Lecturer's Précis - Morton, Hammersley, and Bekerian (1985)
"Headed Records: A Model for Memory and its Failures"
Copyright Notice: This material was written and published in Wales by Derek J. Smith (Chartered Engineer). It forms part of a multifile e-learning resource, and subject only to acknowledging Derek J. Smith's rights under international copyright law to be identified as author may be freely downloaded and printed off in single complete copies solely for the purposes of private study and/or review. Commercial exploitation rights are reserved. The remote hyperlinks have been selected for the academic appropriacy of their contents; they were free of offensive and litigious content when selected, and will be periodically checked to have remained so. Copyright © 2010, High Tower Consultants Limited.
First published online 16:20 BST 18th August 2004, Copyright Derek J. Smith (Chartered Engineer). This version [HT.1 - transfer of copyright] dated 18:00 14th January 2010
Readers unfamiliar with the computer industry concepts of "record", "field", and "access key" should pre-read our e-tutorial on "IT Basic Concepts". Readers unfamiliar with the principles of "random access" to computer data files should pre-read Section 3.1 of our e-paper on "Short-Term Memory Subtypes in Computing and Artificial Intelligence" (Part 5). Alternatively, use the [glossary/tutorial] links as and when you come to them. Readers should also note that there is a very long-standing tradition within neuropsychology that names and their referents are more or less separately stored within the brain, requiring an associative link of some sort to be maintained between them. As a result, the act of generating thoughts appropriate to words is a major part of the human speech perception process, and the act of locating the best word(s) to express a given thought is a major part of the human speech production process [as described in detail in our e-paper on "Speech Errors, Speech Production Models, and Speech Pathology"].
1 - Introduction
The authors began this paper with a problem and a proposed solution. The problem was the unknown nature of the memory processes by which we are able to recognise a face and then somehow locate (or fail to locate) the personal name which belongs to it. In the authors' judgement, neither Anderson's (1976, 1983) associative network approach to memory organisation [for the background here, see Section 3.8 of our e-paper on "Short-Term Memory Subtypes in Computing and Artificial Intelligence" (Part 5)], nor any of the several schema-based [glossary] memory theories are particularly good at explaining how the cluster of declarative memories [glossary] which makes up much of our personal identity might be linked to the arbitrary attribute of our given name. Their solution was to strengthen the network approach by giving more thought to the possible mechanisms by which information is stored and accessed. A theoretical model was accordingly developed to explain how a number of separate "memory units" might act together via a system of "access keys" to deliver more complex memories. They defined the memory unit as "the unit of storage for recallable information in memory" (p6), and they saw the system working as follows .....
"We first supposed, following Norman and Bobrow (1979), that memory consists of discrete units each containing information relevant to an 'event', an event being, for example, a person or a personal experience. Information contained in a memory unit could take a number of forms, with no restrictions being placed on the way information is represented, on the amount being represented, or on the number of memory units that could contain the same nominal information. Attached to each of these memory units would be some kind of access key [and] two features would distinguish it from the memory unit. First, the contents of the access key would be in a different form to that of the memory unit, e.g. represented by a different code. Second, the contents of the access key would not be retrievable ....." (Morton, Hammersley, and Bekerian, 1985, pp3-4.)
They then explained the notions of memory unit and access key in greater detail, as summarised in the following two sections.
2 - Records as Units of Memory
To help make sense of the unit of memory issue, the authors proceeded to borrow the term "record" from computer science [tutorial] [it makes very good sense to do this, because computer science also sees the record as the logical unit of storage]. Guided by this powerful technical metaphor, the problems of biological memory thus become the problems of (1) parcelling up experience-to-be-remembered into appropriately sized Records, (2) storing these Records away physically in some convenient physiological substrate, and then (3) getting hold of them again whenever their content is needed. As to what that content might be .....
"Records could contain a variety of types of information ranging from lists or propositions to attributes of situational frequency or recency [citation] and can be at any level of abstraction. So, when one sees a film, one could have a record containing a précis of the plot while other Records may contain lower level information such as the physical details of one of the characters." (p6.) << AUTHOR'S NOTE: Computer scientists have much the same problem, and invariably have to accumulate different types of record, each contributing its own part to the whole. Indeed, perhaps the most critical stage of systems analysis lies in determining what those types and contributions actually are. The most important specific technique during the early stages of systems analysis is "Entity-Relationship Modelling" (ERM) [tutorial], because it provides a "logical" [that is to say, abstract and machine-independent] view of how the organisation in question operates. >>
The following five key "features" summarise the role played by Records in the model .....
(1) Discreteness: The first important feature of Records is that they are "independent of one another", that is to say, "no connections link Records that happen to be related in terms of their content" (p6).
(2) All-or-None Access: The second important feature of Records is that when you attempt to retrieve one you either succeed or fail. [Or to put it another way, if you can only retrieve part of what you want to know then it must have been stored away as separate records to begin with, not all of which have been retrieved.]
(3) Unlimited Content: The third important feature of Records is that there are no restrictions on "the amount of information they contain, nor on the format for information" (p6).
(4) Duplication Allowed: The fourth important feature of Records is that duplication of content is allowed. Specifically, "the same event or any of its constituents can be repeated in multiple Records" (p6).
(5) Unmodifiability: The fifth important feature of Records is that once information has been stored in one it cannot be changed.
3 - Headings as Access Keys
The authors then chose the word "heading" to refer to the corresponding access key, thus .....
"Headings form the means through which Records are accessed and are made up of a number of distinct elements. Remembering involves first of all searching the Headings with some information [until] a match is found. Only if a Heading is matched will the contents of its Record be accessible for recall." (p6.)
The following five key "features" summarise the role played by Headings in the model .....
(1) Format: The first important feature of Headings is that "the format of information contained in Headers is different from that in Records" (p6).
(2) Content: The second important feature of Headings is that their content "need bear no propositional relationship to the content of its Record" (p6).
(3) Discriminability: The third important feature of Headings is that the information they each contain will determine how easy it is to get hold of the associated Records.
(4) Unmodifiability: The fourth important feature of Headings is that once information has been stored in one it cannot be changed.
(5) Inaccessibility: The fifth important feature of Headings is that their contents "are inaccessible and can never be subject to recall" (p6).
The precise information by which the search process is directed will depend upon the nature of the act of remembering in question. A given memory search might be the result of an explicit question, for example, such as "When did you get to work today?", or it might arise as a goal or subgoal within a complex problem solving task, or it might be the entire problem solving procedure itself, or it might be part of the more general process of perception (as in finding the person details to go with a face you have just recognised). In any event, the search specification - that which you are looking for - is called the "description", and the important thing about a Description is that it "can comprise a number of independent fields" (p7). It is the skill of the searcher in putting together the most promising Description, and the subsequent interaction of that Description with the many available Headings which then determines the success or failure of the ensuing memory search, as now described .....
"The nature of the match required between the Description and a Heading will be a function of the type of information in the Description. If the task is to find the definition of a word or information on a named individual then a precise match may be required at least for the verbal part of the Description. We assume that the Headings are searched in parallel. On many occasions there will be more than one Heading that matches the Description. However, we require that only one Record be retrieved at a time [.....] and the data indicate that the more recent of two possible Records is retrieved." (pp7-8.)
ASIDE (1): The Internet was not available in 1985, of course, but it offers a highly convenient illustration of what the authors were proposing in their model. For a Record, we may think of a single web page [just like the 2000-word document you are currently looking at]. This Record comes with a unique worldwide "resource locator", namely its URL [which for the present document is http://www.smithsrisca.demon.co.uk/ PSYmortonetal1985.html]. If you know this resource locator, then you can key it into your browser software whereupon it will function as a Description. In fact, the Internet software then has to unbundle the URL into an http address component (the first half) and a file name component (the second half). However, the underlying telecommunications network does not use the http component for routing purposes, and converts it instead into a 999.999.999.999 format TCP/IP address [tutorial], and it is this address which enables enquiries to be routed to the correct server. Hence we have a Record with an access key we do not normally get to see (and which - being just a string of numbers - means little or nothing to us whenever we do get to see it), and a multi-component Description.
ASIDE (2): At the same time, the problem faced by search engines such as Google is that they need to cross-reference web enquiries with all the resource locators they have available to them, and they do this using a "keywording" system to maintain [user-invisible] master indexes. This allows end-users to construct Descriptions of variable specificity according to their needs. Thus if you search for the present paper using only on the first author's name, <Morton>, you get 3,420,000 hits [Google, 2nd August 2004], but only one of these will reference the present document. Since your target is statistically unlikely to be at the top of the display every time, it makes sense to add supplementary keywords to the Description, in order to make it more discriminating. When we add the second author's name, for example, and search for <Morton Hammersley> the number of hits reduces dramatically (to 2060), and when we add the third author's name and search for <Morton Hammersley Bekerian> it reduces again (to a much more manageable 62). Again we have a Record with an access key we do not normally get to see, accessed via a multi-component Description.
4 - The Official Worked Example
The authors then illustrated how their "headed record" memory system might operate in practice by working through a number of everyday memory encounters. The first of these was the "What was his name?" memory failure phenomenon .....
"Most people admit to having experienced the situation of feeling able to recount virtually everything they know about a particular individual except their name. One such incident recently occurred to two of the authors of this paper. Someone's research was being discussed. The main results were familiar to the people involved. We knew where the man worked, where he lived, the name of his wife, and the last time he had given a talk at the Applied Psychology Unit. But the person's name eluded us. We knew that we would be able to recognise his name, if produced, and also that we would have been able to reproduce all the information currently available to us had we previously just been given his name." (p2)
Now the neuropsychology of face perception has attracted much theoretical analysis in recent years, and a powerful person perception model has emerged. The lead workers in this area have been the University of Stirling's Vicky Bruce, Cardiff University's Haydn Ellis, and Lancaster University's Andrew W. Ellis. In common with other areas of neuropsychology such as language and number processing, the model proposes a modular parallel processing architecture in which the basic perceptual processes of facial recognition have to work alongside equally important (and potentially older in the phylogenetic sense) processes of "expression analysis" (Bruce and Young, 1986). Within this complex modular set-up, the basic act of recognition relies on "face recognition units", each a memory unit in the sense that phrase has been used above. [For a fuller introduction to this topic, see Ellis and Young (1988; Chapter 4). One would therefore be forgiven for thinking that a person's name - "Bill Smith" in the authors' central example - ought to be the access key to the person Record for that person. This would mean that the name node for <Bill Smith> would be accessible in its own right, and then capable of activating the corresponding person node(s) (we have to allow for plurals here, because two or more people might share the same name), in much the same way that the word "pen" can activate the corresponding object node for a particular type of writing instrument. However, such an arrangement would immediately be inconsistent with the authors' theoretical assertion that access keys should not themselves be retrievable, and so they proposed instead that the name was in fact stored twice, once in a memory unit of its own, and a second time in the irretrievable access key. For a retrieval request to work properly, the access key would be activated by the antecedent train of thought, and both name and person nodes duly triggered.
Other discussion examples were then used to explain the model's relevance to memory theory in general and to eyewitness testimony studies in particular.
5 - Evaluation
To understand why this paper is important, one must firstly understand that psychological theorists have actually been rather slow to import ideas from computer science and apply them to biological memory access .....
ASIDE: Concerned that psychological theory was being held back by this reluctance to explore what computer science had to offer, we included an impromptu list of the main imports and oversights in Smith (1997a). Among the successes we noted the working memory concept, clock speed, the hierarchical, networked, and modular nature of processing, the adoption of the dataflow diagram, and the notion of the semantic network. Amongst the features "more or less overlooked" by cognitive science, we noted little use of the program structure diagram or program flowchart, poor consideration of the role which might be played by interrupt handling in processing hierarchies, patchy exploration of database concepts (particularly schema, subschema, currency, and indexing), patchy use of data modelling principles or the logical-physical design dichotomy, negligible understanding of network protocol concepts (especially as they pertain to peer-to-peer communication), and poor consideration of the use of STM resources in the basic computational cycle. We have recently updated this list in Section 3 of our e-paper on "Short-Term Memory Subtypes in Computing and Artificial Intelligence" (Part 6), if interested.
Morton, Hammersley, and Bekerian should therefore be congratulated for having helped introduce record and access key concepts into psychology. For the implications of the headed records approach to the theory of eye-witness testimony click here or here, and for an introduction to the semantic network data storage mechanisms developed by the computer industry over the last half century, see the story of the "codasyl" breed of database in Section 3.6 of our e-paper on "Short-Term Memory Subtypes in Computing and Artificial Intelligence" (Part 6).
6 - References