Arndt, J. &
Hirshman, E. (1998). True and false recognition in MINERVA2: Explanations
from a global matching perspective. Journal of Memory and Language,
39, 371-391.
This paper uses the DRM paradigm to examine predictions
of MINERVA2, a global memory model.
The DRM paradigm involves presenting word lists in
which all the items in the list (e.g. bed, rest, nap, etc.) are associatively
related to a single non-presented critical lure (e.g. sleep).
A global memory model is one in which the probe item
(i.e. the item on the recognition test) is compared to all items in memory
MINERVA2
Concepts are represented in MINERVA2 as a list (or
vector) of features. The value of the feature can be +1, 0,
or -1. +1 means that when the concept occurs the feature tends to
occur, 0 means that the feature is unrelated to the concept, and -1 means
that when the concept occurs the feature tends not to occur. (e.g.
If bird is the concept the feature feathers might be given a value
of +1, fur might be given a value of -1).
When an item is studied its features are activated
imperfectly. In simulations of the model a learning parameter
provides the probability that any feature will be accurately encoded.
Obviously things like attention, motivation, speed of presentation should
be reflected in this learning parameter.
At test, the similarity between the probe item (i.e.
the item on the test) and each of the studied items is computed.
Similarity is computed in the following way:
S= S[(P*T)/N]
Start with the first feature. Multiply the value
of the probe (P) on that feature and the memory trace (T) on that feature.
The product can equal 1, -1, or 0.
Repeat this process for all the features.
Add up all of these values
Divide by the number of features where neither the
probe nor the trace equals zero.
Note: This is a pretty common sense definition
of similarity. Similarity will tend to be high when the probe and
the trace share a lot of the same features and low when they don't.
The activation level of the memory trace is the cube
of how similar it is to the probe.
A = S3
Now here's the global part. The activation
values are summed across all items that occured in a particular
episode. This value is called the echo intensity and it is
influenced not just by the target item in memory but also by how well the
probe matches the overall gist of the list.
I = S A
Simulations: To generate predictions Arndt
& Hirshman first ran computer simulations of the model to see how independent
variables should influence the output.
First they generated some hypothetical memory representations
for 72 prototypes. These are equivalent to the Critical Lures in
the DRM paradigm. They did this by creating feature lists (i.e. vectors)
that contained 24 features. Each prototype was created by randomly
deciding for each feature whether it would be +1,0,-1.
Based on the prototypes they then created the targets
from the lists. They did this by replacing some of the features in
the prototype with new features (turning a +1 into a -1 or a 0).
The level of distortion is the probability that any feature of the prototype
will be changed.
Conceptually, the level of distortion tells you how
similar the targets are to the critical lure. So when distortion
is high (e.g. 0.85) the targets are not as strongly related to the critical
lure as when distortion is low (e.g. .60).
Ran 1152 simulation subjects in each condition at
each level of distortion. The decision criterion was set at the median
echo intensity for these simulated subjects.
Analyses are performed in terms of d'. In this
case d' is based on studied targets versus unstudied targets and another
d' is based on critical lures from studied lists versus nonstudied lists.
Model Parameters Manipulated & Examined Empirically
Learning rate: This is the probability that
a feature will be accurately encoded at study. The model predicts
that at high learning rates d' for targets should be greater than d' for
critical lures. This effect will be magnified the more disimilar
(on average) targets are from the CL (i.e. at high levels of distortion).
Arndt & Hirschman manipulated this experimentally
by varying the presentation rate (300, 500, 800, 1500, 3000 msec).
Consistent with the simulation, d' for both targets and CLs increased with
longer exposure durations but that d' for the targets exceeded d' for CLs
at the longest presentation rate.
Number of items: Decreasing the number of
items per list should decrease false recognition but not true recognition.
This is because true recognition is produced primarily by a very large
match between the probe and the memory trace but false recognition is produced
primarily by a number of small matches to everything in memory. The
results generally matched the prediction, however, both CL and Target recognition
decreased when the length of the study list was decreased.
Associative Strength: Associative strength
was modeled by assuming a greater learning rate (low frequency words) with
greater distortion from prototype. Prediction of model is that low
associate exemplars would be recognized more but would lead to less recognition
of CLs(Table 1). These predictions were confirmed (Figure 7).
General Findings
High learning rates lead to better discriimination
between targets and exemplars.
Decreasing the amount of associative information
improves memory for exemplars but decreases recognition of prototypes
The ability to discriimate between targets and related
lures is based upon the cubing process in the model which requires a high
match before similarity adds much to the echo intensity. This results
in a model of memory that is global but that can be influenced by sufficiently
strong contextual information (a very rational way for memory to operate).