (Austin, 1987a), do not suffer the long training times associated with conventional MLNs. Thus they are applicable in a wider variety of applications than MLNs. The following sections describe how a rule based expert system may be mapped onto a neural network approach. Following this, a description of ADAM is given. This is concluded with an illustration of the abllities of a neural network approach when applied to uncertain data.