Reder's SAC model has been successfully applied to a variety of memory-related data. Among these are data from feeling-of-knowing tasks (Reder & Schunn, 1996; Schunn, Reder, Nhouyvanisvong, Richards, & Stroffolino, 1997), mirror-effect data from remember/know tasks (Reder, Nhouyvanisvong, Schunn, Ayers, Angstadt, & Hiraki, 1998), data from negative priming tasks (Erickson & Reder, 1998a), and data from implicit memory tasks (Erickson & Reder, 1998b). Within the Reder lab, more studies are being undertaken that will provide data to be fit by SAC. Among these are an investigation of the conditions necessary to obtain the word-frequency mirror effect (Glanzer & Adams, 1985), and to obtain certain transfer appropriate processing (TAP) effects (e.g., Graf & Ryan, 1990; Morris, Bransford, & Franks, 1977)
SAC is a formal model of memory that attempts for broad coverage of a number of memory and cognitive phenomena. SAC's representation is a rather generic semantic network model of declarative memory (derived from, but not identical to, ACT-R&'s semantic memory) that consists of inter-associated concept nodes that vary in long-term strength. For example, there are nodes that represent letters (e.g., a, b, c), words (e.g. cat, climb, tree), sentences (e.g., the cat climbed the tree) or parts of the sentence, numbers (e.g., 5, 17, 31), operators (e.g., +, /, *), whole problems (e.g., 17+31), and events (e.g., reading this sentence). It is assumed that concepts such as cat have as constituents the letters that spell the word as well as semantic information and many other types of perceptual information. Although this model uses a localist, rather than a distributed representation, it is assumed that each concept is associated with a wide variety of features, a subset of which can activate the higher level node.
Figure 1. Example of problem representation used
by SAC
for the Feeling of Knowing experiment of Reder &
Ritter (1992).
Figure 1 illustrates how the nodes representing an arithmetic problem used in Reder and Ritter (1992) (modeled by Schunn et al., 1997) connect the operands and operators to its answer. Nodes that represent numbers may serve as operand nodes for some problems and answer nodes for other problems, (e.g., 31 is an operand in the problem 23 * 31 and is also the answer to 14 + 17).
The generic representational format of SAC gives it the ability to account for phenomena at many cognitive levels (e.g., memory for words, arithmetic operations, and problem-solving strategy). Because SAC has access to the activation values of its nodes, the model is well suited to making predictions concerning implicit memory experiments as well as predicting metacognitive judgments, such as Remember vs. Know judgments and feeling of knowing judgments. It is the detailed specification of how representations change from experience and how activation values are interpreted in particular situations that allows SAC to make specific, quantifiable predictions for many types of tasks.
Although SAC has been applied in such a wide variety of domains, herein I will only consider two of the projects with which I have been involved. The first deals with negative priming, and the second with implicit memory.
Negative priming occurs when an object or location that has been ignored in the past is now to be attended. It is characterized by responses to the object or location that are slowed or less accurate (e.g., Tipper, 1985).
In typical negative priming studies, participants switch from ignoring to attending to an item or location on consecutive displays. To permit investigation of the influence of memory processes on negative priming, we examined the effects of inserting large numbers of intervening displays between the presentation of an item when it was to be ignored (the prime trial) and the presentation of the same item when it was to be attended (the probe trial). Further, we examined the effect of having participants ignore an item repeatedly.
It is unclear what extant theories or models of negative priming would predict under these conditions inasmuch as they typically base their predictions on sensory inhibition. SAC combined with aspects of Logan's Instance Theory (1988), however, predicts that repeatedly ignoring an item should cause it to be progressively easy to ignore and that negative priming can endure over a large number of intervening trials.
In one of these negative priming studies, we used numeric stimuli and the task involved reporting an attribute of the font (bold or outline) of the lesser of the two numbers in the display. In one condition in this experiment, the greater (i.e., to-be-ignored) number was repeated 16 times. Each repetition was separated from the others by at least one trial, and the lesser (i.e., to-be-attended) number in each of these trials was chosen randomly. Nevertheless, repetition of the ignored numbers facilitated participants' responses as can be seen in the figure below. The squares indicate the mean of participants median response times, and Stimulus Types D1—D8 indicate pairs of repetitions (e.g., the first two repetitions are shown as Stimulus Type B2 for build-up 1 & 2). After having been ignored 16 times and at least five trials after having been presented as the greater number, this number was presented as the lesser, to-be attended number in a display. As can be seen in the figure as Trial Type DP (delayed probe) and as is typical in negative priming studies, participants responses were much slower in this case than in unprimed control trials (indicated as Trial Type Ctrl). In this same study, we replicated more typical negative priming results by presenting a number as the greater number on one trial and as the lesser number on the next. Participants' response times are shown below as Stimulus Types I and IP, respectively.
In sum, we found large negative priming effects even after substantial delays from the last negative prime trial (i.e. the sixteenth presentation of a number as the greater of the pair; the median delay was 10 trials, approximately 20 s). This finding is evidence against the common assumption that such effects only occur at relatively short delays (e.g., up to 6.6s in Tipper, Weaver, Cameron, Brehaut, & Bastedo, 1991), and a good mechanistic account of the negative priming must be able to explain it.
Figure 2.
The model for the numeric experiment is implemented in SAC using one node to represent each possible item that participants could see. The representation of this task is shown on the left in Figure 2. Associated with each number was an attend response and an ignore response (Logan, 1988). The model simulated each display by boosting the current activation of the two nodes representing the two numbers. Activation was then sent along links to the two pairs of response nodes (one pair for each number node).
Participants' responses were assumed to be facilitated by activation of the ignore node for the larger number and the attend node for the smaller number. Likewise their responses were assumed to be impeded by activation of the attend node for the larger number and the ignore node for the smaller number. To combine these activations into one value, facilitation for each number was computed by subtracting the activation of the impeding node from the facilitating node. Total facilitation was computed by summing the facilitation of the two numbers. If total facilitation was positive, SAC predicted a speedup on that trial above some baseline, and if total facilitation was negative SAC predicted a slowdown on that trial below the baseline. The baseline performance for each trial was predicted by the power-law of practice (Newell & Rosenbloom, 1981).
Independent of any predicted speedup or slowdown on a given trial, each of the two numbers would have been attended to or ignored. Thus, after each trial, the node and link strengths were updated to encode a memory of the appropriate responses. That is, the node strength of both numbers, the node strength of the attend node for the smaller number and of the ignore node for the larger number, and the link strengths between these nodes were all increased according to Equations 1 and 2. To simulate the data from the immediate negative priming trials, the ones in which the prime and probe trials were consecutive, SAC had to be augmented to update the short-term, current link strengths (for the same nodes that had their long-term link strengths updated) by a constant amount. This gives the links, the associations between chunks in memory, the same type of short-term activation patterns as the chunks themselves. These short-term link strengths then decayed exponentially just as the short-term, current node activation.
The model was fit to the data by allowing just one parameter, a multiplier that linearly scaled activation values to response time values, to vary freely. The model's fit is shown with the empirical data in the figure above. (The circles indicate the model's predictions and the squares indicate the empirical data.) The model successfully accounted for 90% of the variance in the data.
SAC posits that multiple repetitions of study items should yield better performance than single items. According to SAC, implicit memory effects are based on the activation of a concept (word) node alone, whereas explicit memory effects are based on the activation and association strengths between the word-node, an event (episodic) node that represents the event of studying a word on a particular list, and an experimental context node (see Figure 3). In an explicit retrieval task such as recognition, correct recognition of a word (a hit) can arise from two different processes, recollection and familiarity. A hit due to recollection is caused by activation of the event node that has been transmitted from the probed concept node and the experimental context node. A hit due to familiarity is caused by the activation of the probed concept node itself. The activation of the probed concept node, however, may be elevated for reasons unrelated to events within the experiment. For example, words with a high normative frequency tend to have elevated node activations. Thus, when probed, they seem familiar even though they have not been seen previously in the experiment. Recognition responses based on familiarity, therefore, are more susceptible to false alarms than are responses based on recollection.
In an implicit task, word-node activation alone determines responding unless explicit memories affect the process. According to SAC, repeated presentations (within a single list) should lead to better explicit and implicit memory because all the node activations increase according to the same principles with each presentation of a word, and all the node activations decrease according to the same principles as time elapses. Hence, SAC makes a straightforward prediction whereas the prior research is equivocal with regard to this prediction. I will briefly describe one study and how SAC was used to model its results.
Our principle goal was to test whether multiple repetitions strengthened implicit memory, and we wished to use words that had been repeated many times (e.g., 10 vs. 1, not 2 vs. 1). This experiment emerged serendipitously from two prior experiments completed 12 and 18 months earlier. Participants in the previous studies (Reder et al., 1998, Experiments 1 and 2) had been presented with a sequence of high- and low-frequency words in a continuous recognition paradigm wherein on the first presentation of a word the correct response was to identify the word as ``new,'' and on subsequent presentations of the same word, the correct response was to identify it as ``old.'' Thus, the participants were intentionally trying to encode the words as they were presented. Each subsequent presentation of a word was separated from the previous one by at least one different word, and each of the words in these studies was repeated 1, 2, 4, 6, or 11 times. We contacted the participants from these studies via email without revealing why they had been chosen to participate in the current experiment. Because the people with whom they had contact, the rooms in which the experiment was administered, and the format of the experiment were all different from the previous experiments in which words were encoded, we anticipated that it would be extremely unlikely that participants would relate the two studies and unlikely that they could use explicit strategies to recall words that had been presented more than a year earlier. Their task in the present experiment was to complete fragments made from words studied in the previous experiments (e.g., complete d_n_ _au_r to form "dinosaur"). Figure 4 shows the results of the study and the fit of SAC. There is a significant linear trend indicating that more repetitions did, indeed, strengthen implicit memory and that those effects were evident for up to 18 months. These results argue repeated presentations do have greater impact than single repetitions in implicit memory just as they do in explicit memory. Thus, repetition effects should not be used to argue in favor of separate implicit and explicit memory systems. We fit SAC to the data from this experiment to show how a model of memory with a single representation for implicit and explicit memory could account for these results. Although SAC contains a number of parameters, all of these were fixed based on values from previous studies except two (the retrieval threshold and the degree activation noise). These two parameters were fixed by first fitting SAC to the data from control participants who had no experimental exposure to the words. The parameters from that fit were then applied to the model to account for the data from the participants who had been exposed to the words 12 or 18 months previously. Thus, the fit shown has zero free parameters, yet its RMSD is only 0.026.Figure 3. Implicit and explicit memory in SAC. The circles represent nodes and the lines represent linking associations. The lines emanating from the concept node indicate associations with other concepts and with the perceptual properties of the concept. The lines emanating from the context node indicate links to other events during the experiment. Explicit memories are the result of retrieving an event when a concept and context are used as probes whereas implicit memories are a function of the concept-node activation alone.
Figure 4. Proportion of correct fragment completions by frequency. The diamonds indicate the participants' performance, and the squares show the model's predictions.
Erickson, M. A. & Reder, L. M. (1998a). The Influence of Repeated Presentations and Intervening Trials on Negative Priming. In M. A. Gernsbacher & S. J. Derry (Eds.), Proceedings of the Twentieth Annual Conference of the Cognitive Science Society, (p. 327-332). Mahwah, NJ: Erlbaum.
Erickson, M. A., & Reder, L. M. (1998b). More is better: The effects of multiple repetitions on implicit memort across long durations. Manuscript submitted for publication.
Glanzer, M., & Adams, J. K. (1985). The mirror effect in recognition memory. Memory & Cognition, 13, 8-20.
Graf, P., & Ryan, L. (1990) Transfer-appropriate processing for implicit and explicit memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 978-992.
Logan, G. D. (1988). Toward an instance theory of automatization. Psychological Review, 95, 492-527.
Morris, C. D., Bransford, J. D., & Franks, J. J. (1977) Levels of processing versus transfer appropriate processing. Journal of Verbal Learning & Verbal Behavior, 16, 519-533.
Newell, A., & Rosenbloom, P. S. (1981). Mechanisms of skill acquisition and the power law of practice. In Anderson, J. R. (Ed.), Cognitive Skills and their Acquisitions. Hillsdale, NJ: Erlbaum.
Reder, L. M., & Ritter, F. E. (1992). What determines initial feeling of knowing? Familiarity with question terms, not with the answer. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 435-451.
Reder, L. M., \& Schunn, C. D. (1996). Metacognition does not imply awareness: Strategy choice is governed by implicit learning and memory. In L. M. Reder (Ed.), Implicit memory and metacognition (pp. 45--77). Hillsdale, NJ: Erlbaum.
Reder, L. M., Nhouyvansivong, A., Schunn, C. D., Ayers, M. S., Angstadt, P. \& Hiraki, K. (1998). A mechanistic account of the mirror effect for word frequency: A computational model of Remember/Know judgments in a continuous recognition paradigm. Manuscript submitted for publication.
Schunn, C.D., Reder, L.M., Nhouyvanisvong, A., Richards, D.R \& Stroffolino, P. J. (1997). To calculate or not to calculate: A source activation confusion model of problem familiarity's role in strategy selection. Journal of Experimental Psychology: Learning, Memory, Cognition, 23(1), 3--29.
Tipper, S. P. (1985). The negative priming effect: Inhibitory priming by ignored objects. Quarterly Journal of Experimental Psychology: Human Experimental Psychology, 37A, 571-590.
Tipper, S. P., Weaver, B., Cameron, S.,; Brehaut, J. C., & Bastedo, J. (1991). Inhibitory mechanisms of attention in identification and localization tasks: Time course and disruption. Journal of Experimental Psychology: Learning, Memory, and Cognition, 681-692.
This research was supported by National Institute of Mental Heath Grant 1R01 MH52808-01, Office of Naval Research Grant N00014-95-1-0223, and Air Force Office of Scientific Research Grant F49620-97-1-0054.