The Use of Our Language in Higher Cognition
and Other Tasks
Kashima has provided a demonstration of how our language can be used to
propose hypotheses about some tasks in social cognition. We think
that this demonstration is especially useful because social cognition has
traditionally viewed memory as simply involving trace access and has
relied heavily on search (e.g., Srull & Wyer, 1989). To be useful
our functions only have to be more similar to the algorithmic-level
mechanisms than are the functions currently employed in social cognition.
Similarly, Murnane has shown how our language is flexible enough to
extend to more complex tasks involving context. Sloman has
suggested, however, that it would be as easy to use an algorithmic-level
language (also see Lewis).
Sloman's point about the use of an
algorithmic-level language is predicated on the observation that most of
the difficulty involved in extending our specifications to new tasks lies
in the identification of the knowledge and actions needed to perform the
task. These are indeed the difficult parts of the task and it is
also true that once you have identified the knowledge and actions required
you could use any algorithmic-level language that computes our functions.
This use of an algorithmic-level language to describe tasks is
actually very close to what Humphreys, et al. (1989b) attempted and
there are pitfalls to this approach. It is very difficult for some
modellers to accept that algorithmic-level ideas can be used to describe a
task. That is, they have a tendency to reject the approach unless
"you show that your model actually fits the data." There is merit in
fitting data but there are problems. If your purpose is to point
out similarities and differences across a large number of tasks, formal
modelling can be premature. That is, in many cases systematic
experiments comparing the tasks should follow the identification of
similarities and differences and precede a model fitting effort.
Formal modeling can also get in the way of effective communication.
That is, the number of arbitrary assumptions that are needed to actually
fit data can detract from the argument over the identification of the
knowledge and actions required. An illustration of this problem
occurs in Lindsay's (1991) comment on Metcalfe (1990). Metcalfe had
provided an identification of the knowledge and actions involved in the
misleading postevent paradigm. She had also shown that her model
(CHARM) could provide a reasonable fit to the data. In his
commentary Lindsay invoked Pike's (1984) criticism of the mathematics of
CHARM. This completely ignored the fact that, for the purpose of
describing the knowledge and actions required, Pike's and Metcalfe's
mathematics are interchangeable. The knowledge and actions required
could also be described in the language of the target article and this
might better serve to focus the debate.
Kinoshita has compared the use of our language with the language of an
algorithmic-level theory. Not surprisingly she found our language
was inadequate. We have never claimed that our language could
substitute for an algorithmic-level language and our position - like
Marr's - is that theories at all three levels are required (computational,
algorithmic, and implementational). What our language can do is to:
a) help determine if Kinoshita and the other researchers in the area are
asking all of the right questions, b) suggest alternative explanations for
experimental results, and c) indicate the potential relevance of
results.
One cannot conclude that an ceptual fluency (i. e.,
an ahistoric source of information) is involved in single item recognition
without varying the episodes in which different items have occurred and
using instructions to direct subjects to particular episodes. With
rare exceptions (Parkin, Leng, & Hunkin, 1990; Jacoby, 1991) researchers
who have addressed Kinoshita's concerns have not used list- specific
tasks. Thus, they are not asking all of the right questions.
Kinoshita also assumes that a masked identity prime serves to increase
perceptual fluency. Presenting a masked identity prime just prior
to the presentation of a to-be-recognized word is, however, the same
sequence of events as occurs in JPTS. Our specifications for LSIR
and JPTS suggest an alternative to the perceptual fluency interpretation.
That is, they suggest that subjects are computing the intersection
between the to-be-recognized word and the words in the study list in order
to recognize a word. If the to-be-recognized word did not occur in
the study list or if the association with context was not learned then the
intersection will be empty. Under these conditions the presence of
a masked identity prime may result in a nonempty intersection just as it
is assumed to do in JPTS (Humphreys & Bain, 1992). Some support for
this prediction comes from Bernstein and Welch's (1991) observation of a
very high correlation between similarity and recognition
judgments.
As an example of how our theory can identify potentially relevant results,
Jacoby (1983) had subjects study a long list of words and then gave them a
perceptual identification test. When the study list was made
salient via instructions or the use of a short retention interval, the
probability of the correct identification of an old word increased and the
probability of the correct identification of a new word decreased.
This pattern of facilitation and inhibition is the same pattern that is
found when direct (explicit) and indirect (implicit) retrieval
instructions are used with partword cues (Roediger, et al., 1992).
This connection may not have been noticed because Roediger et al.
thought of their results as arising from a generation/recognition process
instead of a more generic intersection process. Jacoby's (1983)
results are unlikely to arise from a generate-recognize process but they
strongly suggest that subjects are combining information from two sources
(perceptual information from the target and list-membership information)
in a manner approximating an intersection. This observation raises
several important questions including: a) Is the use of list-membership
information in this task in any sense deliberate? and b) In what way is
the use of list-membership information when the subject has stayed in the
same setting performing the same task, similar to the use of
list-membership information invoked by instructions?
Oscar-Berman questioned whether our theory could help in understanding
anterograde amnesia. What it provides here is a language for
describing hypotheses about the knowledge and actions needed to perform
tasks. For example, in Humphreys and Dennis' (1994) commentary on
Eichenbaum et al. (1994) our language was used to propose
hypotheses about how some animal memory tasks were being performed.
Previously Humphreys, et al. (1989a) had used the algorithmic-level
language of Humphreys, et al. (1989b) to propose hypotheses about a
wider variety of tasks. Because our language was designed to
subsume the Humphreys, et al. (1989b) language, amongst many other
algorithmic-level languages, it will be relatively straightforward to
translate the Humphreys, et al. (1989a) hypotheses into our
language. It should be apparent that our language is far more
complete, powerful, and precise than is the language of associative
learning, familiarity, configural memory, declarative memory,
etc.
Our analysis also suggests that Oscar-Berman and other
researchers are not asking all of the right questions. For example,
in section 2.5 we showed how a task (AB ABr learning) that requires a
three-way binding can be performed using two-way bindings and a series of
retrievals. This procedure is an unlikely way to perform AB ABr
learning, given the ease with which untrained subjects perform this task.
However, the possibility that hippocampectimized rats are
performing some tasks, which require three-way bindings, by using two-way
bindings and a series of retrievals needs to be explored (Humphreys &
Dennis, 1994).
Miscellaneous Issues and Misconceptions
Clarke draws a parallel between the bottom up versus top down approach and
the split in cognitive science between Searle (1992) and Dennett (1991).
Ours and Marr's advocacy of the importance of a computational-level
theory does not presuppose an acceptance of Dennett's position.
Even in a strong specification the functions are simply abstract
characterizations of the mechanisms/processes that need to be implemented.
Such important issues as the type and number of errors and possibly
even subjective experiences will emerge from the details of the
implementation or even of the neural substrate.
Lewis has questioned whether our distinction between tasks that
functionally do and do not require an episodic input maps onto the
implicit/explicit distinction of Graf and Schacter (1985). Our
distinction directly maps onto the Dunn and Kirsner (1989) distinction
between implicit and explicit tasks. To map onto the Graf and
Schacter distinction we have to assume, as they do, that subject supplied
instructions can turn an implicit task into an explicit task.
Neither our approach nor the Graf and Schacter approach can unambiguously
classify all tasks without some additional assumptions.
Murnane suggests that we are caught in a "bind" because a three-way
binding does not suffice for nested context problems. A three-way
binding sufficed for the task we analyzed and we designed our language so
that it could be extended to arbitrarily complex problems. There is
no "bind" here as Murnane's extension of our analysis shows.
Murray drew a parrallel between our classification scheme and the
classification of psychophysical methods in accordance with the number of
stimuli involved. Yet, there is only a superficial resemblance, as
one of our crucial distinctions between AB ABr learning and CREA involves
the same number of stimuli. In addition, modern psychophysical
theories may assume that the same processes underlie the different
methods. Modern memory theories, however, assume that different
processes underlie memory access with partword and word cues.
In arguing for the inclusion of a filtering process in a
computational-level theory Murray implies that we are assuming perfect
storage. In order to correctly perform a task such as CREA a
subject must recall a word from the list that is related to the cue.
We have described the complexity of the representation that is
required to correctly perform this task. In addition, in describing
the input/output mapping, we have used a language that suggests
similarities and differences between this task and other tasks and may
even provide a rough description of some of the mechanisms and processes
involved. That is, we are trying to describe the components which
are necessary in order to correctly perform the task. If a
representation of sufficient complexity is not available (i.e., if storage
has failed) the task will not be performed correctly.
Tiberghien questions the autonomy of the computational level because the
alternative specifications for LSIR could have arisen from implicit
algorithmic-level hypotheses. This question reflects the tension
that exists in our article between thinking of our specifications as weak
or strong. As a weak specification the alternatives are simply an
allowable rearrangement of a formal mathematical expression. As a
strong specification they are alternative algorithmic-level hypotheses.
Their strength in the latter role comes from the fact that our
language, unlike Tiberghien's language, does not make a commitment to
particular mechanisms (context as a tag) and processes (search).
There are, however, some algorithmic-level hypotheses we cannot express at
the computational level. For example, we could not express the
hypothesis that the degree of similarity between the study and test
contexts was playing a crucial role (Tiberghien expresses this idea as an
association between the list context and the test context). Some
hypotheses must be expressed at the algorithmic level. This
requirement means that they must be expressed in terms of a specific form
of representation and in terms of specific processes. As a
consequence the hypothesis will not generalize to other processes and
other forms of representation. The strength of an hypothesis stated
in our language is that it generalizes to different forms of
representation and different processes. The weakness is that we
lose the fine detail.
Van der Velde et al. argue that our memory access functions are
based on the notion that bounded sets exist. In fact, our preferred
method of implementation is to represent a set as a composite vector where
the elements of the set vary in terms of the strength by which they are
included. Again this is an issue for the algorithmic level and is
not an assumption of our approach.