<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1113">
  <Title>Distributed Memory: A Basis for Chart Parsing</Title>
  <Section position="2" start_page="0" end_page="476" type="intro">
    <SectionTitle>
1. DISTRIBUTED MEMORY SYSTEMS
</SectionTitle>
    <Paragraph position="0"> Constraints on space make it impossible to describe distributed memory systems in detail, and this section merely outlines their salient properties. For a more detailed description of their properties and operation see Slack (1984a, b).</Paragraph>
    <Paragraph position="1"> In distributed memory systems items of information are encoded as k-element vectors, where each element ranges over some specified interval, such as \[-1 ,+1\]. Such vectors might correspond to the pattern of activity over a set of neurons, or the pattern of activation levels of a collection of processing units within a connectionist model. Irrespective of the manifest form of vectors, the information they encode can be manipulated by means of three basic operations; association (denoted by *), concatenation (+), and retrieval (#). The association operator creates a composite memory vector which encodes the association of two items of information (denoting memory vectors by angular brackets, the association of items &lt;A&gt; and &lt;B&gt; is denoted &lt;A&gt;*&lt;B&gt;).</Paragraph>
    <Paragraph position="2"> The concatenation operator creates a composite memory vector encoding individual items or vectors (&lt;A&gt;+&lt;B&gt;). Thus, a single composite vector can encode large numbers of items, and associations between items. The individual elements encoded on a composite vector are accessible through the retrieval operator. When a retrievalkey vector is input to a composite vector the retrieval operator produces a new composite vector which encodes those items of information which were associated with the key vector on the original composite trace. An important property of the retrieval and assocaition operators is that they are both distributive over concatenation. However, only the association operator is associative.</Paragraph>
    <Paragraph position="3">  These basic properties enable distributive memory systems to encode large amounts of knowledge on a single composite memory vector. The capacity limitations of such systems are determined by the noise levels implicit in the output of the retrieval operation. The generated noise is a function of the number of traces encoded on a vector and the size of the vector.</Paragraph>
    <Paragraph position="4"> Two classes of distributive memory can be distinguished, permanent and working. The former provide permanent storage of information, while the latter have no permanent contents but are used to build temporary representations of inputs. An important feature of working distributed memories is that the traces encoded on them decay. Associated with each vector encoded on ~t working memory is a decay, or strength, index. The decay function is event dependent, rather than time dependent, in that the indices are adjusted on the oceurence of a new input. This property provides a useful form of event indexing which is crucial to the chart parsing scheme.</Paragraph>
  </Section>
class="xml-element"></Paper>