Thesis: Integrating Learning into Models of Human Memory

[Simon Dennis]

FTP-host: archive.cis.ohio-state.edu
FTP-file: pub/neuroprose/Thesis/dennis.thesis.ps.Z

The file dennis.thesis.ps.Z is now available for copying from the
Neuroprose archive from the directory pub/neuroprose/Thesis.


	 The Integration of Learning into Models of Human Memory
				   
			     Simon Dennis
				
			     Ph.D. Thesis
		    Department of Computer Science
		       University of Queensland


Since memory was first distinguished as a separate phenomenon from
learning (Melton, 1963), researchers in the area have concentrated on
the memory component. Mathematical models, such as SAM (Raaijmakers &
Shiffrin, 1981; Gillund & Shiffrin, 1984), TODAM (Murdock, 1982), CHARM
(Eich, 1982), Minerva II (Hintzman, 1984) and the matrix model (Pike,
1984; Humphreys, Bain & Pike, 1989), have focussed on the mechanisms of
encoding, storage and retrieval. The affects of variables such as
retention time, number of presentations, spacing of presentations, type
of retrieval test, nature of cues, encoding paradigm and the extent to
which the study context is specified in the test instructions have been
studied empirically and modelled.

The learning component, which was the focus of the field for much of
this century, has received less attention in recent years (Estes,
1991). Despite the extensive empirical database on learning phenomena
(Postman, Burns & Hasher, 1970), attempts to model this data have been
few.  The attempts that do exist have concentrated on specifying
algorithms by which experience might tune the parameters of existing
memory models (Murdock, 1987) rather than attempting to explain how
learning induces the representations, decision criteria and control
processes of memory in the first instance.

How people acquire these components of the memory system has important
ramifications for the study of human retention.  One of the most
critical of these ramifications is the nature of the relationship
between the environment and the mechanism of memory. Recent empirical
work on the the environment of memory has revealed a striking
correspondence between the structure of the environment and the pattern
of performance in human subjects (Anderson & Schooler, 1991).  This
thesis extends this work by studying the environment empirically,
developing a learning mechanism and demonstrating that this learning
mechanism behaves in a qualitatively similar fashion to human subjects
when exposed to an environment that mirrors that with which the
subjects contend.

Analyses of the relevant environments of two touchstone phenomena: the
list strength effect in recognition and the word frequency effect in
recognition were performed to establish the context in which
interactive accounts of these phenomena must be set. It was found that
while low frequency words occur less often than high frequency words,
they are more likely to recur within a context. In addition, the
probability of recurrence was found to increase if a word had occurred
frequently in the current context, but was not affected by the amount
of repetition of words other than the target word.

A learning or interactive model of human memory called the Hebbian
Recurrent Network (HRN) has been developed. The HRN integrates work in
the mathematical modelling of memory with that in error correcting
connectionist networks by incorporating the matrix model (Pike, 1984;
Humphreys, Bain & Pike, 1989) into the Simple Recurrent Network (SRN,
Elman, 1989; Elman, 1990).  The result is an architecture which has the
desirable memory characteristics of the matrix model such as low
interference and massive generalisation but which is able to learn
appropriate encodings for items, decision criteria and the control
functions of memory which have traditionally been chosen a priori in
the mathematical memory literature.  Simulations demonstrate that the
HRN is well suited to the recognition task.  When compared with the
SRN, the HRN is able to learn longer lists, generalises from smaller
training sets, and is not degraded significantly by increasing the
vocabulary size.

To demonstrate that the HRN learning mechanism is capable of addressing
experimental behaviour, the phenomena studied environmentally were
modelled with the HRN. The HRN showed a low frequency word advantage
when it was presented with an environment in which high frequency words
occurred more often, but low frequency words were more likely to recur
within a context. In addition, the HRN showed a null list strength
effect while retaining the list length and item strength effects when
exposed to an environment in which the environmental results were
embedded.

By incorporating a learning mechanism and examining the environment in
which memory models are situated it is possible to produce models that:
(1) can start to address developmental phenomena; (2) can provide a
mechanism to address learning-to-learn phenomena; (3) can address how
internal states attain their meanings; (4) are easily extended to a
wide variety of cognitive phenomena; and (5) account for the striking
similarity between the environmental demands placed upon the memory
system and the performance of human subjects.

---------------------------

The thesis is 218 pages (22 preamble + 196 text).



Simon Dennis      Department of Psychology       mav at psych.psy.uq.oz.au
Post Doctoral     The University of Queensland
Research Fellow   QLD 4072 Australia