Kenneth J. Malmberg, PhD Department of Psychology University of South Florida |
An Annotated Reading List for Recognition Memory This is an annotated list of readings on the topic of recognition memory. It is primarily designed for my students, but it should also be useful to others who want a primer. By now, there are a large number of articles and chapters published on the topic and this reading list will perhaps provide you with a nonrandom entry into this fascinating literature. This is a work in progress; so please check back. Recognition Procedures Recognition is the discrimination of events that one experienced from events that one has not experienced. At first glance, this might seem to be a rather mundane task. But when one appreciates the variety of different forms of recognition memory tasks and the fact that a given stimulus might have been encountered in thousands of different prior contexts, recognition memory is a remarkable faculty indeed. For instance, I might ask if you if you saw your department's Chair yesterday. Presumably, you can fairly accurately determine whether this is so even when you have probably encountered this person many, many times in highly similar contexts. Understanding the ability to know what you have experienced is central to our understanding of human memory. There are three basic varieties of recognition tasks. One might be presented a stimulus, for instance, and be asked if it was encountered in specific context. This is referred to as a yes-no or old-new task. A closely related task is a rating task, whereby subjects are asked to provide a scalar judgment that represents his or confidence that a stimulus was studied. Some of these ratings are usually associated with a “yes” response and the remaining associated with “no” response, and each represents a different degree of confidence that the stimulus was studied. These can be related to a true-false test, in that the subject is asked to judge whether it is true or false an item occurred or how confident the subject is that an item occurred. A different form of recognition test is a multiple choice test. Accordingly, one might be shown two stimuli and asked which one was presented in a specific context. Psychologists referred to this as a two-alternative forced-choice task. Recognition accuracy for the yes-no and ratings tasks is a positive function of the difference between the probability of responding “yes” to previously encountered items versus the probability of responding “yes” to item not previously encountered. These are referred to as hit rates and false-alarm rates , respectively. In the later case, recognition accuracy is simply the probability of choosing the correct alternative. The Relevance of Recognition Memory Research The nature of memory for prior occurrences has been investigated for many years. Some of the earliest published papers on recognition memory are:
However, the earliest days of human memory research paid little attention to the recognition task because the dominate paradigm utilized recall tasks to assess how interference affected memory for past events, how one learned sequences of events, etc. The investigation of recognition memory increased in popularity in the late 1960's, towards the end of the “verbal learning” heyday and towards the beginning of the “cognitive revolution” in psychology. Today, investigations of recognition memory are commonly reported in the literature, and the empirical results that they report have proven to place strong constraints on theories of memory. One of the reasons for the relatively recent uptake in interest in recognition memory is that recognition was viewed by many as a simpler task than, say, serial recall, pair-associate recall, or free recall. That is, researchers following a reductionist approach believed that more specific questions about the nature of memory could be answered by focusing on what they believed was a simpler task (cf. Crowder, 1976). The assumption that recognition is a simpler task was based on at least two related observations. First, recognition almost always is more accurate than recall. Second, accurate recognition does not necessarily require the production of specific information from memory and recall does. We now know that these observations do not hold in all situations, and our motivations for investigating recognition memory are somewhat different today. Indeed, the investigation of recognition memory is more popular than ever. A recent search of PsychInfo with the keywords “recognition memory” turned 4562 unique articles published since 2000. Why is the investigation of recognition memory more popular than ever? The current massive popularity of recognition investigations is in no small part attributable to the fact that the recognition task in its variety of forms is amenable to the methodologies used to relate brain activity to behavior. Investigating the neurological basis of recall performance is, particularly during retrieval, much more difficult utilizing fMRI, PET, or EEG methods. Another reason that recognition memory is popular topic is the finding that recognition memory does not decrease in accuracy with age. Hence, understanding the nature of recognition might hold the keys to understanding the relationship between the brain and the mind and the relationship between age and memory. Here are references for several recent reviews of the literature on recognition memory:
Recognition memory occupies a prominent place in the current zeitgeist of memory research. Before discussing the current state of the science, it is worth reviewing the history of this field in order to better understand its current state better. Early researchers thought that investigations of recognition memory might help them to develop and test their models of recall, but the relevance of many early recognition memory findings was questioned when it was quickly discovered that recognition and recall were affected in different ways by variety of factors. It was clear that recognition was more than just a simple version of a recall task, and these interactions between operational factors and memory tasks have subsequently been a primary source of motivation for many who seek a better understanding of episodic memory. Here is but a short list of articles that address the distinction between recognition and recall. Emotion
Context variability
Word-frequency
Mnemonic Organization
Aging
Serial Position
Amnesia
Alcohol
Recognition as Detection of an Internal Signal Embedded in Noise Findings that show that a variable differentially affects recognition and recall provide a rich empirical basis for evaluating theories of memory. Explaining these interactions is at the heart of memory research. Perhaps the greatest achievement in memory research has been the development of formal models. This is especially true of those theories that attempt to explain the performance of variety of different task within a single theoretical framework. The models developed within such frameworks are often referred to as global memory models . The first formal models of human memory had a much more limited scope. During the early years of recognition memory research, several researchers developed models that were based on signal detection theory (Green & Swets, 1966; Macmillan & Creelman, 1991). Whereas recall was conceived as a threshold-like process (Krantz, 1969), recognition was assumed to be based on continuous random variable, which was often conceptualized as the strength or familiarity of the test stimulus. These papers are representative of this discussion in the early literature on recognition memory.
Global-Matching Models Signal-detection models made at least two significant contributions to our understanding of recognition memory. The first is an empirical contribution: Signal detection provided a means for independently measuring the sensitivity and the bias of recognition memory performance. Signal detection also provided a framework that considered the basis of a detection task to be evidence that varied along a continuous dimension. In recognition memory research, this random variable is typically conceptualized as item familiarity or memory strength. Within such a framework, a comprehensive understanding of recognition requires a way describing how the familiarity associated with a stimulus is generated. Signal detection models are not, alas, models of familiarity. A highly influential class of familiarity models is global-matching models. There are a wide variety of global-matching models. However, they have several common elements. Global-matching models assume that a memory trace is stored for each stimulus studied. These traces may be either holistic or multidimensional representations of the stimulus and the context in which it occurred. At test, global memory models assume that a temporary representation of the test stimulus is constructed. This is referred to as a retrieval cue . The retrieval cue is used to probe memory. The probe consists of a comparison of the retrieval cue to contents of memory, which is assumed to consist of at least those traces stored during study, and perhaps many more. The match between each trace and the retrieval cue contributes to the outcome of the global-matching process, a scalar value, which represents how familiar the retrieval cue seems to be. This value is then compared to a subjectively set criterion. If the familiarity value exceeds the criterion, the recognition response is “yes”; otherwise it is “no”. Generally speaking, the more similar a retrieval cue is to the contents of memory the greater its familiarity value will be. Here are references for several first-generation global-matching models:
The Effect of Interference in a Global-Matching Framework A watershed moment in our understanding of recognition memory occurred in the late 1980's and early 1990's when researchers noticed that the global-memory models predicted that recognition memory for a given item would be negatively affected to a greater degree when it was studied with other more strongly encoded items than when studied with other more weakly encoded items. This prediction was investigated using a mixed-list pure-list paradigm, whereby subjects studied lists comprised of all strong or all weak items and the performance on these lists was compared to the performance for similar items studied on lists comprised both strong and weak items. This is now referred to as list-strength manipulation, and list strength is another manipulation that has different effects on recognition and recall. Whereas many global memory models predicted that strong items would produce stronger interference than weak items for both recognition and free recall, this prediction was only confirmed for free recall. Here are several critical articles on the list-strength effect:
Receiver Operating Characteristic (ROC) Analyses The basis for the predicted list-strength effect for recognition memory is the following assumption: Increasing how well an item is encoded results in an increase in the variability of its contribution to the stimulus familiarity when memory is probed with a new item or foil. This is easy to understand at an intuitive level. Assume at the limit that an item was studied, but almost nothing was stored in memory to represent that event. At test, a foil will never match that trace very well, and hence there is very little variability in the global-match strength of the foils. On the other hand, consider a situation in which an item is studied and a great deal of information is stored about that event. While some foils will not match this trace well, others will match it well, and therefore there is greater variability in the global-match strength of foils when items are well encoded. Many separate-trace global-matching models made this prediction. The relative variances of the target and foil distributions can be measured within the framework of signal detection by constructing a receiver operating characteristic function (ROC). The z -transformed receiver operating characteristic ( z ROC) will equal 1.0 if the variances are the same. For a review of how to construct ROC curves see:
Because many global-matching models predicted that an increase in the strength with which items are encoded would affect the variance of foil familiarity distribution, they predicted that the slope of the z ROC would be systematically affected by a list-strength manipulation. Several important articles on the form of the recognition memory ROC were published. They established that slope of recognition memory zROC is less than unity, it is affected by item strength, but it is not affected by list-strength. As a package these findings were difficult for many global-matching models to handle, and the importance of these ROC analyses to the future direction of recognition memory research cannot by over estimated.
Mirror Effects An equally important discovery in the history of recognition memory research is that recognition memory often improves in a fashion that hits rates (HR) increase and false-alarm rates (FAR) decrease. This pattern of data is often referred to as a mirror effect because the changes in the means of the underlying familiarity distributions mirror each other along the decision axis. Perhaps the most well known mirror effect is the word-frequency effect. Accordingly low-frequency (LF) words are better recognized than high-frequency (HF) words. For the yes-no ratings task the following pattern of hit rates and false-alarm rates constitutes a mirror effect: FAR (LF) < FAR(HF) < HR(HF) < HR(LF). For the two-alternative forced choice task, this the mirror effect: P(HF-old, HF-new) < P(HF-old, LF-new), (R1) P(LF-old, HF-new) < P(LF-old, LF-new), P(HF-old, HF-new) < P(LF-old, HF-new), and P(HF-old, LF-new) < P(LF-old, LF-new). On null-comparison test trials, two foils or two targets from different stimulus classes are presented, and this is a mirror-patterned WFE for the two null comparisons: P(LF-old, HF-old) and P(HF-new, LF-new) > .50. (R2) Thus, old words are chosen more often than new words, HF-new words are chosen more often than LF-new words, and LF-old words are chosen more often than HF-old words. The word-frequency was problematic for many global-matching models because if they assumed that LF words were better encoded than HF words, then they either had to assume that both LF targets and foils would be more familiar than HF target and foils (respectively) or they would have assume that LF and HF foils would be equally familiar. In neither case, would a mirror effect be predicted. Because mirror effects were theoretically important they extensively investigated, most notable by Murray Glanzer and his colleagues.
Bayesian Models of Recognition Memory The list-strength effect, the slope of recognition memory z ROC, and mirror effects challenged extant global matching models. The response of the field to these challenges was nothing less that revolutionary. Several new global matching models were designed to account for the findings that the prior models could not explain. While each member of the class is different, the new class of recognition models was motivated by the assumption that the recognition memory system had evolved over time operating in an optimal or efficient manner. This assumption of the new generation of models reflects a recent movement towards rational models of cognition ( Anderson , 1990). The formalism used to implement the adaptive memory assumption was based on Bayesian definitions of optimality. Within this framework questions are often posed in terms of how should recognition memory and followed with questions about how recognition actually works. A major theoretical assumption of the adaptive memory assumption is that increasing the extent of encoding enhances the contribution of a target to trace to global-match strength and decreases the contribution of a foil trace to global-match strength. That is, increasing the completeness and/or the accuracy with which an item is represented results in a trace that is more similar to the studied item and less similar to the traces representing other studied items. Based on the matching of a retrieval cue to contents of memory comprised of traces formed according to this specification, an optimal decision is made based on prior information about how the recognition system works and posterior information about how the cue matches the contents of memory.
Dual-process Models of Recognition Memory So far, we have only considered familiarity-based models of recognition memory. While they have been very successful in both organizing data and motivating new research, they are often criticized as being overly simplistic. It is intuitively obvious to most people that recognition can occur based on the familiarity of a stimulus but also based on the recollection that a stimulus was experienced in a particular context. The distinction between familiarity-based recognition and recollection-based recognition was perhaps best analogized by Mandler (1980): “Consider seeing a man on a bus whom you are sure that you have seen before; you “know” him in that sense. Such a recognition is usually followed by a search process asking, in effect, Where could I know him from? Who is he? The search process generates likely contexts (Do I know him from work; is he is a movie star, a TV commentator, the milkman?) Eventually the search may end with the insight, That's the butcher from the supermarket!” [p. 252-253] Models that provide an opportunity for the subject to base his decision on either the familiarity of the stimulus or a recollection of the occurrence of a stimulus are often referred to as dual-process models. (Single-process recollection models are rarely proposed, primarily due to the interactions between the performance of recognition and recall tasks that we discussed earlier; but see Yonalinas, 1999; Diller, Nobel, & Shiffrin, 2001 for notable exceptions). Here some the most prominent dual-process models of recognition:
Distinguishing between Single-Process and Dual-Process Models One of the most hotly contested issues in the cognitive psychology literature is whether single-process or dual-process models of recognition are the preferred way to characterize recognition memory. The distinction between single- and dual-process models is whether multiple sources of information can be the basis of a recognition judgment. Sometimes this distinction is rather blurry. For instance, Rotello et al.'s STREAK model assume that global and specific sources of evidence are independent, but that the recognition decision is always based on a weighted combination of them. Thus, multiple sources of information factor into the recognition decision, but the subject presumably is unable to distinguish among them. From the vantage of philosophy of science is critical to note that the single-process model is a special case of the more complex dual-process model. The dual-process model reverts to a single-process model when it is assumed that recollection is not important factor in the recognition decision. Thus, the single-process model is to be preferred unless it can be shown to be insufficient because there is no need to develop overly complicated models to organize out observations.
The Retrieval Dynamics of Recognition Memory In 1973, Reed introduced a method for measuring the accuracy of a binary response at various latencies subsequent to the presentation of a stimulus. Accordingly, subjects are prompted to make a decision within a small (usually around 200 ms) experimentally controlled time window. This procedure has become known as the signal-to-respond procedure and it generates a function relating the latency of responses and the accuracy of those responses. While Ratcliff (1978) might have been the first to utilize the signal-to-respond procedure to investigate recognition memory, Barbara Dosher (1981, 1984) pioneered its use. The findings of these experiments showed that subjects were less likely to reject semantically related intact and rearranged pairs than semantically unrelated intact and rearranged pairs. This suggests that pre-experimental information (i.e., semantic associations) contributes to associative recognition decisions. Importantly, subjects were also more likely to reject unrelated rearranged pairs when each of the items comprising that pair was studied as part of a semantically related pair. Dosher (1984) proposed that subjects were better able to reject the pairs whose constituent items were studied as part of a semantically related pair, because they were able to recall that at least one member of the pair was studied with a different word. The rejection of foils based on recalling some aspect of a study event is now referred to as recall-to-reject (Rotello & Heit, 1999; Rotello, Macmillan, & VanTessel, 2001; Malmberg, Holden, & Shiffrin, 2004). Gronlund & Ratcliff (1989; also see Nobel & Shiffrin, 2001 and Rotello & Heit, 2000) compared the retrieval dynamics of single-item recognition and associative recognition. Subjects studied pair of words and single words. At test, they were presented with intact pairs, single-word targets, rearranged pairs, single-word foils, and pairs consisting of two unstudied words. As expected, hit rates increased and the false-alarm rate decreased with the latency of the responses for all but one type of foil. For rearranged pairs, the false-alarm rates initially increased until about 600 ms subsequent to the probe, after which the false-alarm rates decreased, converging to an asymptote at about 1400 ms subsequent to the probe. Gronlund and Ratcliff (1989) speculated that a recall-to-reject mechanism might be used in order to account for the nonmonotonic relationship between the false-alarm rates for rearrange pairs and the latency of the response. Since Gronlund and Ratcliff, a number of other investigations have reported non-monotonic false-alarm rate functions (Hintzman and Curran, 1994; Light, Chung, Pendergrass, and Van Ocker, 2006; Light, Patterson, Chung, and Healy, 2004; McElree, Dolan, and Jacoby, 1999).
Other Important Issues: Recognition Memory as an Implicit Form of Memory The Cognitive Neuroscience of Recognition Memory Aging Context-dependent Recognition The Relationship between Recognition and Classification The Relationship between Recognition and Lexical Decision The Relationship between Recognition and Perceptual Identification The Relationship Between Recognition and Source Memory Recognition Memory versus Memory Scanning |
|