Home Introduction Prescientific 19th Century Interference Cognitive Primary Learning Representation Encoding Storage Retrieval Implicit Memory Implicit Learning Neural Bases Modeling Development Emotion Personal Social Conclusion

 

Implicit Memory 

The procedural knowledge that underlies conscious experience, thought, and action appears to be unconscious in principle, in the sense that we have no direct introspective access to these rules and skills, under any circumstances, and can know them only by inference from behavior. However, we usually think of the percepts, memories, and thoughts on which these procedures operate, and which these procedures create, as conscious -- at least in principle. We might not be aware of the unconscious inferences that create the moon illusion, or of the automatized reading skills which underlie the Stroop effect. But at least we are aware of the moon and the horizon, the fact that we are looking at them, and the fact that the moon seems larger on the horizon than it does at zenith -- just as we are aware of the words on the page, the color in which they are printed, our intention to read the words and ignore their color, and the difficulty we experience doing so. Mental procedures may be automatized and unconscious, but mental states -- our percepts, memories, thoughts, images, feelings, and desires -- are always conscious. Or are they?

No, they're not -- at least in the case of memory. Although it may be difficult to believe that we are not consciously aware of our percepts, thoughts, feelings, or desires, it is clear that we are not consciously aware of all of our memories at any particular time, any more than we are consciously aware of all the objects in our immediate physical environment. Such awareness would be a cognitive impossibility. We have an enormous amount of knowledge stored in memory, just as there is an enormous amount of material in the environment, but our limited attentional capacity prevents us from bringing more than a very small portion of it into consciousness at any particular time. Estimates of this capacity vary. The pioneering American cognitive psychologist George Miller famously summarized the capacity of conscious attention as "The magical number seven plus or minus two", and for present purposes that estimate will do nicely to capture the limitations on both conscious perception and conscious memory. We can hold only about seven items in consciousness at any particular time. All the rest is outside of consciousness -- available to conscious introspection, perhaps, but not accessible at that particular moment.

The idea of unconscious memory reminds us of Breuer and Freud's precept, that "hysterics suffer from reminiscences" -- representations of past experiences that influence their behaviors, in the form of symptoms, even though they are not accessible to conscious awareness. It is one thing for memories to lie dormant in a sort of passive storage, until they are actively retrieved for use. This is easy to imagine, because we have the experience every day, every moment, of turning out attention outward toward some objects and events and ignoring others. It is quite another thing for memories to dynamically influence our experience, thought, and action outside of phenomenal awareness. This is not part of our everyday experience, because even if it occurs, unconscious influence is by definition unnoticed and unnoticeable. So does it happen? And how would we know?


Availability and Accessibility

In early cognitive psychology, there was some tendency to think of the unconscious as a kind of wastebasket -- the repository of memories lost through decay over time or displacement by other memories. In the wastebasket metaphor, forgotten memories are effectively lost forever, and cannot subsequently be accessed and put to any sort of use. However, it serves no purpose to classify as "unconscious" memories that have been permanently lost and can play no role in influencing the person's ongoing behavior. The notion of "unconscious memory" makes no sense, and serves no purpose, unless there is some possibility that conscious access to the memory in question can be recovered -- or, failing recovery into consciousness, unless there is some demonstrable effect of the memory on the person's experience, thought, or action.


Attention and Memory

One way to construe unconscious memory is to realize that we have more memories "in storage" than we are aware of at any particular moment. Just as we have the ability to direct our attention outward, to selected aspects of the external environment, we have the ability to direct our attention inward, and to deliberately retrieve stored information about the past of which we are not currently aware. I have just thought of the moment at which I first met my spouse, Lucy -- at the intermission of a blues concert at the Memorial Union on the University of Wisconsin campus -- but I wasn't consciously remembering that event a couple of minutes ago.

The deliberate retrieval and reconstruction of stored memories is one way to think about the distinction between conscious and unconscious memories. By virtue of these processes, we make conscious what was previously unconscious. Of course, the same thing can happen spontaneously. Looking out my study window just now, I was watching Halley, our neighbor's cat, cross the street. This reminded me of an event, in September 1997, when two ducks marched into our garage demanding bread (we had fed them before). I wasn't consciously aware of this memory a moment ago, either; it was just sitting there in memory, waiting to be revived under the appropriate conditions, and brought to my conscious awareness.


Cue-Dependency and Retrieval

In light of considerations such as these, Tulving and Pearlstone (1966) distinguished between the availability of a memory in "storage" -- that is, once the memory trace has been encoded through some perceptual or learning experience -- and the accessibility of that memory to any particular attempt to retrieve and use it. Consider, for example, the difference between short-answer and multiple choice tests: we may not be able to generate the correct answer to a question on our own, but we may well recognize it when we see it. The correct answer is obviously available in memory, but it was only accessible on the multiple-choice test. Accessibility is a matter of bringing into consciousness a memory that previously resided in a latent, unconscious state of mere availability.

In the laboratory, the distinction between availability and accessibility is modeled by the "verbal learning" experiment, a procedure in which subjects are exposed to a list of familiar words during the study phase, and later asked to remember the contents of the list during the test phase. By varying the conditions of study, testing, or the interval between study and test, researchers gain experimental control over the three phases of memory processing -- encoding, storage, and retrieval. Traditional laboratory tests of memory take the form of a query, which by its very nature provides some cue information indicating what the subject is supposed to remember. In free recall, this cue information consists of a description of the spatiotemporal context in which the study phase took place: "Tell me what words were on the list you just studied", or "Tell me what was on the list you studied yesterday". In cued recall, the cues are more specific: "The words included some animals and some fruits: can you tell me what they were?". In recognition, the cues include the studied items themselves (copy cues), as well as comparable items which were not studied: "Which word was on the list: rabbit or elephant? Banana or apple?".

01TulWat.JPG (27653
              bytes)The classic finding is that free recall produces the least memories, cued recall somewhat more, and recognition the most of all. Consider a simple experiment in which subjects studied a lists of familiar five-letter words. At the time of test, they were given either no cues, or cues consisting of the first two to five letters of the words. The no-cue condition is equivalent to free recall, while the five-letter cue condition is tantamount to recognition testing; the intermediate conditions represent various levels of cued recall. The figure shows that the subjects remembered more words under cued recall than under free recall, and more was remembered under recognition than under cued recall. Within the cued-recall conditions, more letters in the cue yielded more words remembered.

The Tulving and Watkins experiment shows that memories that are inaccessible on a free-recall test may very well be accessible on a cued recall test, and those that are inaccessible on a cued-recall test may be accessible on a recognition test. The recall failure of recognizable words illustrates the important principle of cue-dependency in memory. Every act of memory retrieval begins with some retrieval cue, provided by the environment or generated by the rememberer him- or herself. The accessibility of any memory available in storage, then, is a function of the amount of information provided by the retrieval cue. The cues in cued recall may be of many sorts: semantic associates, category labels, initial letters or fragments, or similarly sounding items. In the case of recognition, the cue is a copy of the item itself. Whatever the nature of the cue, the more information it provides, the more likely it is that any attempt at retrieval will be successful. The unrecalled-but-recognizable memories needed the right amount of cue information to make them accessible to conscious recollection. Without the proper cues, they were available in memory, but remained unconscious until conditions fostered their retrieval.


Encoding Specificity and Transfer-Appropriate Processing

Cue-dependency is an important determinant of accessibility, but it must be qualified by a further principle of encoding specificity. The encoding specificity principle (known to memory researchers as ESP) states that items are accessible in memory to the extent to which the cues processed at the time of retrieval match the cues processed at the time of encoding. In other words, it is not just the sheer amount of cue information that matters: the retrieval cues have to be of the right sort. According to encoding specificity, the "rightness" of a cue is determined by events that occur at the time the memory was initially encoded.

As an illustration of encoding specificity, Tulving and Thompson asked subjects to study paired associates consisting of a target item, such as COLD, and a weak cue, such as ground -- i.e., ground-COLD. After studying the list, the subjects were presented with the cue and asked to recall the associated item from the study list. This cycle was repeated for another list of word pairs, such as badge-BUTTON. The purpose of the cued-recall procedure was to encourage the subjects to form associative links between cues and targets. Subjects studied a third list of word pairs, such as glue-CHAIR, just as they had the previous two lists. Instead of the usual cued-recall test, the subjects were surprised with a free-association test in which they wrote out the first words that came to mind in response to particular cue words. Some of the cues, such as table, were selected to be strong associates of targets studied in the third phase, and so the subjects naturally included some target words in their responses. They were then asked to examine their list of free associates, and circle any words that they had recognized from the study phase.

02TulThomp.JPG
              (24699 bytes)As shown in the figure, the subjects missed many targets on this recognition test. They actually remembered more targets on a subsequent cued-recall test, when presented with the cues originally paired with the targets during the study phase. Tulving and Thomson's experiments showed that subjects were able to produce, on a cued-recall test, studied items that they missed on a recognition test. This finding reversed an ancient principle of memory:-- that recognition is superior to recall. Their discovery, which they labeled the recognition failure of recallable words, also seems to violate the principle of cue-dependent forgetting, because the "copy cues" provided on the recognition test appear to be richer and more informative than the weak associates provided on the cued-recall test. But in fact, the phenomenon simply illustrates an important qualification. Accessibility is not just a matter of the sheer amount of information provided by the retrieval cue; rather, the cue must provide the right kind of information. Available memories are most easily brought to consciousness when the information processed at the time of retrieval matches the information processed at the time of encoding.

A variant on encoding specificity is the principle of transfer-appropriate processing, familiarly known as TAP. According to TAP, memory is best when the cognitive processes deployed at the time of retrieval match those that had been engaged at the time of encoding. The wording of TAP parallels that of ESP, but the emphasis is on processes rather than cues. Morris et al. illustrated TAP with an experiment in which subjects studied words presented under one of two orienting tasks. In the phonemic condition, they were asked to compare the sounds of two words -- e.g., whether hail rhymed with bail. In the semantic task, they were asked to compare the meanings of the words -- e.g., whether hail was associated with snow. The procedure is a variant on paired-associate learning, involving items such as hail-bail and hail-snow. Later, the subjects were tested for recall cued by new rhymes or associates: pail-_____ or sleet-_____.

03MorrBran.JPG
              (23851 bytes)As shown in the figure, there was overall better memory for items processed semantically, compared to those processed phonemically, at the time of encoding -- an illustration of the levels of processing effect in memory. But there was also an interaction: for items subject to phonemic encoding, retrieval was somewhat better when given phonemic cues; for items subject to semantic encoding, retrieval was considerably better when subjects were given semantic cues. This result is not a product of cue overlap between study and test, because the cues presented at test were not presented at encoding. But despite the change in cues, the processing remained the same, phonemic or semantic. Memory theorists debate whether ESP is a special case of TAP or vice-versa, but we need not get caught up in this debate. For our purposes, the important point uniting both principles is that available memories may be inaccessible to conscious retrieval unless there is an overlap between events that occur at the time of retrieval and those that occurred at the time of encoding.


State-Dependent Memory

A rather dramatic case of encoding specificity is found in the phenomenon of state-dependent memory (SDM) or state-dependent learning (SDL) in experimental subjects who perform memory tasks under the influence of psychoactive drugs. The first convincing demonstration of SDM was by Overton, in a classic series of studies of animal learning. In his experiments, Overton required rats to learn to run a T maze (that is, a maze shaped like the letter T) in order to escape shock. In one condition of the experiment, the rats were rewarded if they turned left at the choice point. Before the learning phase, the rats were drugged by injection of a high dose of barbiturates. Despite their sedation, over 10 training sessions the animals learned to respond perfectly. Then Overton allowed the drug to wash out of their systems.

04Overton.JPG (20453
              bytes)The figure shows what happened when the animals were retested: the animals behaved randomly, turning left or right in equal proportions. It was as if they had forgotten what they had learned. However, a control group, which also learned under the influence of the drug, was re-administered the drug before the retest. These rats reliably chose the proper arm of the maze, showing that they had not forgotten what they had learned. Rather, the learning was state-dependent: performance depended on the congruence between the drug state present during learning, and that present during testing. In terms of the encoding specificity principle, we may infer that the animal's physiological state provided cues which affected the accessibility of available memories.

Overton called his discovery state-dependent learning, but it was not the learning which was state-dependent: it was performance -- or, put another way, the animals' memory for what they had learned. State-dependent memory has also been observed in humans who have been administered psychoactive drugs such as barbiturates, narcotics, amphetamines, anti-anxiety and anti-depressant agents, marijuana, and alcohol before engaging in the study or test phases of a typical memory experiment. Even nicotine (but probably not caffeine!) has been observed to produce state-dependent memory. In each of these cases, subjects who learn under the influence of the drug remember more if they are also tested under its influence, than if they are not. Similarly, subjects who study while drug-free remember more if they are tested in the same state.

05Swan1.JPG (23377
              bytes)A classic illustration of human SDM was provided by Swanson and Kinsbourne, in a study of the effects of Ritalin, an amphetamine-like stimulant, on learning and memory in children. The drug, which usually impairs learning, has paradoxical effects in some patients with attention deficit/hyperactivity disorder (ADHD), improving performance in many domains. Swanson and Kinsbourne employed a variant on paired-associate learning in which the names of cities were paired with the names of animals. On the first day of the experiment, hyperactive children learned half of the list before, and half the list after, taking a standard dose of Ritalin. On the second day, they relearned half of each list in the same, and half in the opposite, condition from which it had been learned originally. As the figure shows, relearning was fastest (i.e., the children made fewer errors) when there was congruence between the two drug states: for example, items initially learned on Ritalin were relearned faster if the child were on Ritalin again. Similar, though weaker, results were obtained for control subjects who were not hyperactive.

As a rule, SDM is not as dramatic in humans as it is in nonhuman animals. Partly this is because ethical considerations prevent researchers from giving human subjects the high drug doses that can be administered to laboratory animals. SDM is also asymmetrical. As a rule, there is more transfer from non-drugged to drugged state than the reverse. Relatedly, there is a main effect of the drug itself: as a rule, psychoactive drugs impair both encoding and retrieval (the paradoxical positive effects of Ritalin on memory in hyperactive children are perhaps the only counterexample). The main effect of the drug may account for the asymmetry. In any event, as a practical matter students who have a few beers while studying should not tank up before their exams! Performance is best if students abstain during both phases. Moreover, SDM effects are cue-dependent: they are most commonly obtained on tests of free recall, and rarely with cued recall or recognition. Encoding specificity does not always trump cue-dependency; apparently, the cues available in the latter procedures are powerful enough to outshine those provided by the drug state.

06EichMet.JPG (21808
              bytes)Drugs are not the only way to induce state-dependent memory. Comparable effects have been observed when normal subjects shift emotional state between encoding and retrieval -- a phenomenon known as mood-dependent memory. In one experiment, Eich and Metcalfe induced happy or sad moods by asking their subjects to listen to music, subjects studied a list of words in one mood, and then received a test of free recall in the same or different mood. As shown in the figure, recall was better when subjects were tested in the same mood as that in which they had studied the list. Similar effects have been found in psychiatric patients diagnosed with manic-depressive illness, as they go in and out of various phases of their illness.

07GoddBadd.JPG
              (23743 bytes)Drug- and emotion-state-dependent memory both have to do with the subject's internal physiological or mental state. Accessibility also shifts when the external environment shifts, a phenomenon known as environment-dependent memory. In another classic experiment, Godden and Baddeley (1975) asked deep-sea divers to study a list of words either on the beach or 15 feet underwater. They then received a free-recall test of memory in the same or opposite condition. The figure shows that memory was best when study and test took place in the same environment.

These effects suggest that drug-induced SDM is a special case of context dependency in memory: memory is best when the context (internal or external, physiological or mental) in which retrieval is attempted matches the context in which encoding occurred. Context dependency follows naturally from the principles of encoding specificity and transfer-appropriate processing. The accessibility of available memories depends on the match between the context in which memories are encoded, and the context in which they were retrieved.


Reminiscence and Hypermnesia

Accessibility can vary even in the absence of changes in retrieval cues. When Tulving gave subjects a series of recall tests after a single study trial, he observed inter-trial shifts in memory for individual items. Some words, recalled on an early trial, were forgotten on later ones, and vice versa. The inter-trial recovery of memory was generally matched by inter-trial forgetting, so that net recall did not change. With lonion intervals, or longer delays between tests, inter-trial forgetting probably would have exceeded inter-trial recovery, resulting in the time-dependent forgettin nghau, there are circumstances under which inter-trial recovery can exceed inter-trial forgetting, resulting in a net t of recall over time. Ballard, who originally observed this phenomenon, named it reminiscence -- the opposite for oblivescence, his wonderfully evocative term for forgetting.

08Erd.JPG (28262 bytes)More than half a century later, Matthew Erdelyi (1984, 1986) extended Ballard's findings in an extensive program of research on what Erdelyi labeled hypermnesia, or the opposite of amnesia. An early study by Erdelyi and Becker (1974, Experiment 1) illustrates his essential procedure and findings. The subjects studied either a set of 40 line drawings of familiar objects or a list of the 40 corresponding words, and then received a series of three recall tests. Between the two tests, one group was asked to engage in free association, writing down whatever occurred to them; another group was asked to think silently, and a control group proceeded from test to test without any interpolated activity. Figure 5. shows the results: for pictures, though not for words, net recall increased from trial to trial. This is the basic hypermnesia effect.

Erdelyi's research was partly inspired by Freud's alleged success in inducing the recovery of traumatic memories in hysterical patients, as reported in the Studies of Hysteria. Erdelyi also argued that pictorial stimuli, and imagistic representations in memory, were especially susceptible to hypermnesic effects; this, in turn, was taken as consistent with the classically Freudian view of the unconscious as the repository of imagistic, nonverbal memories. However, free association is not necessary to produce hypermnesia: as Erdelyi & Becker themselves showed clearly, a simple retest will do. Nor are pictures and images privileged with respect to hypermnesia: a number of investigators subsequently observed hypermnesia for lists of words, including highly abstract words that are difficult to recode as images. As a general rule, any experimental manipulation that promotes good recall also promotes hypermnesia. The important point about hypermnesia is not that it supports Freud's view of the unconscious, or the validity of psychoanalytic theory and procedures. The important point is that hypermnesia illustrates the emergence of previously unconscious memories into consciousness.


Unconscious Memory

Cue-dependent retrieval, encoding specificity and transfer-appropriate processing, state-, mood-, and context-dependency, and hypermnesia all illustrate one form of unconscious memory. Memories available in storage may be inaccessible on particular attempts at retrieval, even under conditions that would seem to be objectively optimal (such as recognition testing). In that sense they are unconscious. This condition is not permanent, however, because a later shift in circumstances may render these memories accessible to conscious recollection. Thus, the memories are not unconscious in the "wastebasket" sense of being permanently unavailable to conscious recollection, but are unconscious in the more interesting sense of being temporarily inaccessible.  Following Freud, we might well call them preconscious, because they are available to consciousness, at least in principle, but not consciously accessible at any particular time.  

This is fine so far as it goes, but the question of unconscious memories goes beyond the fluctuating accessibility of information available in memory storage. For one thing, the notion of "preconscious" memories implies a kind of passive storage -- that preconscious memories are somehow latent, and do not dynamically interact with other ongoing cognitive processes until they have actually been accessed, and brought into conscious awareness.  The really interesting question has to do with the dynamic activity of unconscious memories: can traces of the past influence ongoing experience, thought, and action without our being aware of them? This is the question of implicit memory.

Explicit memory is the conscious recollection of some past event. If I am passed on the highway by a classic Bugatti convertible, of the sort in which the dancer Isadora Duncan died, I am likely to remember the experience, at least for a while. Implicit memory is much broader, referring to any effect of a past event on the person's subsequent experience, thought, and action. If I am passed on the highway by that same automobile at a time, and later reply "Bugatti" when asked to name a make of European sports car (when "Jaguar" or "MG" would probably come to mind more easily), I am showing implicit memory. Intuitively, we think of explicit and implicit memory as going together. When someone asks me why I thought of the Bugatti, I can tell them about my experience on the highway. But evidence from both amnesic patients and normal subjects suggests that they can be split apart. But I might also think of the Bugatti without realizing why I am doing so. 

Under those circumstances, explicit and implicit memory can be dissociated from each other.  In other words, subjects can show implicit memory for a prior event in the absence of conscious recollection of that event -- or, more loosely, independent of conscious recollection.


Implicit Memory in Amnesic Patients

Initial evidence for a dissociation between two expressions of memory, explicit and implicit, was provided by studies of patients with the amnesic syndrome first described by Sergei Korsakoff (1854-1900), a Russian neurologist and psychiatrist (Korsakoff, 1889a/1996, 1889b). The characteristic feature of this syndrome is a specific deficit in memory for recent events, in the absence of impairments in sensory or perceptual functions, intelligence, or language. Korsakoff originally observed this memory deficit in patients suffering the ravages of chronic alcoholism, and the amnesic syndrome in alcoholics -- often accompanied by confabulation, or a tendency to make up memories -- is traditionally known as Korsakoff's syndrome.

The appearance of Korsakoff's syndrome in alcoholics is a consequence of an inadequate diet, which ultimately causes a deficiency in thiamine (vitamin B1); new cases are rare since the introduction of vitamin-enriched bread and other prepared foods in the late 1940s and early 1950s. However, the amnesic syndrome can result from other causes, including tumors, cerebrovascular accidents, and traumatic brain injury. 

Patients like these, for whom the brain injury can be precisely dated, permit a decomposition of the amnesia into two components. The anterograde amnesia covers the "postmorbid" period following the injury, while the retrograde amnesia covers the "premorbid" period before there was any brain damage. All patients with the amnesic syndrome suffer a profound anterograde amnesia, which persists until death. For example, Patient H.M., who died in 2008 at the age of 82, remembered essentially nothing that has happened to him since his surgery in 1953 when he was 27 years old. The retrograde amnesia, when it occurs, is more time-limited, and may spare remote memories, encoded well before the injury occurred. For example, although H.M.'s memory for his young adulthood, before his surgery, is rather fragmentary, he can remember much of his childhood.

09Brain.JPG (27718
              bytes)It is now known from lesion and brain-imaging studies of both humans and nonhuman animals that the amnesic syndrome comes in two major forms, each resulting from damage to different structures in the forebrain (see the figure). The classical Korsakoff's syndrome is associated with damage to the diencephalon, a set of structures that includes the thalamus and hypothalamus, and particularly to the mammillary bodies. Other cases of amnesic syndrome, such as H.M., result from damage to the hippocampus and surrounding structure in the medial portion of the temporal lobe, including the parahippocampal cortex surrounding the hippocampus, and the entorhinal cortex and perirhinal cortex surrounding the amygdala -- structures which have come to be known as the medial temporal lobe memory system. In either case, the damage must be bilateral, affecting corresponding structures in both cerebral hemispheres, to produce a full-blown amnesia; unilateral damage produces more specific memory deficits.

Early research on amnesic patients employed traditional memory paradigms, intended to document the extent of the patients' memory deficits.  For example, research on serial-position effects showed that amnesic patients had normal recency effects but diminished primacy effects.  These results were typically interpreted in terms of the multistore model of memory, meaning that the amnesia impaired long-term memory but spared short-term memory.  In fact, this selective impairment of long-term memory was cited by Atkinson & Shiffrin as evidence for their structural distinction between memory stores.   Specifically, the amnesic syndrome was interpreted as resulting from consolidation failure: study items were held in short-term memory, but could not be copied into long-term memory.  

Neuropsychological studies of the amnesic syndrome promised to give us our first clues about the biological basis of memory.  But as the studies accumulated, the picture became more complicated - -chiefly because, on some memory tests, amnesic patients showed no deficits at all.  The story begins with a series of experiments reported by Elizabeth Warrington and Leonard Weiskrantz.


Priming in the Amnesic Syndrome

WarrWeis1970.jpg
              (57158 bytes) Warrington and Weiskrantz discovered that priming is spared in amnesic patients who do not remember the priming event. When they study a list of familiar words, and are asked to recall them shortly thereafter, amnesics show gross impairments in memory compared to controls. But quite different results may be obtained when they are asked to identify briefly presented words, or to complete a word stem or other fragment with a meaningful word. Under these circumstances, amnesic patients were more likely to produce words from the previously studied word list than from control lists that had not been studied. This phenomenon is known as a priming effect, in which performing one task (in this case, studying a word list) facilitates later performance of another task (in this case, completing stems or fragments with meaningful words). The normal controls also showed priming, and the extent of priming did not differ between the two groups. Warrington and Weiskrantz subsequently confirmed and extended their initial findings, as did others (e.g., Brooks & Baddeley, 1976).

Priming effects were known before Warrington & Weiskrantz, and they have become an important tool in the study of cognition.  For example, in a classic study, Meyer & Schvaneveldt (1977) employed priming in a study of the structure of semantic memory.  

Observation of semantic priming in a lexical decision task suggests that concepts in semantic memory are linked in an associative network.  Processing one item activates its internal representation in memory, and this activation can spread to other, related representations, facilitating task performance.  

Priming is no surprise when subjects consciously remember the primes.  What's interesting in the Warringnton & Weiskrantz experiment is that priming still occurred when the subjects -- the patients were amnesic, after all - -did not remember the primes.  So priming in amnesia reveals two expressions of episodic memory, which were best articulated by Dan Schacter (1987), following earlier work by Graf & Schacter (1984).

The Warrington & Weiskrantz studies were the first to reveal that explicit and implicit memory can be dissociated -- that implicit memory can occur in the absence of explicit memory.

As Schacter (1987) noted, the distinction between explicit and implicit memory has long been in the literature, especially -- you should forgive me for putting it this way -- implicitly.

Priming is an expression of episodic memory, because it requires that some trace remain of the experience of performing the prior task. But recall and recognition are explicit expressions of episodic memory, because they make explicit reference to some past event, and explicitly require subjects to cast their minds back to some action or experience that occurred sometime in the past. The experimenter asks "What were the words you studied?", and the subject replies with "I remember ostrich, flannel, garden, and baby." By contrast, memory is only implicit in priming: the experimenter does not ask the subject to recollect anything about the past, and the subject's task performance is focused on the here and now. The experimenter asks "What is this word?" and the subject replies "Ostrich", or whatever. The subjects do not have to remember the past, but memory for the past is implicit in their priming performance. Priming shows that available memories are not merely latent, but rather are dynamically active, able to influence performance on other ongoing tasks.

10GrafSqui.JPG
                (25753 bytes)The difference between explicit and implicit expressions of memory is underscored by an experiment which corrected a subtle ambiguity in many early studies of priming in amnesia. The stem- and fragment-completion tasks often used to assess priming are analogous to cued recall, in that the stem or fragment may serve as a cue for recalling the entire target word. Therefore, when a completion task yields more studied targets than a free recall task, we do not know whether this is because priming is a qualitatively different expression of memory, or merely because the completion task provides quantitatively more cue information than the free recall task. The fact that amnesics perform poorly on tests of recognition memory, which provide even more cue information than cued-recall tests, argues against this possibility. Still, a properly controlled comparison of explicit and implicit memory would hold the cues constant at the time of the test, and vary task requirements. In their research, Graf and his colleagues asked amnesic patients and controls to study a list of words individually printed on index with three-letter stems under two conditions. In the explicit memory test, the subjects were asked to examine a list of word stems and "try to think of a word from the cards with the same beginning letters" (Experiment 3, p. 172); in the implicit memory test, they were asked merely to "write the first word that comes to mind" beginning with those letters" (Experiment 1, p. 168). The figure shows the results of the comparison. Amnesics showed poorer cued recall than controls, but slightly better stem completion. The fact that the stem-completion test yields more evidence of memory than the stem-cued recall test has to do with the nature of the tests, not differences in available retrieval cues. If the stems are used as cues for conscious recall, they yield little by way of explicit memory; but if they are used as cues for priming, they yield substantial evidence for implicit memory.


Priming Following Electroconvulsive Therapy

Other forms of amnesia also show a dissociation between conscious, explicit and unconscious, implicit memory. Chief among these is the amnesia which occurs as a side-effect of electroconvulsive therapy sometimes used to treat acute cases of depression and other psychotic disorders, especially those that are unresponsive to drugs. Although precise procedures may vary from clinic to clinic and from case to case, in modern practice ECT entails the delivery of a burst of electrical current (perhaps 100 volts at 500 milliamperes) through electrodes applied to the skull of the patient. The passage of the current across the brain induces a convulsive seizure resembling that of epilepsy, which in turn results in memory deficits resembling those associated with concussive blow to the head. First, the patients remember nothing that occurred during the time they were anesthetized, including the convulsive seizure itself, or during the post-ictal confusional period extending for about 30 minutes after the shock is delivered. After the patients have recovered their orientation, however, they show a retrograde amnesia covering the period of time immediately prior to the shock, and an anterograde amnesia covering the time immediately afterward. Bilateral ECT, in which the electrodes are placed over the left and right temporal lobes, produces more memory difficulties than unilateral ECT, in which one electrode is placed over the nondominant (usually right) hemisphere, and the other at the vertex of the skull. The precise mechanism of this amnesia is unclear, but the fact that one or both electrodes are placed over the temporal lobes suggests that the functioning of the medial temporal-lobe memory system is somehow disrupted. The memory deficits are unrelated to the outcome of ECT. In that sense they are an annoying side-effect, but they have also provided researchers with an opportunity to study another form of organically based amnesia.

11SquiShim.JPG
                (25990 bytes)Unlike the amnesic syndrome, both anterograde and retrograde components dissipate after the course of treatment has ended; but while the ECT is proceeding, the amnesia selectively impairs explicit memory, while sparing implicit memory. On the anterograde side, Squire, Shimamura, and Graf (1985) asked a group of patients to study a list of words 45 minutes after delivery of the ECT. The figure  compares their performance on two tests of memory. For a test of recognition, the subjects were presented with sets of three words and asked to choose from each set the one they had studied earlier. On this test, the patients performed very poorly compared to controls, barely better than chance levels of 33% correct. For the test of stem completion, they were asked to complete three-letter stems with any English word; each stem could be completed by an item from the study list or by any of at least 10 other familiar words. On this test patients and controls alike were more likely than chance to complete the stems with list items, compared to chance performance of 10%; more important, the magnitude of this priming effect was the same in the two groups.

12DorfECT.JPG
                (26442 bytes)On the retrograde side, Dorfman and her colleagues performed a similar experiment, except that the patients studied their wordlists within the 30 minutes immediately prior to the treatment. Approximately one hour after treatment, the subjects were given a variety of memory tests. As shown in the figure, the patients performed poorly on a test of stem-cued recall, compared to controls, but equally well on the test of priming in stem-completion. Thus, both the anterograde and retrograde amnesias induced by ECT impair explicit memory but spare implicit memory. The patients cannot consciously recollect the words they have studied, but traces of these words unconsciously affect their performance on other, ongoing tasks.

At this point, dissociations between explicit and implicit memory have been confirmed in a wide variety of amnesic populations.


Implicit Memory in "Normal" Subjects

The quotation above from Leibniz suggests that the dissociation between explicit and implicit memory can be observed in normal, neurologically intact human subjects, as well as in amnesic patients.  And so it can.  In contrast to the population dissociations discussed earlier, these are functional dissociations.

Kihl1980.jpg (24150
                bytes)The dissociation between explicit and implicit expressions of memory is most dramatic in amnesic patients, but it can be seen even in individuals who are neurologically intact. Posthypnotic amnesia dissociates explicit from implicit memory, in much the same way that the amnesic syndrome or ECT does. For example, subjects who learned a list of words during hypnosis will be more likely to use these words as free associates or category instances, even though they cannot remember the learning experience itself. This happens even though there has been no damage to the medial temporal-lobe memory system, or any other insult, injury, or disease affecting brain tissue. In some respects, this can be characterized as a population dissociation, because the contrast is between two groups -- one that is amnesia and the other not.  

I will discuss posthypnotic amnesia in some detail later, but for now I turn to somewhat more prosaic examples in which explicit and implicit memory are dissociated in "normal" neurologically intact laboratory subjects.

The dissociation between explicit and implicit memory was dramatically illustrated by a series of studies of savings in relearning, a measure of memory invented by Ebbinghaus for his classic studies of memory for nonsense syllables. In one series of experiments, Ebbinghaus would memorize a list of nonsense syllables and then let a period of time pass, inducing some degree of forgetting. Then he would attempt to learn the same list again. Not surprisingly, it was easier to learn the list the second time, but the difference between the two attempts became the first quantitative measure of memory. 

13Nels.JPG (21290
                bytes) Nelson's subjects did something similar, memorizing lists of 20 paired associates consisting of a two-digit number and a familiar word (e.g., 48-book and 36-party). Study-test trials continued until the subjects met a strict criterion for learning. After a four-week retention interval, with no further study, the subjects were given two tests of memory. For the cued recall test, the subjects were presented with the stimulus terms (e.g., 48-) and asked to provide the corresponding response terms (e.g., book). For the recognition test, they subjects were presented with each stimulus term paired with all 20 response terms, and asked to select the correct pairing. For the relearning portion of the experiment, Nelson asked his subjects to learn a second list of words composed of three types of items: those that had been successfully recalled, those that were recognized but not recalled, and those that were neither recognized nor recalled. Within each group, half the items were the same as those presented during the original acquisition phase; for the remainder, the original cue was paired with new targets (e.g., 48-party). The figure presents the results for the two groups of unrecalled items, recognized and unrecognized.

In these experiments, special interest attaches to those items that were forgotten on both cued recall and recognition tests -- i.e., items for which the subject had no explicit memory. Nonetheless, the subjects found it easier to relearn old compared to new pairings. To be sure, there was less saving in relearning unrecognized than for recognized items, but the fact that there were any savings at all indicates that the unrecalled-and-unrecognized items remained available in memory. Nelson's study is a classic in the memory literature because it definitively demonstrated that savings in relearning is a more sensitive measure of memory than recall or recognition. But the importance of savings goes beyond the question of sensitivity. Because the relearning task asked subjects to master an ostensibly new list, and not to remember a list from the past, and because the old items were not consciously accessible, savings in relearning task constitutes an implicit expression of memory for the past event. As such, Nelson's experiment was the first experimental demonstration that implicit memory could be spared when explicit memory was impaired in normal, neurologically intact subjects.

14JacDal.JPG (21946
                bytes)Nelson's study went to great lengths to insure that normal subjects would have poor explicit memory. Another early demonstration took another tack, employing Craik and Lockhart's (1973) levels of processing paradigm. In the first phase of the experiment, the subjects were asked make "shallow" orthographic or "deep" semantic judgments about a list of words. Half the subjects then received a recognition test of explicit memory. The remaining subjects were simply asked to identify words presented on a computer screen. In this perceptual identification task, the words are flashed so briefly that it is difficult to make them out, but it is easier to identify previously studied items compared to items that are entirely new -- another kind of priming effect. As the figure shows, level of processing had a substantial effect on explicit recognition, but it had no significant effect on the magnitude of priming in perceptual identification task. This illustrates another form of dissociation between explicit and implicit memory: an experimental manipulation -- in this case, level of processing during encoding -- has an effect on recognition but not on priming in perceptual identification.

So far we have seen two different ways in which explicit and implicit memory can be dissociated. In the case of priming in amnesic patients, and Nelson's study of savings in relearning, we see the influence on subsequent task performance of items that are not accessible to conscious recollection. In the Jacoby and Dallas study, an experimental manipulation that affects conscious recollection does not affect priming.  Does the reverse ever occur? 

15GrafMod.JPG
                (28142 bytes) In an early study, Graf and his colleagues presented words for study either visually or aurally, followed by written tests of free recall and stem-completion. As the figure shows, the amnesics were grossly impaired on the explicit test, but showed levels of priming that were comparable to controls. However, the change in format from aural presentation to written test had no effect on recall, while it suppressed priming to some degree. This is another experimental dissociation: a shift in modality between study and test affects implicit memory, but not explicit memory. The effects of the modality shift are sometimes referred to as hyperspecificity, meaning that implicit memory is more closely tied than explicit memory to the specific modality or format in which the experience in question originally occurred.

 

Source Amnesia

We do not just remember our past experiences. We also learn from them, and the fate of this newly acquired knowledge can provide another perspective on the dissociation between explicit and implicit memory. If I am driving down the highway and we pass an exotic-looking car, and my spouse tells me that it's a Bugati, I may volunteer the information that Isadora Duncan was killed in a Bugati, when her scarf got caught in the car's open wheels. If she then asks me where in the world I learned such an obscure fact, I will reply that the incident was portrayed in the 1968 film of Duncan's life, starring Vanessa Redgrave in the title role (I don't believe everything I see in the movies, but this is actually true). Remembering some piece of knowledge, and remembering the source of that knowledge, in my past experience, seem to be two different types of memory. In fact, Endel Tulving has distinguished between episodic memory, or our autobiographical memories of particular personal experiences, and semantic memory, or our repository of context-free, generic knowledge.

Although we acquire new semantic memories through experience - a process called learning, Tulving has argued that these two forms of memory are mediated by separate memory systems in the brain. On the other hand, both episodic and semantic memory are both forms of declarative knowledge. Semantic memories are clearly propositional in nature, for example:

Columbus discovered America

or

Automobiles have engines.

But episodic memories are also propositional in character. Every episodic memory consists, first, of a proposition (or perhaps a set of propositions) describing an event, such as:

John gave a present to Lucy.

This sentence, as written, has something of the character of a semantic memory, representing little more than the raw fact that one person touched another. More elements are needed to make this semantic memory into an episodic memory. First, we need to add additional propositional knowledge, representing the episodic context in which the event occurred:

John gave a present to Lucy in the restaurant on her birthday.

Second, and most critically, to make an episodic memory we need to add further propositional knowledge representing the role of oneself in the event represented: this link to the self creates the autobiographical character of episodic memory. In linguistic terms, the self can appear in one of four ways in an episodic memory:

as the agent or patient of some action:

I gave a present to Lucy (etc.)

Lucy gave a present to me (etc.).

or as the stimulus or experiencer of some state:

I made Lucy happy (etc.).

Lucy made me happy (etc.).

Any autobiographical memory can be represented in this fashion, as a proposition or set of propositions which refer to some event involving oneself.

While my conscious recollection of a particular learning experience is clearly an explicit memory, in a sense my knowledge of what I learned through that experience is an implicit memory for that same experience. I can now do something -- like reflect on a piece of knowledge, or convey it to someone else -- that I was not able to do before the learning occurred. These considerations raise the further question of whether one's (implicit) memory for a piece of semantic knowledge can be dissociated from one's (explicit) memory for the episodic source of that knowledge. Now, there is a trivial sense in which this is true. Although there are some salient exceptions, we do not usually remember where and when we picked up particular facts about the world, or learned how to spell certain words, or what they mean, and nobody expects us to do so. However, we do ordinarily remember the source immediately after we have a learning experience. Accordingly, in this context we may say that explicit and implicit memory are dissociated when a person remembers a piece of information, but quickly -- more quickly than would ordinarily be expected -- forgets the conditions under which it was learned.

This phenomenon is called source amnesia, a term initially coined to describe instances where a hypnotized subject experiencing posthypnotic amnesia remembered new (and obscure) factual knowledge learned during hypnosis, but could not remember the hypnotic circumstances in which that learning occurred. However, the basic observation of source amnesia, if not the label, has been in the neurological literature on amnesia since the very beginning. Korsakoff himself reported a number of observations suggesting that "even though the patient has no memory of traces of the impressions he receives, these traces persist and probably influence, in some way, his unconscious intellectual activity". On one occasion, Korsakoff was administering electrical shocks to a patient for therapeutic purposes:

Every time I asked him what I would do to him he remained perplexed and answered that he did not know. I would urge him to look at the table where the case that enclosed the machine was placed. Then he told me that I was probably here to give him electro-shocks. I know that he had only encountered this machine during his illness. Consequently, if he had not retained some trace of memory of the case containing the machine, he could not have guessed so quickly (p. 9).

Based on these sorts of observations, Korsakoff concluded, first, that amnesia reflected an inability to evoke or reproduce memories, rather than an inability to retain them. However, the traces were too weak to be reproduced consciously, and could only be reproduced unconsciously.


Claparede's Patient

Perhaps the most famous demonstration of semantic memory as implicit memory was reported by Edouard Clarparede (1873-1940), a pioneering Swiss psychologist. Claparede, who was a neurologist and psychiatrist as well as a psychologist, spent some time observing and testing patients at the local psychiatric hospital -- among whom was a woman with Korsakoff's syndrome. The patient's memory deficit was so profound that, after five years of hospitalization, she was still unable to recognize her doctors and nurses as familiar. However, Claparede noticed that she was able to navigate around the hospital, and showed other evidence of learning. She could find her way to the ward's bathroom, though she could not tell anyone where it was, or describe it. Although she could not remember a particular nurse's name, or specify her role, she behaved appropriately with her and other staff members, and generally followed the domestic routine of events in the hospital. Claparede also reported (unfortunately without providing any details of the experiment), that the patient showed savings in relearning, in the absence of conscious recollection of the previously learned material

...What is worthy of our attention here was her inability to evoke recent memories voluntarily, while they did arise automatically, by chance, as recognitions.

When one told her a little story, or read various items of a newspaper to her, three minutes later she remembered nothing, not even the fact that someone had read to her; but with certain questions one could elicit in a reflex fashion some of the details of those items. But when she found these details in her consciousness, she did not recognize them as memories but believed them to be something "that went through her mind" by chance, an idea she had "without knowing why," a product of her imagination of the moment, or even the result of reflection (pp. 374-375).

On one occasion, Claparede introduced himself to this patient, and while shaking hands (because she was amnesic, she failed to recognize him as familiar), pricked her with a pin he had hidden in his palm. As he tells the story:

The light pain was forgotten as quickly as neutral perceptions; a few minutes later she no longer remembered it. But when I again reached out for her hand, she pulled it back in a reflex fashion, not knowing why. When I asked for the reason, she said in a flurry, "Doesn't one have the right to withdraw her hand?" and when I insisted, she said, "Is there perhaps a pin hidden in our hand?" To the question, "What makes you suspect me of wanting to stick you" she would repeat her old statement, "That was an idea that went through my mind," or she would explain, "Sometimes pins are hidden in people's hands." But never would she recognize the idea of sticking as a "memory" (p. 375).

A single episode of experience changed the patient's knowledge about the world -- she now knew that sometimes people hide pins in their hands -- even though she had no conscious recollection of the experience by which she acquired that knowledge. As Claparede put it in his autobiography, "the contents of memory can remain unrecognized, even when they are capable of being reproduced or of starting adapted reactions". This is exactly what we mean when we say that implicit memory can be dissociated from explicit memory.


Source Amnesia in Neurological Patients

Although Claparede's story was quickly absorbed into the clinical lore of neurology, formal study of source amnesia in brain-damaged patients had to wait more than a half century. In the course of a test of general knowledge, Schacter and his colleagues informed an amnesic patient of certain obscure facts, such as that the American president Theodore Roosevelt held the world's record for shaking hands. When queried later, he correctly recalled or recognized more than half the items he had been taught. However, he was rarely able to correctly state the source of this new knowledge.

In a more formal study, Schacter and his colleagues taught amnesic patients a wide variety facts. Some of these were well-known "true facts" about well-known people -- e.g., that Al Capone was a gangster; others were "false facts" about well-known individuals -- e.g., that Bob Hope's father was a fireman; still others were "false facts" about individuals the patients did not know -- e.g., that Alice Reznak (in fact, a fictional name made up by the experimenters) was addicted to nicotine. This procedural shift was designed to vary the familiarity of the individuals named in the test, and to insure that some of the "knowledge" in question was actually acquired, and not merely refreshed, during the experiment (a false fact about Bob Hope is unlikely to have been learned from a movie magazine, but it could have been linked to other, pre-existing knowledge about the Hollywood comedian). In addition, the information was presented by one of two experimenters. The information was presented in question-and-answer format. One of the experimenters would ask a question, such as "What job did Bob Hope's father have?". If the patients knew the answer (or thought they did), they were asked where they learned it. After a delay of about 30 or 120 seconds, during which the patient responded to other such questions, the questions were repeated as a test of memory.

16Source.JPG (23100
                bytes)The basic results of the experiment are shown in the figure. The patients correctly remembered about only a small amount of the new information. However, they correctly remembered the source of this new information for only a minority of these remembered items. Most of the attribution errors were false attribution to some extra-experimental source -- for example, saying that they learned it on television. They remainder were correctly attributed to the experiment, but attributed to the wrong experimenter. By contrast, a second experiment showed that neurologically intact subjects, tested at intervals of 10 minutes to 1 week, correctly remembered the source of newly acquired information most of the time. When these normal subjects made source errors, most were of the intra-experimental variety, in which the two experimenters were confused with each other. Source amnesia, as represented by extra-experimental errors in attributing the source of correctly remembered items, did occur in normal, nonamnesic subjects, but it occurred far more frequently in amnesic patients.

Similar findings were reported by Shimamura and Squire, who taught their subjects general-information items of the sort used in the popular game, Trivial Pursuit. The amnesic patients, who included both neurological patients with amnesic syndrome and psychiatric patients receiving bilateral ECT as a treatment for depression, had poorer recall of new facts presented during the experiment. When they did successfully recall the new information, however, they were much more likely to erroneously identify the source of their knowledge. Interestingly, a later study by Janowsky, Shimamura, and Squire found that source amnesia was especially prominent in neurological patients with damage to the frontal lobes of the cerebral cortex, even when the patients were not otherwise amnesic.


Cryptomnesia and Reincarnation

Outside of hypnosis, source amnesia is rare in neurologically intact individuals, but it occurs dramatically in occasional episodes of unconscious plagiarism, or cryptomnesia. For example, Mark Twain, the American author, took the dedication of Innocents Abroad from Oliver Wendell Holmes' Songs in Many Keys. As Twain himself told the story, at a dinner celebrating Holmes in 1879:

Two years before, I had been laid up a couple of weeks in the Sandwich Islands, and had read and re-read Doctor Holmes' poems till my mental reservoir was filled up with them to the brim. The dedication lay on the top, and handy, so, by-and-by, I unconsciously stole it....  Well, of course, I wrote Doctor Holmes and told him I hadn't meant to steal, and he wrote back and said in the kindest way that… he believed we all unconsciously worked over ideas gathered in reading and hearing, imagining they were original with ourselves.

Some cases of alleged reincarnation have the flavor of cryptomnesia, as in the famous case of "Bridey Murphey". At a party on November 29, 1952, Morey Bernsein, a Colorado businessman and amateur hypnotist, hypnotized "Ruth Simmons" (as she was called in the book, to protect her privacy), an acquaintance who had on previous occasions proved to be an excellent hypnotic subject. Using a technique called hypnotic age regression, he suggested that she relive the first year of her life -- and, when that was successful, to try to go back even further. The subject, whose real name was Virginia Tighe, then began speaking in Irish-accented English, recalling events from her previous life as a young girl named Bridget (Bridey) Murphy McCarthy, who had lived in County Cork, Ireland during the 19th century. Bernstein continued his "experiment" (p. 8) over several sessions, extracting a great deal of material about Bridey's life and her surroundings -- even though Ms. Tighe had never lived in, or even visited, Ireland.

This past-life regression, apparently providing evidence for reincarnation, caused a sensation when it was initially reported in a series of articles in Empire, the Sunday magazine for the Denver Post, and later in a book by Bernstein himself. Several newspaper reporters even traveled to Ireland, where they were able to verify much of the material that Ms. Tighe had recalled under hypnosis -- although at least one Irishman weighed in with a dissenting view. Another group of reporters, instead of following Bridey back to Ireland, followed Ms. Tighe back to her childhood in the Midwest, where they discovered many plausible sources of her memories -- including lessons in Irish dialect, song, and dance, and contact with a number of people with Irish connections -- including a Chicago neighbor named Bridie Murphy Corkell. One interpretation of the whole body of evidence is that Ms. Tighe absorbed a fair amount of Irish cultural knowledge as a child, which was revived in the imaginary context of hypnosis -- without corresponding revival of the source of that knowledge. Like other cases of its type, and setting aside the possibility of deliberate deception, very likely this apparent past-life regression was source amnesia masquerading as reincarnation.

Given the multicomponent view of episodic memory described earlier, it is easy to see how source amnesia occurs. Under ordinary circumstances, a person who is told that Bob Hope's father was a fireman is likely to remember both the fact and the telling:

Today I heard on TV that Bob Hope's father was a fireman.

But if the episodic and self-referent components of episodic memory are fragile for some reason, the knowledge represented there may be lost, so that all the person remembers is the occupation of the actor's father, and not how he or she came by that information.


Learning in Amnesia

The literature on source amnesia shows that amnesic patients can acquire new semantic knowledge, despite an anterograde amnesia that impairs their ability to remember the learning experience itself. Other studies show the amnesic subjects can acquire new procedural knowledge as well. The notion of procedural knowledge has its origins in the work of the French philosopher and psychologist Henri-Louis Bergson (1859-1941), who in his classic treatise on Matter and Memory asserted that "The past survives under two distinct forms: first, in motor mechanisms; secondly, in independent recollections". Bergson illustrated this distinction with the example of learning a lesson (p. 89-91):

The memory of the lesson, which is remembered in the sense of learnt by heart, has all the marks of a habit.... like every habitual bodily exercise, it is stored up in a mechanism which is set in motion as a whole by an initial impulse, in a closed system of automatic movements…. The memory of each several reading, on the contrary… has none of the marks of a habit. Its image was necessarily imprinted at once on the memory…. It is like an event in my life…. On the contrary, the memory of the lesson I have learnt… is no longer a representation, it is an action. And, in fact, the lesson once learnt bears upon it no mark which betrays its origin and classes it in the past; it is part of my present, exactly like my habit of walking or writing…. I might believe it innate, if I did not choose to recall at the same time, as so many representations, the successive readings by means of which I learnt it.

Bergon's point was picked up later by the British analytical philosopher Gilbert Ryle (1900-1976) who distinguished knowing how from knowing that, as well as by Terry Winograd and John Anderson, in their distinction between declarative and procedural knowledge. Although it is not entirely clear that episodic and semantic memories are represented differently, both types of knowledge being declarative in nature, a better case can be made for a representational distinction between declarative (whether episodic or semantic) and procedural knowledge. Declarative knowledge is factual in nature, and can be represented in sentence-like propositions, while procedural knowledge concerns mental and behavioral operations, and can be represented by systems of IF-THEN statements known as productions.

Following this reasoning, it seems possible that a person could acquire a new skill through learning and practice, but be unable to remember the learning and practice sessions themselves. In fact, among normal individuals this is almost certainly the case: each of us possesses many skills, acquired long ago in childhood, whose acquisition we can no longer remember. But can we observe the same sort of phenomenon over shorter intervals of time -- retention intervals that would not ordinarily induce forgetting of the learning experience? Apparently so. In 1845, even before Korsakoff's syndrome received its name, the British physician Robert Dunn described an amnesic woman who learned to be a dressmaker after the onset of her illness, "even though she had no recollection from day to day what she had done". More than 100 years later, Moira Williams and George Talland, in their pioneering experimental studies, also saw some evidence of learning in the amnesic syndrome.

17Maze.JPG (21977
                bytes)Theorists began to take notice of the learning capabilities of amnesic patients with research on H.M., the amnesic patient introduced earlier. As part of a continuing, long-term investigation of this patient, Milner and her colleagues employed a visual "stepping stone" maze consisting of a rectangular array of metal bolts embedded in a wooden base. On the task, H.M. was required to traverse a particular pathway from start to finish, touching each bolt on the path in turn. When he touched a bolt that was not on the pathway, he was informed of his error. Early studies with a 10x10 maze involving 30 points and 11 "turns", showed no reduction in errors over repeated trials, indicating that H.M. was unable to learn his way through the maze. However, the path in question was quite complex, and Milner et al. reasoned that he might do better with a less complex maze, only 4x6 and involving only 9 points and 3 turns, whose path did not exceed the capacity of short-term memory (which, for H.M., was found to be normal). The figure summarizes H.M.'s performance on this task.

As can be seen, H.M. did indeed show evidence of learning, progressively reducing his errors on the 100 trials of first day of training. After a day of rest, he showed some regression in performance, but reached a criterion of learning after 55 more trials. On subsequent days of testing, H.M. continued to improve his performance, showing considerable "savings" in relearning and quickly reaching criterion, even after a retention interval of six days. H.M.'s performance was by no means normal: neurologically intact control subjects learn the maze more quickly and retain it longer. But the fact that he showed improvement from trial to trial within a day of testing, and some degree of savings across days, provides clear evidence of procedural learning of the following sort:

IF I wish to leave the starting point,

THEN move down two points;

IF I have moved down two points,

THEN move right four points;

and so on until he reached the goal point.

Milner et al. also found that H.M. could learn a tactual maze while blindfolded, while Corkin found that he was capable of learning a "pursuit-rotor" maze which requires the subject to maintain contact between a stylus and a point on a rotating disk. However, H.M. was completely unaware that he had engaged in a large number of trials, over extended days, while learning these skills.

At about the same time, Warrington and Weiskrantz, found evidence of learning on another, more perceptual task in which subjects were asked to name the objects depicted in fragmented line drawings. On the first trial, the subject is shown a picture that was almost unrecognizable. On successive trials, more and more details are in -- until, on the fifth trial, the whole picture, easily identifiable, would be shown. Warrington and Weiskrantz also devised a similar task involving the identification of fragmented words. On both tasks, amnesic patients showed progressive improvement across trials within a session, in identifying wholes from fragments, and considerable savings across sessions conducted on separate days. Corkin and her colleagues confirmed these findings with H.M..

18Pursuit.JPG
                (22752 bytes)Likewise, Brooks and Baddeley found that amnesics took progressively less time to traverse the Porteus Maze (a common intelligence test), and to assemble the same jigsaw puzzle over repeated trials. However, there was no transfer to a new jigsaw puzzle. In each case, some degree of learning was retained over an interval of one week. Perhaps the most striking finding of these investigators was on the pursuit-rotor task studied by Corkin, as the figure shows, there was no difference between amnesics and controls in performance on this task, or in retention over a one-week interval, Thus, the amnesic patients showed clear evidence of learning even though they did not recognize the tasks as ones they had previously worked.

Even more dramatic evidence of spared learning abilities in otherwise densely amnesic patients was provided by Squire and Cohen, in a study of mirror-reading. In their experiment, subjects were presented with sets of 3 mirror-reversed words, such as:

xxxxx xxxxx xxxxx

19Mirror.JPG (26577
                bytes)Whereas pursuit-rotor learning is a perceptual-motor task, Cohen and Squire argued that mirror-reading requires a more "cognitive" pattern-analyzing ability. In addition, while Brooks and Baddeley's subjects performed the same task over and over, Cohen and Squire tested their subjects on both repeated and nonrepeated triads. In this way, they could trace their subjects' ability to learn to decode specific word triads, and their ability to learn a generalized perceptual-cognitive skill. The figure shows the results for Patient N.A., who became amnesic after receiving an accidental stab wound to the diencephalon. Similar results were obtained for a group of neurological patients with Korsakoff syndrome, and for another group of psychiatric patients receiving ECT for the treatment of depression.

It can be seen that N.A.'s performance is practically indistinguishable from that of controls, on both repeated and nonrepeated word triads. He learned to read both repeated and nonrepeated words printed backwards, at a rate comparable to that of controls. Nevertheless, N.A. and the other amnesic patients were unable to recognize the words that they had learned to read backwards; nor did they have any awareness that some of the words had been repeated as much as 20 times during the experiment. The only appreciable difference between amnesics and controls was that the amnesics showed substantial forgetting of repeated words on the first trial of each day after the first, although they quickly regained lost ground. Note, however, that this forgetting occurred only for the repeated words, additional testimony to their lack of explicit memory. For the generalized perceptual-cognitive skill represented by reading nonrepeated mirror-reversed words, there was no forgetting across days for either amnesics or controls. The patients learned, but they did not remember what they had learned.


Formulating the Explicit-Implicit Distinction

Based on findings such as these, Graf and Schacter proposed a distinction between explicit and implicit memory. As noted earlier, explicit memory entails conscious recollection of some past event. Explicit memory tests, such as free recall, cued recall, and recognition, always refer to some past event, and always require conscious recollection of the past. By contrast, implicit memory refers to any effect of a past event on some subsequent task entails conscious recollection of the past. Implicit memory tasks, such as stem- and fragment completion, perceptual identification, savings in relearning, the learning of new factual material, and the acquisition of skills such as pursuit-rotor performance and mirror-reversed reading, do not logically require conscious recollection of the past. All the subjects have to do is perform the task presently before them. Still, traces of past events and experience reveal themselves in the subjects' task performance.


Precursors to the Explicit-Implicit Distinction

Warrington and Weiskrantz (1968) discovered implicit memory in amnesic patients, and Graf and Schacter (1984) gave it its name. But as Schacter (1987) has amply documented, what we now recognize as the explicit-implicit distinction already had a history within both philosophy and psychology. For example, in his groundbreaking treatise on memory, Uber das Gedachtniss, Hermann von Ebbinghaus invented the technique of savings in relearning precisely because he wanted a measure of memory that was not restricted to voluntary, conscious recollection.

Ebbinghaus wanted to circumvent the limits of voluntary recollection in his research, but -- perhaps because he earlier had been critical of von Hartmann's view of the unconscious as a force permeating the entire universe -- he lacked a substantive interest in unconscious expressions of memory. More than a decade earlier, however, the German physiologist Ewald Hering, who was influenced by Goethe, Kant, and Schopenhauser as well as by Fechner and Helmholtz, offered a prescient view of unconscious memory that has not been widely recognized. Hering distinguished, first, between two forms of conscious memory.

The word "memory" is often understood as though it meant nothing more than our faculty of intentionally reproducing ideas or series of ideas. But when the figures and events of bygone days rise up again unbidden in our minds, is not this also an act of recollection or memory. We have a perfect right to extend our conception of memory so as to make it embrace involuntary reproductions of sensations, ideas, perceptions, and efforts.

Here, then, we have a distinction between two forms of conscious recollection, one which is the product of intentional, controlled activity and one which occurs more or less automatically. But then Hering goes on to consider an expressly unconscious form of memory in the sense of latent traces of past experience, available in storage but not presently accessed:

I was conscious of this or that yesterday, and am again conscious of it to-day. Where has it been meanwhile? It does not remain continuously within my consciousness, nevertheless it returns after having quitted it. Our ideas trend but for a moment upon the stage of consciousness, and then go back again behind the scenes, to make way for others in their place (p. 70).

Hering also gives an account of Helmholtz's unconscious inferences:

The perception of a body in space is a very complicated process. I see suddenly before me, for example, a white ball. This has the effect of conveying to me more than a mere sensation of whiteness. I deduce the spherical character of the ball from the gradations of light and shade upon its surface. I form a correct appreciation of its distance from my eye, and again I deduce an inference as to the size of the ball. What an expenditure of sensations, ideas, and inferences is found to be necessary before all this can be brought about; yet the production of a correct perception of the ball was the work only of a few seconds, and I was unconscious of the individual processes by means of which it was effected, the result as a whole being alone present in my consciousness (p. 72).

In later passages, Hering notes that processes that were once conscious can become unconscious by virtue of practice.

How long does it not take each note to find its way from the eyes to the fingers of one who is beginning to learn the pianoforte; and, on the other hand, what an astonishing performance is the playing of the professional pianist. The sight of each note occasions the corresponding movement of the fingers with the speed of thought - a hurried glance at the page of music before him suffices to give rise to a whole series of harmonies; nay, when a melody has been long practised, it can be played even while the player's attention is being given to something of a perfectly different character over and above his music (p. 73).

From this observation, Hering deduces that skilled performance must be mediated by an unconscious form of memory encoded in the nervous system itself:

Our perceptive faculties must have remained always at their lowest stage if we had been compelled to build up consciously every process from the details of the sensation-causing materials tendered to us by our senses; nor could our voluntary movements have got beyond the helplessness of the child, if... the motor nerve system had not also its memory, though that memory is unperceived by ourselves. The power of this memory is what is called "the force of habit"… (p. 74). [T]he reproductions of organic processes, brought about by means of the memory of the nervous system, enter but partly within the domain of consciousness, remain unperceived in other and not less important respects (p. 75).

Still, for Hering, "unconscious" memory either refers to automatized habits of mind -- to procedural rather than declarative knowledge -- or to the biological substrate of memory -- the physiological memory trace, or engram. In the former case, the label "unconscious" seems appropriate. In the latter instance, however, memory is no longer a mental concept, and so -- as with Hartmann's Absolute and Physiological levels of the unconscious, the term "unconscious" is really a misnomer.

More than two centuries before either Ebbinghaus or Hering, the 17th-century French philosopher Rene Descartes, in The Passions of the Soul, observed that a frightening or other negative experience may be "imprinted" on a child's brain "to the end of his life", without "any memory remaining of it afterwards" -- a neat anticipation of Breuer and Freud's (1893-1895/1953) later pronouncement (without apparently having read Ebbinghaus) that "hysterics suffer from reminiscences". In the 18th century, the German philosopher Leibniz, already introduced in the context of subliminal perception, also discussed priming effects: "Often we have an extraordinary facility for conceiving certain things, because we formerly conceived them, without remembering them".

Something like the explicit-implicit distinction in memory was especially well developed in 19th- and 20th-century French philosophical psychology. For example, Maine de Biran (1766-1824), a French philosopher of the "empiricist" school, distinguished between a "representative" form of memory, involving conscious awareness, and "mechanical" and "sensitive" memory involving unconscious motor and emotional habits. While recall and recognition are clearly examples of the former, priming and savings might serve as examples of the latter. Similarly, as noted earlier, the French philosopher and psychologist Henri-Louis Bergson (1859-1941) asserted that "The past survives under two distinct forms: first, in motor mechanisms; secondly, in independent recollections". For Bergson, recollection entails the conscious recollection of "memory-images" (p. 92) representing specific past events. A habit or skill, by contrast, "bears upon it no mark which betrays its origin and classes it in the past; it is part of my present" (p. 91).

Aside from Bergson, William MacDougall (1871-1938), a pioneer in the study of motivation, was the first to distinguish between explicit and implicit memory within experimental psychology. McDougall argued that recognition is a fundamental cognitive function because "[t]he biological function of Mind is to bring the past to bear upon present action, guiding it in anticipation of the future psychology". In addition, MacDougall distinguished between a primitive form of implicit recognition, by which we recognize an object or event as somehow familiar, and a more developed form of explicit recognition, in which "that reference to the past which is implicit in all mental activity is more explicit and more prominent than in other acts" (p. 309, emphasis original). Perhaps echoing Descartes, MacDougall wrote:

The dog that runs away at sight of the man who kicked him yesterday recognizes him implicitly. It is probably true to say that the perceiving of the same man merely evokes again the fear-impulse which yesterday was evoked by his brutal act. The dog does not think "That is the man who kicked me yesterday; I will get out of his way lest he kick me again.... [T]he similarity of the effect… is the essential ground of recognition (p. 308).

 

The Origins of Priming

Explicit and implicit memory are both expressions of what psychologists call episodic memory -- traces, or representations, of particular events that are encoded and retained in memory. These traces are available to influence ongoing experience, thought, and action, even if they are inaccessible to conscious recollection. While tests of explicit memory are virtually limited to various forms of recall and recognition, almost any task can serve as an implicit memory task, provided that we can demonstrate that changes in task performance are systematically related to the subject's past experience. Priming effects are especially clear indices of memory, because they result from a single presentation of some stimulus. Therefore, they represent memory for a single discrete episode in the person's experience, as opposed to an extended series of trials with the pursuit-rotor or reading mirror-reversed words. For that reason, a great deal of research on implicit memory has focused on priming effects.

Although research on explicit memory has been going on for more than a century, research on priming is of relatively recent vintage. A search of the PsycINFO database, a virtually complete index of the journal literature in psychology, conducted in the summer of 2001, revealed more then 4700 items, almost 3/4 of which appeared since 1990. The term was originally introduced in a theory of classical conditioning, where the conditioned response was considered to be a prime for the unconditioned response. Later, priming was used to describe the enhancement of audiogenic seizures in rats by the previous presentation of another, non-convulsive stimulus, the preparation of the body for complex sequences of behavior, and the effect of non-ejaculatory penile intromissions on the copulatory behaviors of male rats. Later, investigators of the "pleasure centers" in the brain discovered that rats would not engage in self-stimulation unless they had already received a free "priming" shock, independent of their behavior.

It was not until the mid-1960s that the term was imported into research on cognitive processes, in a study of the effects of extraneous cues on word-association performance. Shortly thereafter, Segal employed priming in a comparison of "indirect" and "direct" test of memory. Her subjects studied a list of words for one, two, or four exposures, and then were tested after a delay of six or twelve minutes. Free recall served as the direct test of memory, while the indirect test was whether studied items occurred as responses in a free-association task. She reported that both exposure and delay affected free recall, while neither variable had any impact on free-association performance. We can now see Segal's experiment as the earliest demonstration of a dissociation between explicit and implicit memory, similar to those reported by Jacoby and Dallas and by Graf et al.. Unfortunately, Segal neither used the term nor drew any conclusions from her experiment about unconscious memory. Rather, like Ebbinghaus with his "savings" technique, she seems mostly to have been interested in developing a test of memory that did not depend on self-reports.

In general, priming comes in two broad forms.  

Either way, priming is an effect of memory.  Apparently, some trace of the previous activity persists, and interacts with subsequent task performance.  In the cases that concern us, episodic memory is implicit in the priming effect, because some past event influences processing in the present.  This effect must, logically, be mediated by a representation of the past event -- which is what an episodic memory is.

Priming itself comes in two broad forms, depending on the relation between the prime and the target..

In theory anyway, repetition priming is mediated by what John Anderson would call a perception-based representation of the prime -- that is, a representation of the physical (including spatial and temporal) properties of the prime.
In theory, semantic priming is mediated by what Anderson would call a meaning-based representation of the prime -- that is, the sort of representation that is created when a stimulus has been analyzed for meaning, not just physical appearance.

It is important to understand that repetition and semantic priming are not necessarily associated with particular tests.  Even stem- and fragment-completion tests can be employed to study semantic priming, depending on how the test is constructed.  Consider, for example, a subject who studies the prime doctor.

In all these instances, explicit and implicit memory are dissociated when implicit memory is spared even when explicit memory is grossly impaired.  Dissociations are very popular in cognitive psychology, cognitive neuropsychology, and cognitive neuroscience, because they seem to imply that performance on the two dependent variables in question -- e.g., the explicit and implicit memory tasks -- is mediated by qualitatively different underlying processes, including qualitatively different brain systems.  Statistically, all dissociations take the form of interactions between independent variables, and these interactions themselves take three quite different forms.

As originally clasified by Lashley (1952) and Teuber (1955):


There are single dissociations (Teuber called them simple dissociations, in which a single independent variable, A, affects dependent variable Y but not another dependent variable, Z.  The literature on implicit memory provides numerous examples where amnesia or level of processing (for example) affects explicit memory (e.g., recall or recognition) but not implicit memory (e.g., priming).  However, there are occasional findings of the opposite type: for example, modality shift appears to affect implicit memory (e.g., priming) but not explicit memory (e.g., recall).

More convincing are double dissociations (I call them twin dissociations, because they're really tantamount to two single dissociations), in which one independent variable, A, affects one dependent variable, Y, but not another one, Z; while a second  independent variable, B, affects Z but not Y. Double dissociations are the "Holy Grail" of cognitive neuroscience, because they seem to indicate that performance on the two tests, explicit and implicit, is mediated by different memory systems.  But they are also very rare in captivity.  Yes, there can be triple and quadruple dissociations, but these are even rarer.



A variant on the
double dissociation is the crossover dissociation, in which a single
independent variable has opposite effects on two dependent variables.


In the present context, the dependent variables are various explicit memory tasks, chiefly recall or recognition, and various implicit memory tasks, such as priming.  The independent variables come in two broad forms.

But for now, let's return to population dissociations involving amnesic patients and non-amnesic controls.


Kinds of Priming

At this point, a fairly large number of priming procedures have been employed in the study of implicit memory. As already described, subjects can be asked to identify masked or briefly presented words, complete word stems or fragments, or provide free associations. But the number of priming tests is limited only by the ingenuity of the investigator, and the tasks can be carefully tailored to the specific needs of the experimenter, so some classification is in order.

Perhaps the earliest such distinction is between direct and indirect priming. In direct priming, the target is a token of the prime (or, for that matter, vice-versa). For example, a subject might see the prime

bashful,

and then be asked to identify the same word,

bashful,

when briefly presented on a computer screen, decide whether the item is a legal English word, or to complete the stem

bas___

or the fragment

b_s_f_l

with the first word that comes to mind. Such priming is also called repetition priming, because the target repeats the prime -- although, technically, repetition priming should be reserved for those cases where the prime and the target are absolutely identical. Direct priming is, by far, the most studied form of implicit memory, as in the pioneering studies of Warrington and Weiskrantz.

In indirect priming, by contrast, the prime and target are not identical. For example, in a free-association task subjects might study the word touch; at the time of test, they would be presented with the word soft and asked to respond with the first word that came to mind. Touch is, in fact, a relatively low-frequency response to soft; hard is much more frequent, so if a subject says "touch", and a control subject does not, that is especially convincing that priming has taken place. Similarly, in a category instance generation task, subjects might study a word like sheep and later asked to generate instances of the category four-footed animal. Instances like dog, cat, lion, and tiger are much more frequently given than instances like sheep, so when sheep appears more frequently than in control subjects, again that is an instance of priming.

Indirect priming is exemplified by semantic priming on free association and category generation tasks, but there are other procedures as well. For example, subjects might be asked to answer a general-knowledge question such as:

What was the name of one of the dwarfs

in the children's story, Snow White?

 

. Alternatively, subjects might be asked to solve anagrams of studied items, name pictures representing studied words, or spell orally presented homophones such as

pes

that had previously been presented in a disambiguating context such as

war-peace or piece of pie

. Even tests that would ordinarily represent direct, or repetition, priming can be modified to assess indirect, or semantic, priming -- depending on exactly how they are constructed. For example, a subject might study a word like doctor, and then be presented with an item like nurse for perceptual identification, lexical decision, or stem-completion.

An alternative classification of priming tasks has been provided by Roediger and his colleagues, who have distinguished between perceptual and conceptual priming. Perceptual priming can be mediated by a perceptual representation of the priming stimulus -- that is, a memory that requires only a "shallow" or "bottom-up" analysis of the sensory or perceptual features of an item, but not a "deep" or "top-down" analysis of its semantic or conceptual features -- by an analysis of what the item looks like or sounds like, for example, rather than what it means. Perceptual priming can be mediated solely by a perceptual representation of the event -- a representation that holds information about the object's perceptual and spatial features, but lacks information about meaning or category membership. By contrast, conceptual priming cannot be mediated by a mere perceptual representation of the priming stimulus. When an item like doctor primes the processing of an item like nurse, the memory trace that mediates the priming contains more than just information about the surface features of the prime. The prime must have been analyzed for meaning as well.

Compared to repetition priming, semantic or conceptual priming has rarely been studied in amnesia. In an early example of such a study, Gardner and his colleagues taught subjects category-instance pairs in the form of

The first word is a tree. It is oak.

20SemanPrim.JPG
                (23102 bytes)As the figure indicates, amnesic patients showed a profound amnesia when measured in terms of free or cued recall, compared to control subjects. However, both amnesics and controls showed a semantic priming effect. When asked to generate instances of various categories, they were more likely to produce target items after the study experience than they had been on a baseline test before the study trials. In fact, the amnesics showed more priming than the controls. Similarly, Shimamura and Squire found intact priming when amnesics generated free associates to the cue terms of previously studied paired associates such as table-chair, and Schacter obtained normal levels of priming in an experiments studied common idioms such as sour-grapes, and then were asked to provide the first word that came to mind when presented with the first word of the idiom.

StateSpew.JPG
                (38073 bytes)Thomson et al. (2010 showed that semantic priming could be dissociated from explicit memory in normal, neurologically intact subjects as well.  Under the cover of a classroom discussion of the distinction between availability and accessibility in memory, these investigators incidentally primed subjects with state names, such as "D might make you think of Delaware".  These sorts of primes were intended to bias subjects' memory for state names.  Some 4 to 8 weeks later, the subjects were asked to "List as many of the 50 states as you can remember".  Sure enough, the states that had been primed during the earlier classroom discussion were recalled more quickly than those that had not been primed (the technical term for this is spew order).  When asked to speculate about any sources of bias, the subjects generally failed to remember the earlier priming episode.  OK, it's not a perfect demonstration, but there have been so few studies looking at semantic priming in amnesia t-- a situation that I'll moan more about later -- hat I'll grasp at anything!


 

Dissociating Implicit Memory from Explicit Memory

In priming studies, any advantage in performance for previously studied items over control items reflects memory for the prior exposure. This is in line with the definition of implicit memory as any change in experience, thought, or action that is attributable to a previous event, independent of conscious recollection of that event. But it is one thing to construe priming effects as evidence of memory, and quite another to consider them as evidence of implicit or unconscious memory. In the Segal experiment, for example, priming was considered to be just an alternate measure of memory, like recognition or savings. In fact, evidence of priming (or savings), in and of itself, is not evidence of implicit memory. This evidence only emerges when the effects occur in the absence of conscious or explicit memory -- in other words, when implicit and explicit memory are dissociated from each other. This dissociation can take many forms. There are population dissociations, in which one group of people, such as amnesic patients, or patients receiving ECT, show a different pattern of performance across explicit and implicit tests of memory. And there are experimental (sometimes known as functional) dissociations, in which one experimental variable, such as level of processing or modality shift, has a different effect on the two types of memory test.

In this context, the term "dissociation" is a special case of what statisticians call an "interaction" between two variables. In an experiment with two independent variables, each of these variables can exert a main effect on the dependent variable. That is, Variable A can significantly influence the dependent variable, and so can Variable B. These two effects can be independent of each other, or they can interact. If the interaction term is significant, the effect of one independent variable depends on the level of the other independent variable. When we say that amnesia dissociates explicit and implicit memory, we mean that the effect of one independent variable -- whether the subject is amnesic -- on memory performance (the dependent variable) depends on another independent variable -- whether the memory test is explicit or implicit in nature.

It is not entirely clear how the term dissociation came to be used to refer to statistical interaction. The term "dissociation" has a long history in psychiatry and clinical psychology, of course. Pierre Janet and others used the term to refer to cases of hysteria, fugue, and multiple personality disorder where one set of ideas was "split off" from the rest, resulting in a state of divided consciousness (Janet actually referred to this situation as desagregation, which was translated into English as "dissociation"). This meaning of dissociation is preserved in the dissociative disorders recognized by modern psychiatry, such as dissociative (or psychogenic) amnesia, dissociative (or psychogenic) fugue, and dissociative identity disorder (or multiple personality disorder). According to the Diagnostic and Statistical Manual of the American Psychiatric Association (APA), the central feature of the dissociative disorders is an alteration in consciousness affecting memory and identity.

In the context of implicit memory, however, "dissociation" lacks any necessarily pathological implication: it means simply that the association between two variables, which normally covary with other, has been somehow disrupted. As far as I can tell, William James (who was himself influenced by Janet) introduced this usage into experimental psychology, and it was quickly picked up by others. For example, Carr and Allen reported on the conditions under which accommodation and convergence could be dissociated in the perception of depth, while Downey discussed the dissociation of auditory and vocal imagery in handwriting. The usage was later revived in psychopharmacology, to describe the selective effects of drugs on different behavioral functions, and in neurology and neuropsychology, to describe the selective effects of brain damage. Thus, hippocampal damage dissociates explicit and implicit memory.

For theoretical reasons that will become clearer later, double dissociations are a sort of "Holy Grail" in the study of explicit and implicit memory. For one thing, they address the issue of whether the apparent dissociation is actually an artifact of something uninteresting (at least in this context), like task difficulty. It is easier to manipulate performance on a test that is moderately difficult than on one that is very easy or very hard. If a test is easy, the subject will be likely to pass it no matter what the experimenter does; if a test is hard, the subject will fail it under most circumstances. But if a test is moderately difficult, whether the subject passes or fails can be influenced by relatively small changes in the testing situation. For example, it might be that perceptual identification is such an easy test of memory that people show priming even for items that have been subject to very shallow processing. In principle, it would be possible to match tasks for difficulty level, but this is an arduous chore. If we can show that the variable in question had opposite effects on the two tests, explicit and implicit, that would relieve the suspicion that the lack of an effect on implicit memory was somehow an artifact of differential task difficulty.

There are lots of variations on the logic of dissociation, and it is not important that we go over each and every one of them. It is important to understand, however, that these dissociations do not have to occur, and the experiments could come out otherwise. Priming could yield more evidence of memory than free recall, just as cued recall and recognition do, without the pattern of population and experimental effects changing. The finding of dissociations -- that the pattern of performance differs from one group or experimental condition to another -- tells us that explicit and implicit memory are qualitatively different. Amnesia impairs free recall and recognition, but has little or no effect on repetition or semantic priming. Amnesic patients can retain facts they have recently learned, but do not remember that they recently learned them. They display skills they have recently acquired, but do not remember acquiring them. The one class of effects involves conscious recollection of the past, whereas the other involves unconscious influence of the past. Dissociations between explicit and implicit memory constitute the first evidence that mental contents, not just mental processes, can influence experience, thought, and action outside of conscious awareness.


Theories of Implicit Memory

Implicit memory is unconscious memory. That much seems clear from the evidence cited earlier. But it is not enough to establish dissociations between two measures of memory, one unconscious and the other conscious. We also want to know why the dissociation occurred. What makes the difference between memories that can be consciously recollected, and memories that can only be unconsciously expressed?


First, A Note on Competing Terminology

Let us begin with the very name of the phenomenon. It has been said that a scientist would rather use a colleague's toothbrush than his or her vocabulary. Perhaps for this reason, other researchers have used different terms to capture essentially the same distinction -- resulting in some confusion among readers, even if the writers in question knew what they meant.


Direct vs. Indirect Memory

On the one hand, some authors interpret the dissociations discussed here in terms of a distinction between direct and indirect memory. Recall and recognition are direct measures of memory, while priming is an indirect test. This is fine, but the direct-indirect distinction has already been used to distinguish between different kinds of priming. If we construe priming as an indirect measure of memory, and accept a distinction between direct and indirect priming, then we are trapped into accepting a distinction between "direct indirect" and "indirect indirect" tests of memory -- a rather unpalatable prospect.

In fact, there is a sense in which the direct-indirect distinction is orthogonal to the explicit-implicit one, as depicted in the table below. Repetition priming and semantic priming are direct and indirect measures of implicit memory, respectively. Neither task explicitly refers to a past event, and neither requires conscious recollection of that event. But in repetition priming the same items appear at study and test, while in semantic priming they do not. Free recall, cued recall, and recognition are clearly direct measures of explicit memory, because they require conscious recollection, and the same items appear (one way or another) at study and test. But are there any indirect measures of explicit memory?

Relations Among Direct and Indirect,

Explicit and Implicit, Tests of Memory

 

 

Explicit

Implicit


Direct

Free Recall

Cued Recall

Recognition

Repetition Priming 

(in all its forms)

Indirect

Retroactive Inhibition?

Proactive Inhibition?

Savings in Relearning?

Semantic Priming

 

Retroactive Inhibition?

Proactive Inhibition?

Savings in Relearning?

There are at least three possible candidates. In retroactive inhibition (RI), subjects study two lists, A and B, in that order; when they are tested on list A, they remember fewer items than if they had not also studied B. In other words, memory for B indirectly affects explicit memory for A, and does so retroactively. In proactive inhibition (PI), subjects also study A and B, and remember fewer items from B than they would have had they not also studied A. Memory for A indirectly affects explicit memory for B, and does so proactively. Savings in relearning is another possible case: subjects learn A and then relearn A again, but during the relearning phase their recollections are focused on the new instantiation of the list, not the old one. Savings can be interpreted as a measure of implicit memory, because the relearning phase does not explicitly refer to the initial learning phase, even though memory is implicit in the savings. Similar considerations apply to retroactive and proactive inhibition. Perhaps the way the test is framed makes a difference. If the instructions of a PI or RI experiment explicitly require that the subject recall both lists, or if the subject is told that he will be relearning a previously learned list, that procedural variation may effectively shift the test from implicit to explicit in nature.


Declarative vs. Procedural, and Declarative vs. Nondeclarative Memory

Another way of dividing up the memory universe is to distinguish between declarative and procedural memory, with "declarative" roughly meaning "explicit" and "procedural" roughly meaning implicit. As with the direct-indirect distinction, however, the declarative-procedural distinction is already in use, in terms of factual versus skill knowledge. Some manifestations of implicit memory are clearly procedural in nature, such as learning the Porteus maze or pursuit-rotor task, and mirror-reversed reading. Simple classical conditioning, which is also preserved in amnesia, also qualifies as procedural, in that the contingency between conditioned and unconditioned stimulus can be represented as a production: If CS Then US. However, it is not clear that priming effects are truly procedural in nature. On the one hand, repetition priming on a perceptual identification task could be taken as an instance of perceptual learning, and thus arguably procedural. On the other hand, priming has long been interpreted within associative network models of memory that are essentially declarative in nature.


Declarative vs. Nondeclarative Memory

For these kinds of reasons, Squire abandoned the declarative-procedural distinction for a distinction between declarative and nondeclarative memory. This tack acknowledges that some expressions of memory that are preserved in amnesia are not procedural in nature, but it doesn't solve the basis problem, which is that the declarative-procedural or declarative-nondeclarative distinction confuses phenomenology with representational format. Procedural knowledge is surely unconscious. But declarative knowledge can also be unconscious, as evidenced by priming and source amnesia. The meaning of declarative in ordinary language, as something one can "declare", should not be confused with the technical meaning of the term in psychology and cognitive science, as referring to the propositional format for knowledge representation. If we are going to talk about consciousness, we have to talk about phenomenology; and if we have to talk about phenomenology, the explicit-implicit distinction does least to confuse issues.


Voluntary vs. Involuntary Memory

Sometimes, explicit and implicit memory are characterized as voluntary and involuntary expressions of memory, respectively.  The idea is that free recall, cued recall, and even recognition are things that people do voluntarily, deliberately, with malice aforethought -- they are trying to remember something.  By contrast, priming is something that just happens, as a product of spreading activation or something.  But that distinction is not quite right, because conscious recollections can occur involuntarily as well.  The origins of the Crovitz-Robinson technique for sampling autobiographical memory, discussed in the lectures on Personal Memory, is in a game that Sir Francis Galton, Darwin's famous cousin and the father of psychometrics, used to play.  He'd go out for a walk, and let the things he encountered on the walk bring memories to mind (Crovitz called his technique "Galton's Walk", and wrote a book about it).  Later authors developed collections of people's spontaneous, involuntary autobiographical memories -- as in Esther Salaman's A Collection of Moments: A Study of Involuntary Memories (1970).

So, while priming occurs involuntarily, conscious recollection can be either voluntary and involuntary, either deliberately and spontaneously retrieved. So, this too is a bad choice -- though that doesn't mean that involuntary conscious recollection isn't interesting: it is very interesting indeed.

The point of implicit memory isn't that it occurs involuntarily.  The point of implicit memory is that it's unconscious.  The fact that's unconscious means, perforce, that it is expressed involuntarily.  After all, you can't voluntarily control something that you're not consciously aware of.  But the salient feature of implicit memory, what makes it really interesting, is that it is unconscious memory.


So Why Not Just Call It Unconscious Memory ?

The explicit-implicit distinction is preferable, in my view, because it centers on the question of the subjects' awareness of the memory in question, rather than their ability to verbalize it or the nature of the memory test. However, why not simply distinguish between conscious and unconscious memory? After all, that is what implicit memory is all about. Schacter (1987) considered this possibility, but rejected it on the ground that the term unconscious is "saturated with multiple and possibly misleading meanings" (p. 502). Very likely, it seems, Schacter wished to avoid the taint of Freudian psychoanalysis. At the same time, we can detect in the preference for explicit and implicit a shadow of the positivistic reserve described by the philosopher Owen Flanagan. Even now, many psychologists and cognitive scientists are reluctant to talk about consciousness, and they are especially nervous about anything unconscious. As we can see in the automaticity literature, this is starting to change, but even there (and I think out of the same positivistic reserve) researchers and theorists prefer to characterize the processes in question as automatic rather than unconscious.

Whatever the reason, the implicit-explicit distinction is here to stay, it is preferable to all other attempts to avoid speaking of consciousness and the unconscious, and I happily adopt it in the rest of this book. Still, it should be clear that implicit memory is unconscious memory. Now that dissociations between explicit and implicit memory have been convincingly established, the principal problem is to understand the nature of unconscious memory.


Are Implicit Memories Just Weak?

Perhaps the simplest theory of implicit memory is that unconscious memories are too weak to be accessible to conscious awareness and expression. This is also, perhaps, the earliest theory of implicit memory -- one that was held by Korsakoff himself.

In his seminal reports on the amnesic syndrome, Korsakoff himself described a number of observations suggesting that "even though the patient has no memory of traces of the impressions he receives, these traces persist and probably influence, in some way, his unconscious intellectual activity" (Korsakoff, 1889a/1996, p. 9). On one occasion, for example, Korsakoff was administering electrical shocks to a patient for therapeutic purposes:

Every time I asked him what I would do to him he remained perplexed and answered that he did not know. I would urge him to look at the table where the case that enclosed the machine was placed. Then he told me that I was probably here to give him electro-shocks. I know that he had only encountered this machine during his illness. Consequently, if he had not retained some trace of memory of the case containing the machine, he could not have guessed so quickly (p. 9).

Based on these sorts of observations, Korsakoff concluded, first, that:

[Amnesia] is not due to the inability to retain information, but the inability to reproduce it (p. 14).

In contemporary terms, a deficit in retrieval, rather than in encoding or storage -- a point of view later echoed by Talland and by Warrington and Weiskrantz themselves. But why can't the memories be retrieved? Korsakoff argued that the traces were simply too weak to be reproduced consciously, and could therefore only be reproduced unconsciously.

[W]e notice a whole series of traces that can not in any way be reestablished in consciousness, either actively or passively, but that continue to exist in the unconscious life and that direct the general way of thinking of the patient, suggesting this solution as opposed to that one. This seems to me to be the most interesting characteristic of this illness (p. 13).

* * *

[T}here is a massive amount of evidence that points to the fact that our actions, the direction of our attention, etc., are influenced by past impressions that are unconscious to us. We can explain these phenomena only if we assume that these traces are reproductions, but that they are so weak that they do not pierce into the conscious state. In order for a processes (sic) to become conscious, it has to possess a certain degree of intensity and duration. If it does not have these properties, it will remain unconscious. It follows that all the reproductions of impressions that do not reach this degree of intensity and duration remain unconscious; but this does not mean that they do not influence our train of thought, our moods: the memory of these unconscious reproductions can have a profound effect... (p. 17).

In arguing for the unconscious influence of weak memory traces, Korsakoff seems to have had in mind something like Leibniz's notion of petites perceptions or Herbart's notion of the limen, except in the domain of memory as opposed to sensation and perception. In much the same manner as "subliminal" sensations and percepts, memory traces, lying below some threshold for consciousness, can influence a person's experience, thought, or action outside do the same thing outside conscious awareness and conscious control.

The notion of weak, "subliminal" memory traces did not survive in the form that Korsakoff articulated it, but it found expression in a later hypothesis, discussed by Warrington and Weiskrantz themselves. One way of interpreting the differences between free recall, cued recall, and recognition is in terms of the differential sensitivity of these tests to "weak" memory traces, where "weakness" is defined in terms of the amount of information about the event contained in the memory trace. Free recall can gain access only to "strong" traces that contain a great deal of information. Cued recall and recognition, by supplying more information in the queries that instigate retrieval, compensate for information lacking in the traces themselves, and thus permit access to memories that would otherwise be lost. Perhaps, then, priming tests succeed in gaining access to memories because they contain enough cue information to overcome the informational weaknesses in the memory traces themselves.

Unfortunately, such an argument does not survive close analysis. First, as Warrington and Weiskrantz themselves noted, fragment-completion performance is better than item recognition in amnesics, even though by any standard recognition provides richer, more informative retrieval cues than fragment-completion does. Recognition by nonamnesics is not always superior to cued recall, but the fact that amnesics perform poorly on memory tests that provide both less (free recall) and more (recognition) cue information than stem- or fragment completion strongly suggests that the dissociation between explicit and implicit memory is not simply a matter of having enough cue information to permit access to weak memory traces.

Moreover, Graf and his colleagues, following up on an informal report by Warrington and Weiskrantz, showed that the same cues could lead to either good or poor memory performance, depending on the precise nature of the instructions put to the subjects. When amnesics are asked to complete word stems with items from a previously studied word list, they perform more poorly than normal controls. But when the patients are asked to complete word stems with the first word that comes to mind, they are just as likely to generate a word from the study list. The preservation of implicit memory, then, is not simply a function of the cues available to the subject. Rather, the issue is whether the subject is required to consciously recollect the past. This, of course, returns us to the original question: What is it about conscious recollection? Or about unconscious memory, for that matter?


Modern Theories of Implicit Memory

There are almost as many theories of implicit memory as there are researchers studying the phenomenon. However, the extant theories can be grouped into a relatively small number of categories that continue to compete with each other.


The Activation View

One popular early theory was that implicit memory reflected the activation of knowledge already stored in memory. This activation theory has its roots in a general class of "associative network" models of memory. According to these theories, memories are created by establishing associative links between nodes that represent knowledge already stored in memory. Thus, an event like A hippie touched a debutante in the park on Thursday would be encoded in episodic memory by activating nodes in semantic memory representing the concepts hippie, debutante, and touch, along with other nodes representing the episodic context in which the event took place (e.g., park and Thursday). These nodes are already established in semantic memory by virtue of prior learning about such objects as hippies and debutantes. In episodic memory, these nodes are linked together, under a new node that represents the event as a whole. Later, a query like What did the hippie do? would activate the "hippie" node. Activation then spreads to other nodes associated with it in the semantic network, such as long hair and VW bus, permitting the person to retrieve abstract, conceptual knowledge about hippies. If activation also spreads to the node representing the event in which the hippie touched the debutante, the episodic memory would be retrieved. But if it did not, then episodic retrieval would fail.

Even if episodic retrieval failed, however, priming effects could still occur within such a system. For example, the initial presentation of the sentence, The hippie touched the debutante, would activate a node representing hippie, and so long as that node remained active, it would be easier for the subject to process that word on perceptual identification or lexical decision tasks, and more likely for the subject to generate that word (as opposed to hippopotamus) when presented with the stem hip___. Activation of old knowledge results in priming, regardless of whether the person can remember that the new fact that the hippie touched the debutante. Now, if we assume that amnesia results from a failure to encode and retain new associations but has no effect on associations established before the onset of the amnesia, and that priming reflects only the activation of old associations, it is clear that amnesics should perform normally on priming tasks, in spite of their failures of conscious recollection.

The appeal of the activation view is that it provides an easy account of spared priming in amnesia: priming occurs because the amnesia does not affect the activation of the pre-existing knowledge out of which new memories are constructed. Early evidence favoring the activation view of implicit memory was obtained in an ingenious experiment conducted by Diamond and Rozin, employing patients with Korsakoff syndrome. The patients, and intact controls, were first given six study trials with three different types of word lists: paired associates of monosyllabic words, such as late-man; familiar disyllabic words such as baby and certain, and disyllabic pseudowords formed by recombining the disyllabic words to form unfamiliar but pronouncable strings such as batain and cerby. Each study trial was followed by a memory test, in which the subjects were asked for free recall of the list items.

001DiaRoz.JPG
                (27279 bytes)The figure shows the results. Not surprisingly, the amnesic subjects were severely impaired in recalling the items of the study lists. However, on some trials the subjects were presented with the first syllable of the disyllabic words and pseudowords, and asked to guess at the associated second syllable, as in the experiments by Warrington and Weiskrantz. On this test, the amnesic patients performed much better, almost at the level of the controls -- but only on the disyllabic words, not on the disyllabic pseudowords. Priming occurred, but only for old, familiar words. This is support for the activation view of priming: priming does not occur for the pseudowords, because there are no corresponding representations available in semantic memory to be activated during encoding.

002SquireLong.JPG (26201 bytes)A subsequent experiment showed that performance on the priming test diminished over a two-hour retention interval. Later experiments confirmed this result. For example, Squire et al. studied the performance of amnesics and controls on a number of different priming tasks: in Experiment 1, stem completion where there were several possible completions for each stem (e.g., mot_____ could be completed by motel, the target, or lures such as mother or motor); in Experiment 2, stem completion where there was only one possible completion (e.g., jui_____ can be completed only by juice); and in Experiment 3, fragment completion in which there was only one possible completion (e.g., a__a__in can be completed only by assassin). The figure shows the performance of the amnesic patients in those conditions involving the "deepest" processing at the time of encoding. In all three experiments, there was substantial priming on the immediate test, which fell back to baseline levels after only two hours. By contrast, priming in control subjects (not shown in the figure) was still apparent after four days, at least in Experiments 2 and 3. Squire et al. plausibly explained the appearance of long-term priming in the control subjects as an artifact of explicit memory: they could deliberately use their conscious recollection of the study phase to help them complete the stems and fragments. The amnesic subjects, however, having little or no conscious recollection, had to rely exclusively on priming -- priming which dissipated in fairly short order.

Rozin dubbed this phenomenon the "hot tubes effect", by analogy to a vacuum-tube radio or cathode-ray television set or computer monitor. Such devices take some time to warm up, but if they are turned off and immediately turned on again they come on more quickly than they did the first time, because the tubes are still warm. If there is a long delay before the appliance is switched on again, the warm-up takes longer, because the tubes are cold. If implicit memory is based on this sort of activation, it should dissipate over time -- as indeed it seems to do.

The activation view was more formally expressed by Mandler's (1980), who argued that item recognition was mediated by two processes.  

The same activation that generates the feeling of familiarity also generates priming effects -- a point to which we will return later.


The Memory-Systems View

The activation view provides a reasonable account of the basic dissociation between explicit and implicit memory: priming is intact despite impaired recall of events, it occurs only for old memories, and it decays over time. However, some researchers found that priming could persist over relatively long delays. 

SquireLongDelay.JPG (51105 bytes) For example, Squire and his associates (1987) found that priming in amnesic patients could last for as long as 4 days, when the primes were encoded under "deep" processing conditions.

 

 

003TulLong.JPG
                (28111 bytes) Tulving and his associates, working with the patient K.C., who became amnesic after suffering a head injury in a motorcycle accident, found significant priming on a word-fragment completion task that lasted as long as 12 months! In this experiment, K.C. studied short sentences of the form Vacationer attracted mosquito or Medicine cured hiccup, accompanied by a photograph a scene (such as a person in a boat on a lake or a man in a hospital bed). K.C.'s task was simply to say whether the sentence fit the picture. Shortly thereafter, K.C. was presented with fragments of the last word in the sentence, such as h-c-p, and asked to complete the fragment with an English word. This procedure was repeated five more times at weekly intervals, using the same targets each time, for a total of six session. The figure shows that K.C. performed better with old compared to new targets on each of these sessions -- clear evidence of priming. On the seventh session, K.C. received the fragment completion test without any study phase, to determine whether priming persisted over an interval of one week: It did.

After 14 more sessions, on Session 22, K.C. received a test in which the three-word sentences were presented visually, but the sentence frame presented orally, revealing substantial levels of cross-modal priming. After a 12-month interval, K.C. returned to the laboratory for a final series of tests in which he was presented either with fragments of the target words studied in Session 22 (for a test of perceptual priming), or the sentence frames in which these words were presented (for a test of conceptual priming). Remarkably, K.C. showed significant priming on both tests. Because K.C. is the most densely amnesic patient ever studied -- he remembers absolutely nothing that has happened to him in his entire life -- it seems unlikely that his long-term priming was an artifact of residual explicit memory -- especially over the 1-year interval. However, the persistence of the priming effect may have been a result of the fact that the same material was repeated, in various forms, over many weekly sessions -- a condition known as "spaced practice" that is especially good for long-term retention. Even so, a demonstration that priming can persist for weeks or months under any circumstances suggests that a simple activation account is not enough to account for implicit memory, and that a new approach is in order.

The new theoretical perspective was soon found in the emerging field of cognitive neuroscience. Whereas 19th-century theorists tended to think of the brain as a general information-processing system (not that they used these precise words), observations of brain-damaged patients by neurologists and neuropsychologists quickly led to the conclusion that the brain is modular in structure, with different parts performing different mental and behavioral functions. So, for example, the fact that damage to Broca's area of the frontal lobe and damage to Wernicke's area of the temporal lobe produce different patterns of aphasia suggests that different brain modules are involved in the expression and comprehension of spoken language. Along the same lines, the anterograde amnesia displayed by H.M. and other amnesic patients suggests that the MTMS played an important role in encoding new memories. But because implicit memory is spared in amnesia, it seems that the MTMS was important only for encoding memories in consciously accessible form. Priming is spared in amnesia because it is mediated by some other brain system which remains intact, despite the destruction of the MTMS.

The dominance of the "modularity" viewpoint within contemporary cognitive neuropsychology and cognitive neuroscience has meant that the memory-systems viewpoint has many adherents. Chief among these have been Larry Squire and Stuart Zola-Morgan and Endel Tulving and Daniel Schacter. As an example, the following table presents one classification of brain modules associated with different aspects of memory, as suggested by neuropsychological and brain-imaging studies :

Memory System

Brain System

Conditioning

Cerebellum

Semantic

Prefrontal Cortex, Premotor Cortex

Procedural

Motor Cortex, Supplemental Motor Area

Explicit

Hippocampus, Medial-Temporal Lobe Structures

Implicit

Neocortex

 

Operating in the same spirit, Schacter and his colleagues have offered a somewhat different classification of memory systems :

Memory System

Brain System

Working Memory

Prefrontal Cortex

Semantic Memory

Temporal Lobe

Episodic Memory

Prefrontal Cortex, Medial Temporal Cortex

Perceptual Representation System

Extrastriate Visual Cortex

Procedural Memory

Supplementary Motor Area

In both cases, the brain systems listed as associated with the memory systems should be considered only as samples of a much more complicated reality. For example, Tulving's model of Hemispheric Encoding/Retrieval Asymmetry (HERA) proposes that the encoding of episodic memories is mediated by left prefrontal cortex of the left cerebral hemisphere, while the retrieval of episodic memories is mediated by corresponding structures in the right hemisphere. Similarly, the precise localization of structures associated with the perceptual memory system depends on the sensory modality in which perceptual information is represented. Finally, while motor skills are probably mediated by portions of the motor cortex, other areas are probably involved in cognitive skills. Delineating all the different possibilities is beyond the scope of this book. It is enough that the reader get the general idea of multiple memory systems in the brain.

With respect to explicit and implicit memory, the Squire-Zola and Tulving-Schacter proposals should be viewed as complementary rather than competitive, and in what follows I focus on their work for expository purposes. Squire and Zola have focused on the hippocampus and the other structures of the medial temporal lobe memory system (MTMS) that, they feel, are critical for explicit memory. They also suggest that priming effects are mediated by structures in the neocortex. Tulving and Schacter agree about the role played by the medial temporal lobe memory system in explicit episodic memory, but have proposed that repetition priming is mediated by a number of modality-specific perceptual representation systems located in the neocortex. Perceptual representation systems create modality-specific representations of the perceptual structure, but not the meaning, of a stimulus -- for example, information about what an object looks like, but not information about its category membership; or information about what a word sounds like, but not about its meaning. On this view, the function of the hippocampus and other structures of the MTMS is to bind these separate perceptual representations into a unified representation of the event as a whole. When the MTMS is compromised, this binding function is lost, and with it the ability to explicitly remember past events. However, the perceptual representations themselves remain intact, and can serve as a basis for repetition priming in the absence of conscious recollection.

So far, studies of repetition priming in both brain-damaged patients and normal subjects, supplemented by brain-imaging studies using such techniques as positron emission tomography (PET) and functional magnetic resonance imaging (fMRI) have yielded evidence for at least three different perceptual representation systems. First, there is a visual word-form system, perhaps located in extrastriate cortex of the occipital lobe, which supports priming on visual tasks such as stem- and fragment-completion, perceptual identification, and lexical decision. There is also an auditory word-form system, associated with perisylvian cortex of the temporal lobe, which supports priming on auditory tasks such as the perceptual identification of spoken words embedded in white noise. Finally, there is a structural description system, perhaps based in the inferior portions of the temporal lobes, that represents the spatial relations among the parts of nonverbal objects.


Pictures, Words, and Impossible Figures

004PicWord.JPG
                  (22928 bytes) Initial evidence for a structural description system, ostensibly supporting repetition priming memory for nonverbal materials, came from studies of the priming of pictures and words. For example, Weldon and Roediger studied repetition priming for pictures and words. In the study phase of the experiment, subjects were presented with either familiar words or line-drawings of the objects they represented. In the test phase, they were asked to complete word fragments, or analogous fragments of the pictures. The figure shows the results of the study. Compared to baseline performance, pictures primed picture-fragment completion more than word-fragment completion, while words primed word-fragment completion more than picture-fragment completion. According to the logic of the multiple-memory systems view, this twin dissociation suggests that the same perceptual representation system which "describes" the visual form of printed words cannot be the same perceptual representation system which "describes" the visual form of pictures of the objects these words represent.

006ImpFigs.JPG
                (30370 bytes)The interpretation of repetition priming in terms of perceptual-representations systems is supported by a series of studies of implicit memory for so-called "impossible figures" -- that is, images of objects that cannot actually exist in three-dimensional space, much like the drawings of the Swiss artist M.C. Escher. The figure presents some examples. In one of their experiments, Schacter and Cooper asked subjects to view possible and impossible figures under one of two orienting conditions. In one condition, they were asked to judge whether the object faced to the left or to the right; this instruction was intended to emphasize the global structure of the figure. In another condition, they were asked to judge whether the object had more horizontal or vertical lines -- an instruction intended to emphasize the local structure of the figure. Later, subjects were given explicit and implicit tests of memory for the objects they had viewed earlier. The explicit test was a standard test of yes/no recognition. In the implicit test, the subjects were asked to judge whether the object was "possible" or not, following a very brief exposure of 100 milliseconds. Schacter and Cooper judged that repetition priming on this object-decision task would lead subjects to say that old, studied items were "possible" more often than new, unfamiliar items -- whether they were actually possible or not.

007RecPerf.JPG
                (24785 bytes)The figure summarizes the subjects' performance on the recognition test of explicit memory. The subjects were pretty good at discriminating studied from unstudied items, regardless of either "possibility" or orienting task.

 

 

008ObjDec.JPG
                (26216 bytes)Somewhat different results were obtained with the priming task of implicit memory, displayed in the figure. The object-decision task revealed a clear priming effect for possible objects studied in the "left-right" orienting task, and a weaker priming effect for possible objects studied in the "horizontal-vertical" condition; but there was simply no evidence of priming for impossible objects in either condition. This lack of priming for impossible objects is consistent with the idea that the PRS retains a description of the spatial relations among object components: such a description cannot be achieved for impossible figures, for the simple reason that they are impossible to assemble in three-dimensional space. The fact that there was more priming for possible objects in the left-right task than in the horizontal-vertical task is also consistent with the theory of the PRS, because the left-right task emphasizes the global structure of the object -- that is, the relations among its parts. The horizontal-vertical task, which does not focus on the relations among the horizontal and vertical parts but only asks the subject to count or estimate them, is essentially irrelevant to the function of the PRS.

009PrimeObj.JPG
                (30472 bytes)These results are consistent with the idea that repetition priming depends on a perceptual representation system, of which the structural description system is one example. Conditions that engage the structural description system, such as global judgments of possible objects, produce strong repetition priming effects. Conditions that do not engage the structural description system, such as local judgments of possible objects or judgments of any kind concerning impossible objects, produce little or no priming. Paradoxically, however, a series of studies by Ratcliff and McKoon did find priming of impossible objects in a series of experiments that minimized the influence of explicit memory by requiring subjects to perform a distraction task during the decision phase, or by imposing a fast deadline on their judgments. The figure shows the basic results of their study. When Ratcliff and McKoon replicated Schacter and Cooper's procedures, they got Schacter and Cooper's results: priming for possible but not impossible objects. But recall that Schacter and Cooper also obtained fairly good recognition performance for both types of objects. In other words, their subjects' explicit memory was pretty good. Under the load and deadline conditions that minimized the contribution of explicit memory to the object-decision task, they obtained priming for both possible and impossible objects. Ratcliff and McKoon offer an alternative interpretation of their findings, in terms of response biases created by explicit memory -- specifically, a bias to call objects "possible", whether they are actually possible or not. This interpretation remains controversial, but the fact that repetition priming can be observed for impossible objects after all, even under conditions that minimize the contributions of explicit memory, appear to be inconsistent with the memory-systems view that repetition priming is mediated by a special perceptual representation system.


Modality Specificity? 

The notion that visual and auditory priming are mediated by different perceptual representation systems is supported by the modality effects: priming is reduced when the modality of presentation shifts between study and test phases of an experiment. Moreover, visual priming appears to be reduced when surface features of the words change between study and test, as between different typefaces, while auditory priming is reduced when the voice in which the word is spoken shifts between male and female. Results such as these suggest that the perceptual representations involved in repetition priming can be very specific indeed. Unfortunately, however, the claim for modality specificity, much less hyperspecificity, appears to have been overstated.

In the first place, a modality shift between study and test may reduce implicit memory, but it does not abolish it. In the study by Graf et al., words were presented either visually or aurally at the time of study, and then presented visually at the time of testing. In contrast to the explicit test of stem-cued recall, where there was no effect, shifting modality cut performance on the stem-completion test of implicit memory by about half. But there was still significant priming, despite the modality shift. In fact, comprehensive reviews of this literature have revealed few instances where a shift in modality abolished priming entirely; mostly, priming was simply reduced when the modalities did not match between study and test.

005ModShift.JPG
                (25833 bytes)A study by Rajaram and Roediger illustrates the effects of modality shift. In this study, study items were presented in the form of printed words, spoken words, or pictures of the objects represented by the words. This was followed by one of four tests of implicit memory: word-stem completion, word-fragment completion, perceptual identification, and anagram solution. The results are presented in the figure. Compared to visual presentation, auditory presentation reduced, but did not abolish priming. Compared to visual presentation of words, however, visual presentation of the corresponding pictures virtually eliminated any evidence of priming. In fact, the virtual absence of picture-to-word or word-to-picture priming is one of the best pieces of evidence favoring the involvement of low-level representations (as in the multiple-systems view), or low-level processing (as in the transfer-appropriate processing view) in implicit memory.

At the same time, it should be understood, again, that almost all of the research on modality specificity has used one version or another of repetition priming as the measure of implicit memory: stem- and fragment-completion, lexical decision, or perceptual identification. Arguably, repetition priming can be mediated by a relatively low-level perception-based representation that is more or less closely tied to the modality of presentation, thus as it might be to other item-specific information such as typeface. Even in the study of Rajaram and Roediger, however, modality shift had relatively little impact on the least "perceptual" of the tests employed. To be sure, a test item like lepanthe has all the same letters as the study item elephant, but by any standard its physical resemblance has got to be less than for corresponding word stems (ele_____) or fragments (e_e_h__t). Priming on word-stem and word-fragment completion tests may be mediated by a relatively low-level perceptual representation that is relatively closely tied to the physical appearance of the studied item, but priming on an anagram completion test is less so.

When evaluating claims for the modality specificity of implicit memory, it is important to bear in mind that repetition priming is not all there is to implicit memory. Implicit memory includes various forms of semantic priming, which go beyond perception-based representations and bottom-up, data-driven processes. Precisely because semantic priming is mediated by meaning-based representations that do not preserve information about perceptual attributes, modality specificity -- much less hyperspecificity is unlikely to be observed. Unfortunately, reviews of this literature few studies of the effects of picture-word shifts, auditory-visual shifts, on semantic priming. In one such study, Srinivas and Roediger compared the effects of visual and auditory presentation on visual tests of anagram solution and category generation. Modality shift reduced repetition priming somewhat, as it did in the later study by Rajaram and Roediger already discussed. However, it had no effect on semantic priming.

The lack of research on modality effects on semantic priming is yet another example of how theories of implicit memory have been distorted by the field's narrow focus on repetition priming. Much is made of the role of low-level, perceptual representations and processes on implicit memory, but the only way implicit memory is measured is with tasks that cannot help but be affected by low-level, perceptual representations and processes. Investigators in the field rarely look at anything else. In fact, within the domain of implicit memory it seems that we will have to distinguish at least three types of priming:

Repetition priming, because it requires only perceptual analysis of the priming event, and depends only on the encoding of a perception-based representation, may be more or less modality specific. But the other forms of priming, because they go beyond mere perception, are likely to be more or less independent of the degree of overlap between the modality of study and the modality of test.

Even if it were to be shown conclusively that repetition priming were mediated by one or more perceptual representation systems, it would not be correct to conclude that perceptual representation systems mediate implicit memory in general. This is because repetition priming is not the only form that implicit memory can take. Amnesic patients can show semantic priming as well. Semantic priming cannot be mediated by a perceptual representation system, for the simple reason that semantic priming is mediated by meaning, and perception-based representations do not contain information about the meanings of the things they represent. In order for semantic priming to occur, memory processing must go beyond physical appearance and spatiotemporal relationships to deal with semantic and conceptual relationships as well. If so, then separate memory systems are required to deal with various forms of implicit memory -- repetition priming might well be mediated by one or another perceptual representation system, but semantic priming would require mediation by something closer to semantic memory.

This same semantic memory system presumably allows amnesic patients to acquire new semantic and conceptual knowledge, as when Claparede's patient learned that people sometimes hide pins in their hands, or Schacter's patients learned that Bob Hope's father was a fireman. By the same token, a preserved procedural memory system would allow patients to acquire new mental and motor skills such as pursuit-rotor learning or mirror-reversed reading. But in either case, damage to the brain system responsible for explicit, episodic memory would prevent these patients from remembering the learning experiences through which they acquired this semantic and procedural knowledge. In addition, they might lack metamemory for the fact that they possessed this knowledge at all.


The Transfer-Appropriate Processing View

The memory-systems view offers a powerful account of dissociations between explicit memory and various forms of implicit memory, major evidence for the operation of these separate memory systems comes from experimental dissociations -- the same dissociations that gave rise to the distinction between explicit and implicit memory in the first place. And that is one of the problems: every time we encounter a new dissociation we are tempted to postulate a new memory system or two to account for it. Although Schacter and his colleagues have offered supplementary criteria for documenting memory systems, dissociations still "constitute necessary conditions for postulating memory systems". The problem is that there are lots of dissociations, so there is a danger of proliferating memory systems uncontrollably. Accordingly, some theorists have attempted to account for experimental dissociations without postulating a vast array of memory systems, and without falling back on the "hot tubes effect" of activation.

Accordingly, Roediger and his colleagues have interpreted dissociations between explicit and implicit memory, and within implicit memory, in terms of transfer-appropriate processing, or TAP. The principle of transfer-appropriate processing states that memory is best when the cognitive processes deployed at the time of retrieval match those that had been engaged at the time of encoding. The implication is that dissociations occur when the tests in question require different kinds of processing. Specifically, Roediger noted that most tests of implicit memory, such as stem- and fragment-completion, are perceptually driven, in that they require access only to information about the surface features of an object -- the sort of information given by "bottom-up" or "data-driven" processing. By contrast, all tests of explicit memory, such as free recall or recognition, are conceptually driven, in that they require access to contextual or semantic information given by "top-down" or "symbolic" processing. Dissociations between explicit and implicit memory tests occur, not because performance on them is mediated by different memory systems, but because they require different sorts of processes.

On the assumption that explicit memory tests are conceptually driven while implicit memory tests are perceptually driven, the transfer-appropriate processing view can easily account for the typical dissociation between explicit and implicit memory. But it has the extra advantage that it can also account for dissociations between tasks within each domain, such as the dissociation between word-fragment-completion and picture-fragment-completion described earlier. Words require verbal processing while pictures require nonverbal processing. According to TAP, then, priming effects on the completion of word fragments should be dissociable from priming effects on the completion of picture fragments -- as indeed they are. But of course, this dissociation might also be explained by the theory of multiple memory systems: the two fragment-completion tasks are dissociable because words and pictures are processed by different memory systems.

Stronger evidence for TAP comes from studies where all the conditions are within the verbal domain. In one such study, Blaxton had subjects study words such as BASHFUL, presented either visually or auditorially. Then the subjects received one of four memory tests: two explicit and two implicit, with one of each pair focusing on perceptual processing and the other requiring conceptual processing. All the tests were presented visually. Examples of the test conditions are:

010ModSfhit2.JPG (27743 bytes)The figure shows the results of the study. As can be seen, performance was better on the data-driven tasks when the modality of presentation (visual) matched the modality of test (visual versus auditory), and this was true regardless of whether the tests were explicit or implicit. But presentation modality had little or no effect on the two conceptual tests. Put another way, the distinction between perceptual and conceptual processing created dissociations between two implicit memory tasks, and between two explicit memory tasks. This pattern of results is not what we would expect if implicit memory were mediated by one memory system and explicit memory by another memory system. But it is exactly what we would expect if performance on perceptual tests of memory, whether explicit or implicit, is enhanced by congruence in processing demands, visual-visual versus auditory-visual, between encoding and retrieval, while performance on conceptual memory tests is unaffected by such congruence. Summarizing a large body of accumulated research, Roediger and McDermott concluded that the TAP view could account for all the dissociations between explicit and implicit memory observed in normal, neurologically intact, human subjects.

A major asset of the TAP view is that it can account for dissociations between implicit and explicit memory tasks without requiring a proliferation of memory systems. All that is needed is a simple distinction between perceptual and conceptual processing: dissociations occur between tasks that require different kinds of processes. However, this doesn't really help us to understand implicit memory. While it is true that implicit memory is usually measured by repetition priming effects of a sort that are arguably mediated by bottom-up, data-driven, perceptual processing, there is more to implicit memory than repetition priming. Implicit memory also extends to semantic priming, and other expressions of memory that appear to require top-down, concept-driven, semantic processing. And, at least so far as this book is concerned, it's the nature of implicit memory that we're trying to explain. It doesn't help to conclude that most instances of explicit memory are the product of conceptual processing, while most instances of implicit memory are the product of perceptual processing. What we really want to know is what makes a memory explicit or implicit. If some explicit memories can be the product of perceptual processing, and some implicit memories can be the product of conceptual processing -- and that is what Blaxton's study clearly shows -- then there has to be something else going on.

At best, the TAP view can explain only the most familiar sort of explicit-implicit dissociations, such as the dissociation between stem-cued recall and stem-completion, or between fragment-cued recall and fragment-completion. But even here TAP proves unsatisfactory, because it is not at all clear that stem-cued recall, or indeed any explicit memory task, really involves conceptual processing. To take the "Conceptual/Explicit" instruction employed by Blaxton, there is nothing particularly conceptual about the instruction "What were the words in the list you studied?". The subjects have not been given any semantic cues that might aid their recall, and they haven't been asked to think about the meaning of the list items in any way. And nothing particularly conceptual about remembering that the word bashful appeared on that list. A successful subject does not even have to know what the word means. True, the likelihood of remembering such an event is increased by semantic processing at the time of encoding. But all that recall or recognition really require is for the subject to remember that bashful was studied at a particular time and in a particular place -- and this information is arguably perceptual in nature. In fact, when Anderson distinguished between perception-based and meaning-based representations, he included knowledge about spatial and temporal relationships in the former category, along with visual images, cognitive maps and information about quantity, cognitive maps, and the like.

Another problem with the TAP view of explicit and implicit memory is that it misses the core feature of the explicit-implicit distinction, which is conscious awareness. Explicit memories are consciously recollected, while implicit memories are not. The distinction between perceptual and conceptual processing does not fit into this framework, for the simple reason that we can be conscious of perceptual experiences without analyzing them for meaning, and even when those experiences are meaningless. For example, abstract art, such as that produced by Jackson Pollock or Mark Rothko, does not "represent" anything at all -- that is what makes it abstract. Viewing such a work is a purely visual experience, and it is quite conscious. While TAP can explain many explicit-implicit dissociations, and has the extra asset that it can also explain dissociations within the explicit and implicit categories, it has nothing to say about the feature of implicit memory that interests us most -- that it is unconscious.


The Process-Dissociation View

One view of explicit and implicit memory that does put consciousness at the center is Jacoby's process-dissociation framework. Jacoby assumes that every task includes both unconscious (automatic) and conscious (controlled) components, and it is the purpose of the process-dissociation procedure (PDP) to measure their relative strengths. With respect to explicit and implicit memory, Jacoby assumes that explicit memory tasks emphasize conscious, controlled, deliberate processing, while implicit tasks emphasize unconscious, automatic, and involuntary processing. Like the transfer-appropriate processing framework, the process dissociation framework is an alternative to a multiple-memory-systems view of explicit and implicit memory; but unlike transfer-appropriate processing, it focuses on the essential distinction between explicit and implicit memory -- that implicit memories, being unconscious, cannot be subject to conscious control. Therefore, they must influence the person's ongoing experience, thought and action automatically.

To illustrate this position, consider a study in which Jacoby and his colleagues have employed the method of opposition to document the automatic influence of implicit memories. In this experiment, subjects began by studying a list of words presented visually under one of two conditions. In the full attention condition, that is all they did; in the divided attention condition, they also detected targets in a string of digits presented auditorially. Then they were engaged in a stem-completion test of memory, where half the stems were drawn from the study list and half were new, unstudied items. In the Inclusion condition, they were instructed to complete each stem with an item from the study list -- or, if that failed, with the first word that came to mind. In the Exclusion condition, they were instructed to complete each stem with any item except one from the study list. The logic of the instructions is that the list items produced in the Inclusion condition include some that are consciously remembered (i.e., explicit memory) plus others that automatically come to mind by virtue of priming (i.e., implicit memory). But the list items produced in the Exclusion condition include only those items that automatically come to mind despite a failure of conscious recollection -- simply because if subjects consciously recollected any of these items, they should have excluded them from the list. In terms of Jacoby's formulas:

 

Inc = C + A(1-R)

 

and

 

Exc = A(1-R)

 

where

Inc = p(targets produced on Inclusion task)

C = p(targets consciously recognized)

(1-R) = p(targets not consciously recognized)

A = p(targets automatically generated)

 

Thus, by algebra:

C = Inc - Exc

A = Exc / (1 - C)

011AutCon1.JPG
                (33797 bytes)The principal results of the experiment are shown in the figure, which also presents the cell means so that interested readers can perform the calculations themselves. Compared to those who studied the list in the divided attention condition, subjects in the full attention condition produced more target items under Inclusion instructions, and fewer targets under Exclusion instructions. This is to be expected, as the full attention maximized encoding, and thus conscious recollection. Applying the algebra of the process dissociation procedure, it appears that dividing attention reduced the controlled component of memory performance to zero, while leaving the automatic component intact. Put another way, all of the targets produced by the subjects in the divided-attention condition, whether under Inclusion or Exclusion instructions, were automatically generated by implicit memory, and none of them were the product of conscious recollection. A later experiment reported by these authors confirmed that the subjects in the divided-attention condition had no conscious recollection of the words they had studied. Given that these subjects were neurologically intact college students, the result is truly stunning.

To some extent, the process-dissociation framework is more of a framework for analyzing data than it is a theory of underlying memory processes. Jacoby believes that proponents of both multiple memory systems and transfer-appropriate processing mistakenly identify tasks with the processes that underlie them. For example, the multiple-systems approach identifies explicit memory with the operations of the medial temporal-lobe memory system, while the transfer-appropriate processing approach seems to identify implicit memory tasks with perceptually driven processing. By contrast, the process-dissociation approach takes as a given the idea that there are two kinds of processes, automatic and controlled, involved in every cognitive task, and seeks only to estimate the contributions of each of these processes to performance in some particular case. In the example cited, the PDP indicated that the memory performance of subjects in the divided-attention condition was mediated entirely by automatic priming effects -- in other words, that the ostensibly memories mediating subjects' performance are really unconscious after all

One problem with the PDP framework, however, is that its mathematics assumes that automatic and controlled processes are independent of each other. This assumption is necessary for the mathematics to work properly, but some commentators have found it unreasonable. For example, in Mandler's two-process theory of memory, memories are encoded by a process in which automatic activation and integration of pre-existing knowledge is followed by effortful elaboration of these knowledge structures to form a complete episodic memory. Unconscious activation and integration can occur without conscious elaboration, but conscious elaboration cannot proceed in the absence of prior unconscious activation and integration. More broadly, unconscious processes may be redundant with conscious ones, such that conscious processes being brings their unconscious predecessors along "for free". Put another way, implicit memory may be independent of explicit memory, but explicit memory may not be independent of implicit memory.

For example, Joordens and Merikle have argued that unconscious processes were redundant with conscious ones: that "conscious processes are always associated with correlated unconscious processes" (p. 463). Because unconscious processing can take place in the absence of conscious processing, but conscious processing brings unconscious processing along for free, we are left with a situation of only partial independence: unconscious processes may operate independently of conscious processes, but conscious processes cannot operate independently of unconscious processes. Based on this assumption, they offered a revised calculation of the strength of unconscious processes.

Whereas Jacoby et al. (1993) estimate unconscious processes as:

A = Exc / (1 - C)

Joordens & Merikle estimate them as:

A = Inc

 

012AuCon2.JPG
                (35510 bytes)Based on the assumption of redundancy, Joordens and Merikle arrived at somewhat different estimates for unconscious influences, as shown in the figure. By their calculation, divided attention reduces both conscious and unconscious contributions to performance.

 

 

Curran and Hintzman have also questioned Jacoby's assumption of independence between controlled and automatic processes. In three of their five experiments, subjects studied words presented either for 1 or 10 seconds, and then completed word stems under standard inclusion and exclusion instructions. When these investigators calculated estimates of A and C according to Jacoby's procedures, they found a significant negative correlation between them: when C increased, A decreased. Thus, in these experiments at least, the controlled and automatic contributions to priming were not independent, but rather were correlated with each other, and the assumption of independence is violated. This finding supports the position of Joordens and Merikle, that independence cannot be assumed -- it must be demonstrated independently. When the automatic and controlled processes underlying task performance are not, in fact, independent, Curran and Hintzman show that application of the process-dissociation procedure can lead to incorrect estimates of their relative contributions.

Jacoby's response to such demurrals is, essentially, to remind his critics that the proof of the conceptual pudding is in the empirical eating. That is, the equations of the process dissociation procedure, based on the independence assumption, yield results that are congruent with those of studies employing measures of explicit and implicit memory. As noted earlier, it is generally agreed that experimental manipulations that have significant effects on explicit memory typically have null or minimal effects on implicit memory. In the same way, experimental manipulations that have significant effects on controlled processes typically have null or minimal effects on automatic processes. In addition to this empirical congruence of results across methodologies, Jacoby argues that the results of the process-dissociation procedure are entirely reasonable, given our theories of how the mind works. For example, automatic processes were unaffected by normal aging. And in the study just described, automatic processes were independent of level of processing. Priming ought to be highly automatic and recall ought to be largely controlled, as indeed they prove to be in Jacoby's work. Jacoby's result, that automatic and unconscious contributions to memory performance are invariant across manipulations of attention, is in line with the traditional definition of automaticity. On the other hand, as both Joordens and Merikle and Curran and Hintzman point out, the assumption of independence is only an assumption -- an assumption, furthermore, that is incorporated into Jacoby's algebra. The findings of Jacoby et al. (1993) cannot be interpreted as independent evidence for the independence assumption, because this assumption is built into the formula that provides the evidence.

In fact, on purely theoretical grounds, the independence assumption seems extremely unlikely to be correct. Consider, for example, how people go about a task such as remembering items previously presented on lists. For more than 25 years, the dominant view of memory retrieval has been some variant of a generate-decide model. According to this model, subjects begin the process of memory retrieval by generating plausible items in response to available retrieval cues, and then deciding whether each candidate meets a criterion for calling it a memory. Thus, after studying a word like motel, subjects who are subsequently presented with the stem mot___ will generate plausible alternatives such as mote, motel, moth, mother, motif, motley, motor, mottle, and motto, and then decide that motel is the word they're actually looking for. The two-process, generate-decide model has been challenged by some theorists, but it retains its appeal because it is the natural way to go about the process of memory retrieval. Perhaps for that reason, it has been incorporated into many computer simulations of memory, such as Human Associative Memory (HAM), a predecessor of ACT and SAM.

Note, however, that Jacoby's instructions in the Inclusion condition do not permit this strategy. Recall that the inclusion condition is intended to represent explicit stem-cued recall, which Jacoby argues is a joint product of automatic and controlled processes. A natural approach to this task, and one in line with the generate-decide strategy, would be to instruct subjects to think of words that begin with the stem, and then determine whether any of them seem familiar -- a variant on the generate-decide strategy. But Jacoby's subjects are not permitted to do this. Instead, they are instructed to use the stem as a cue to remember a list item directly; only if this direct retrieval strategy fails are they allowed to complete the stem with the first word that comes to mind. In other words, Jacoby's instructions require subjects in the inclusion condition to temporarily suppress or ignore whatever is automatically generated by the stem-cues. These task demands, to the extent that subjects can comply with them at all, create a kind of barrier between automatic and controlled processes -- one that is not there in the ordinary course of everyday remembering. Now, experimenters often ask their subjects to do unusual things, in the hope of observing some underlying process in greater detail, but in so doing they risk creating experimental situations that lack ecological validity. By assuming independence in its algebra, and enforcing independence in the instructions given to subjects, Jacoby's process-dissociation procedure may well give a misleading picture of the reasons for the dissociation between explicit and implicit memory.


Testing the Theories

The extant theories of implicit memory can be compared and contrasted in terms of two different dimensions, yielding the two-way classification depicted in the table below.

Mechanism

Memory Systems

Single

Multiple

Activation

Mandler

Squire (?)

Acquisition

Roediger

Jacoby (?)

Schacter & Tulving

Squire (?)

 

The most obvious feature is whether the theory is predicated on the existence of multiple brain systems for memory. The theories proposed by Squire and by Schacter and Tulving obviously propose that explicit and implicit memory are mediated by different memory systems in the brain. By contrast, the theories of Mandler and Roediger at least tacitly assume that explicit and implicit memories reflect the operation of a single memory system, but differ in terms of the requirements that they make on that system. For example, Mandler holds that implicit memory requires only activation and integration, while explicit memory requires elaboration: in his view, priming occurs by virtue of activation. Roediger holds that implicit memory is mostly the product of perceptually driven processes, while explicit memory is mostly the product of conceptually driven processes, as shown by dissociations between different tests of implicit memory. In this respect, Jacoby's process-dissociation view is ambiguous. Jacoby holds that implicit memory is mostly the product of automatic processes (whether perceptual or conceptual in nature), while explicit memory is mostly the product of controlled processes, as evidenced by the results of the process-dissociation procedure. These processes could, in principle, operate within the confines of a single memory system. But because Jacoby insists that the processes are independent, the PDP view is also open, if not friendly, to the multiple-systems view.

The second feature distinguishing among the theories is not quite so obvious, and has to do with the nature of the memory representations that mediate priming. For Mandler, these representations are pre-existing knowledge structures stored in something like semantic memory, which are activated by the priming episode. By contrast, the remaining theories seem to assume that each new experience creates a separate trace in memory, and priming is mediated by this newly acquired trace. In other words, when subjects encounter a word like assassin during the study phase, they encode a unique representation of this experience, independent of all other times when they encountered this word. It is this unique episodic representation, not just a generic representation of the word that has been activated and tagged with some sort of marker, that mediates priming. Although it may seem that the creation of multiple traces of closely similar events is a highly uneconomical way to remember things, in fact a great deal of evidence favors this view of memory representation.

Taking the two dimensions together separates the competing theories into four, distinct categories, one of which may be empty. Mandler's activation view posits that priming occurs by virtue of the activation of pre-existing traces stored in the same memory system that supports conscious recollection. Squire has described priming in terms of activation, but he obviously believes that priming and conscious recollections are mediated by memory traces encoded by, if not in, separate memory systems in the brain. Schacter and Tulving share Squire's multiple-systems view of memory, but also believe that experience, no matter how similar to previous experiences, is represented by a separate trace stored in memory. Roediger and Jacoby also appear to prefer the acquisition view, as part of their overall theoretical emphasis on the role of perceptual processing in the encoding of episodic memories. With respect to multiple memory systems, Roediger is obviously opposed, and highly skeptical of the proliferation of new memory systems with each new dissociation, while Jacoby is somewhat less clear on this point.


Implicit Memory for Novel Information

In the current theoretical environment, most attention has focused on the debate between the multiple-systems and transfer-appropriate processing views, with the general impression that the activation view is passe. However, research on priming for novel information suggests that the activation view still has some life in it. The issue is simple: if priming occurs by virtue of the activation of pre-existing knowledge structures, then we should not be able to obtain priming for items that subjects have never encountered before. In fact, a strict interpretation of the modification view would seem to forbid a person from ever learning something new, and thus require that people enter the world accompanied by an innate fund of semantic and lexical knowledge. The acquisition view permits people to learn about entirely unfamiliar objects and events, which is probably the best reason for preferring it in the first place. But in any event, it would seem impossible to explain priming for truly novel information in terms of the activation of pre-existing memory traces, because by definition such traces cannot exist.

In fact, studies of priming for novel information have yielded a mixed bag of results. 

DiamondRozin2.JPG (53976 bytes) Perhaps the earliest attempt to address it was the study by Diamond and Rozin described earlier. Recall that these investigators obtained significant priming for familiar disyllabic words such as baby and certain, but not for novel disyllabic pseudowords like batain and cerby. This finding is consistent with activation views such as Mandler's. If priming were mediated by newly acquired memory traces, there is no reason why subjects could not have encoded representations of the pseudowords as well as the words. Several subsequent studies also found little or no priming with nonword letter strings, compared to familiar and legal words, although there were also some successes with both normal subjects and amnesic patients. On the nonverbal side of the ledger, Musen and her colleagues found significant priming for random dot patterns, while Schacter, Cooper, and their colleagues have found priming for line drawings of novel, but possible, objects. The success of these studies of nonverbal materials is consistent with acquisition views, but leave open the puzzle of why priming is not observed with novel verbal materials.

013BowLeg.JPG
                (25624 bytes)An interesting perspective on this issue is provided by two papers reporting studies performed independently, at roughly the same time, by investigators working at roughly the same place, and published in successive issues of the same journal. In his first experiment, Bowers asked subjects to study letter strings under a levels-of-processing manipulation. Some of the list items were familiar four-letter words such as KITE; others were novel four-letter non-words, such as KERS, that nonetheless followed the rules of English orthography. As the figure shows, Bowers observed significant priming for both types of items on a subsequent test of perceptual identification. By itself, this finding is compatible with the activation view, because the activation in question is not limited to lexical entries representing words. The activation could also extend to particular letter combinations, that would already be familiar to subjects from other words. Thus, the letter string KERS occurs in such familiar words as BAKERS and FAKERS, not to mention the (Los Angeles) LAKERS. In his Experiment 3, Bowers replicated significant priming for the legal non-words, but he also found significant priming for illegal non-words, such as XYKS, that do not conform to the rules of English. On the assumption that such priming could not be mediated by the activation of either lexical or prelexical items preexisting in memory, Bowers concluded that his results favored the acquisition view.

To some extent, Bowers' experiment brought the literature on verbal priming more closely in line with the literature on nonverbal priming. But as he himself noted, there was a still a gap. Bowers found priming for illegal non-words, but, Schacter, Cooper, and their colleagues have consistently failed to find priming for illegal objects, as represented by those Escher-like "impossible objects" whose surfaces, edges, and contours violate the "rules" of three-dimensional geometry. If the test of the activation and acquisition views is the fate of truly novel objects, and priming for illegal non-words counts as evidence against activation and for acquisition, why doesn't the absence of priming for illegal objects count as evidence for the reverse -- for activation and against acquisition? Moreover, Bowers' evidence for priming of illegal non-words is somewhat ambiguous. Whereas the priming for words and legal non-words was independent of level of processing, priming of illegal non-words was greater with phonetic than with structural encoding. Because level of processing affects conscious recognition, it is possible that the priming of illegal non-words was contaminated by explicit memory. For that reason, Bowers weakened his conclusion, and left open the possibility that "If future experiments fail to obtain priming for illegal non-words or if it is demonstrated that legal and illegal nonword priming effects are mediated by different mechanisms..., then a modification framework will be supported" (p. 546).

As it happened, Dorfman reported just a set of experiments. Whereas Bowers had defined illegal non-words as rather informally, as "letter sequences [that] do not occur within words", Dorfman's approach was informed by structural theories of language. According to linguistic theory, the words of a language are not merely strings of letters arranged in a particular way, but rather are composed of sub-lexical components such as phonemes, morphemes, stems, roots, prefixes, and suffixes. Phonemes are the basic sound units of language, while morphemes are the basic units of meaning. A word like STRANGERS, for example, consists of three morphemes: STRANGE, meaning "unfamiliar", er, MEANING "one who", and S, meaning "more than one"; thus, "strangers" are people who are unfamiliar. Linguistically speaking, "legal" words are not just those whose letter combinations follow the rules of English orthography; rather, "legal" words are those that follow the rules of English phonology and morphology. When we encounter a new word, we understand its meaning by parsing it into its constituent morphemes. And when we coin a new word, we build it up from morphemes whose meanings are already familiar to us.

This example illustrates how learning can occur, if (as theorists like Mandler propose) memories are formed from the activation of pre-existing knowledge. The answer is that new learning always capitalizes on what the learner already knows. The Swiss psychologist Jean Piaget understood this, when he described cognitive development as a process of assimilating experience to pre-existing cognitive schemata, and then adjusting cognitive schemata to accommodate experiences. Similarly, Jerome Bruner, the pioneering cognitive psychologist, asserted that "every act of perception is an act of categorization", meaning that new experiences are processed through a conceptual framework established by prior experiences. In this sense, nothing is ever wholly new: it is only a new combination of the old. Similarly, Mandler's dual-process theory of memory holds that words are encoded through an automatic process of activation and integration of their sub-lexical components, followed by an effortful process of elaboration that establishes new relationships among activated structures. In Mandler's theory, activation and integration form the cognitive basis of implicit memory, while elaboration forms the cognitive basis of explicit memory. But activation and integration come first, and they can't occur unless there's something to activate and integrate.

014DorfLeg1.JPG
                (27050 bytes)In her experiments, Dorfman created three types of novel words. Some, like GENVIVE, were created out of familiar morphemes: GEN, meaning "kind", as in genius, general, and gender; and VIVE, meaning "life", as in survive, revive, and vivify. Others, like FASNEY, were created out of familiar syllables, phoneme combinations that have no particular meaning: FAS, as in fasten, fascinate, and fascist, and NEY, as in chimney, journey, and kidney. A "pseudo-syllabic" nonword, such as ERKTOFE, was composed from letter combinations that were neither morphemes nor syllables in English. The subjects first studied a list of non-words, followed by tests of explicit and implicit memory. For the explicit test, the subjects memory yielded the usual levels effect, with better recognition for items studied under the "deep" processing condition. For her implicit memory test, Dorfman invented a "word-judgment" task in which subjects were presented with two items, one old and one new, and asked to judge which seemed to be a "better" English word; of course, this is a variant on the "mere exposure" paradigm introduced by Zajonc. The figure shows the basic results of her experiments. In her first two experiments, Dorfman found a clear priming effect for both morphemic and syllabic items which was independent of levels of processing. In her third experiment, in which she added the pseudo-syllabic items, however, there was no priming at all. A fourth experiment also found no priming for pseudo-syllabic items; nor did an additional experiment reported in a subsequent paper.

Other results reported by Dorfman found that the priming of morphemic and syllabic pseudowords was not affected by a change in typeface (Experiment 4), but that priming was eliminated when the test items represented repairings or reversals of the sub-lexical components (Experiment 5). Thus, priming for morphemic and syllabic pseudowords appeared to be mediated by integrated representations of the sub-lexical components, not just a perceptual representation of their constituent letter strings.

015DorfLeg2.JPG
                (25513 bytes)One problem with Dorfman's original experiments has to do with her implicit memory task, which asked subjects to judge the extent to which her pseudowords might be good English words. The problem is that pseudo-syllabic items are really bad English words, because they lack both the morphemic and syllabic structure of English. Given the choice between an old and a new pseudo-syllabic pseudoword, it's hard to ask people which one is better, because they're both really awful. Accordingly, it might be that a different implicit memory task, one that is not so tied to linguistic judgments, might reveal priming for truly novel materials after all. Accordingly, Dorfman extended her findings to a more conventional perceptual identification task like that used by Bowers. Instead of judging how good a word a pseudoword was, they were simply asked to identify it when presented briefly followed by a pattern mask. The figure shows that this new experiment replicated the findings of the previous one: both experiments yielded significant priming for morphemic non-words, but neither yielded priming for pseudo-syllabic items (although Experiment 1 showed priming for syllabic pseudowords, Experiment 2 did not).

Taken together, Dorfman's studies show that sub-lexical structure is an important determinant of priming for unfamiliar material such as non-words. She consistently obtained priming for morphemic non-words, sometimes obtained it for syllabic non-words, and never obtained it for pseudo-syllabic pseudowords that were neither morphemic nor syllabic in nature. But sub-lexical components are not the whole story: what is really important is what is made of them by means of integration. Having encoded an integrated pseudoword like GENVIVE, priming occurs for GENVIVE but not for VIVEGEN. It's the "new" representation unit created by activation and integration of "old" sub-lexical components that mediates the priming. These results, in turn, reinforce the point that the notion of "novel information" is something of a misnomer. We always process the new in terms of the familiar, just as Dorfman's subjects processed completely unfamiliar pseudowords in terms of their underlying morphemic and syllabic components. When priming occurs for novel materials, such as pseudowords, it appears to occur by virtue of the activation and integration of more elementary components, such as sub-lexical morphemes and syllables, that already exist in memory. Where the activation and integration of more elementary components cannot occur, because they do not already exist, priming does not occur. This conclusion is in line with the classical activation theory of implicit memory, and contrary to those acquisition theories that predict implicit memory for truly novel materials.


Do We Need Multiple Memory Systems?

By any standard, acquisition views such as transfer-appropriate processing and process dissociation are more popular than the activation view as an alternative to the multiple-systems view of implicit memory. However, Dorfman's research strongly suggests that there is still life in the activation view. What about the multiple-systems view. As even some processing theorists concede, the processing view has difficulty accounting for implicit memory in amnesic patients; and in fact, the most vigorous proponents of the multiple-systems view have been those theorists who have studied dissociations between explicit and implicit memory in brain-damaged patients, where the role of different brain systems stands out in bold relief. Moreover, the memory-systems view can incorporate the processing view, by the simple assumption that certain brain systems perform perceptual processing, while other brain systems perform conceptual processing. So now it is important to confront the memory-systems view directly: are multiple memory systems necessary to account for dissociations between explicit and implicit memory in amnesic patients?

A negative answer to this question comes from an investigative approach to memory and cognition that is often ignored by neuroscientists: the mathematical and computational modeling of memory and other cognitive processes. But if you want to see what can be accomplished, theoretically, if you combine behavioral and computational modeling with neuropsychological and neuroscientific evidence, you'll have to turn to the lectures on Mathematical and Computational Modeling of Learning and Memory.

Believe me, it's worth it.


Is There Explicit Memory Without Implicit Memory?

Because the multiple-systems view holds that explicit and implicit expressions of memory are mediated by separate and independent brain systems, it follows that the dissociation between explicit and implicit memory should go in both directions -- that is, just as we can observe implicit memory without explicit memory, so we ought to be able to observe explicit memory without implicit memory? Evidence for the former dissociation abounds, but is there any evidence for the latter?

020GabRec.JPG
                (23906 bytes)To my knowledge, the only such case that has been reported in the literature as of this writing is M.S., a patient whose right occipital lobe, including all of Brodman's areas 17 and 18, and some of area 19, was surgically removed as a last-ditch treatment for epilepsy (Gabrieli et al., 1995). The surgery succeeded, in that M.S. is now free of seizures and continues to work as a business executive, but of course he has a left homonymous hemianopsia - a blind spot covering his entire left visual field, leaving his right visual field intact. In one experiment, M.S. was asked to read aloud a set of visually presented words (his intact left occipital lobe allowed him to see them clearly), followed by tests of perceptual identification and recognition. As the figure shows, M.S. performed normally on the recognition test, compared to intact normal subjects and patients with focal lesions of the cortex that spared the left occipital lobe; amnesic patients, as expected, performed poorly on recognition.

021GabIdent.JPG
                (25463 bytes)This figure shows corresponding data for a test of perceptual identification, in which studied items and controls were visually presented for a brief period followed by a masking stimulus. In terms of response latencies, the control subjects, as well as the focal-lesion and amnesic patients, all showed priming on this task. However, there was no evidence of priming by M.S. Similar findings were obtained in another study in which L.H., a patient who suffered severe damage to the right occipital lobe in an automobile accident, was compared to the amnesic patient H.M.. Gabrieli and his colleagues concluded that "there was a double dissociation between explicit memory for words (impaired in the amnesic patients but intact in M.S.) and implicit memory for words (impaired in M.S. but intact in the amnesic patients)". These findings were confirmed in a subsequent study.

Double dissociations are the holy grail of cognitive neuropsychology, because they provide the most convincing evidence that performance on the tasks in question is mediated by separate and independent brain systems. However, the first thing to be said about the study of M.S. is that it does not really show a double dissociation. As defined earlier, in a double dissociation a single independent variable exerts opposite effects on two dependent variables. For example, some form of brain damage, or some experimental manipulation, might reduce implicit memory but increase explicit memory. In the case of M.S., what we really have is a case of two single dissociations: one form of brain damage, associated with the amnesic syndrome, impairs explicit memory but spares implicit memory; and another form of brain damage, involving the visual areas of the right occipital cortex, impairs implicit memory but spared explicit memory.

022CatCued.JPG
                (22507 bytes)But even that conclusion goes too far, because it identifies implicit memory with repetition priming, and there is more to implicit memory than that. In fact, a later experiment in the series showed that M.S. had normal levels of implicit memory, so long as he is tested with a conceptual rather than a perceptual task. In this procedure, M.S. and a group of intact control subjects were asked to study atypical exemplars of various conceptual categories, such as banjo or harmonica for musical instruments or cauliflower or radish for vegetables. Using a levels-of-processing manipulation, the subjects made orthographic judgments about half of the items, and semantic judgments about the remainder. For the test of explicit memory, the subjects were presented with the category labels as cues to recall items from the study list. As shown in the figure, M.S. showed normal levels of category-cued recall.

023CatGen.JPG
                (23360 bytes)For the test of implicit memory, the subjects were presented with the category labels, and were asked to list the first examples that came to mind. In contrast to the perceptual identification test employed in the first experiment, both M.S. and the controls showed a significant priming effect, at least for those targets that had been studied under the "deep" processing condition.

 

 

The fact that conceptual priming is spared indicates that M.S. does not, in fact, have impaired implicit memory. The fact that M.S. has impaired perceptual priming and spared conceptual priming is interesting, and may well indicate that performance on these tasks is mediated by separate brain systems, but apparently the same brain system underlying explicit memory is the same as the one involved in conceptual priming. Thus, Gabrieli et al. conclude that "A memory system mediating visual implicit memory... is separable from the memory systems mediating explicit memory for words… and conceptual implicit memory for words" (p. 81). The use of the plural "memory systems" (emphasis added) suggests that Gabrieli et al. believe that separate memory systems mediate conceptual implicit memory and explicit memory, but their study of M.S. provides no evidence for this. Whatever the right-occipital-lobe memory system does is most likely related to a highly specific aspect of visual processing, much like Tulving and Schacter's perceptual representation system. But damage to this area does not preclude other aspects of visual processing. After all, in order to show conceptual priming and explicit memory, M.S. had to be able to read printed words. Nor do the functions of this area, if indeed it exists at all, extend to the whole of implicit memory.


Beyond Repetition Priming

Now that the dissociation between explicit and implicit memory has been demonstrated to the satisfaction of almost everyone (there are always a few naysayers), we now wish to understand the difference between these two expressions of memory. In this respect, it must be said that the field is in a rut. Writing in 1987, Schacter rejected the activation view of Rozin and Mandler, and counterposed the theory of multiple-memory systems which he favored against Roediger's. That is where the field was in 1990, and in 1993, 1994, and in 1997 and that is where we are today. The activation view, once held in low regard, has received a new lease on life by virtue of its ability to explain the circumstances under which implicit memory occurs for novel stimuli. The chief point in favor of the multiple-systems view, other than its ability to bask in the reflected glory of neuroscience, has always been its ability to account for the profound dissociations observed in amnesic patients : even proponents of the transfer-appropriate processing view admit this. Along these same lines, it has proved difficult to apply Jacoby's process-dissociation procedure to analyzing the performance of amnesic patients, for technical reasons having to do with how its calculations are made. But the multiple-systems view is compromised by the successful simulation of amnesic dissociations by computer models of memory that assume just the existence of a single memory storage system.

In such a situation, the temptation is to adopt an eclectic view holding that implicit memory is the product of relatively low-level, automatic processes that create a perception-based representation of an event in memory, while explicit memory is the product of relatively high-level, deliberate processes that create a meaning-based representation and tie it to other representations; and to agree that these perceptual and conceptual processes are to supported by brain systems that are to some extent separate and distinct. Unfortunately, this eclecticism, however much it helps avoid fistfights at scientific conferences, won't do, for the simple reason that there is more to implicit memory than repetition priming. Semantic priming, source amnesia, and the acquisition of new declarative knowledge are all aspects of implicit memory observed in both amnesic patients and normal subjects, and they cannot be explained by such a theory.

Today, more than 15 years after the concept was announced, we are no closer to a theoretical understanding of implicit memory than we were then. One reason for this sad state of affairs is the almost obsessive focus of the field on repetition priming in word-stem and fragment completion, perceptual identification, and lexical decision. The vast bulk of research on implicit memory is concerned with repetition priming. Repetition priming is a very useful paradigm for those investigating perceptual processes, and the discovery of various perceptual representation systems in the brain may be a genuine accomplishment of this line of work. At the same time, it has to be said that repetition priming is the most boring form of memory -- little more than the "memory" of a wire, which, when once bent, is easier to bend again in the same direction. We can hope that, as the study of implicit memory progresses, investigators will broaden their horizons, and consider phenomena of implicit memory other than repetition priming. When they do, their theories are sure to change, and to advance.

At the same time, it has to be said that theoretical development in implicit memory has been retarded by a kind of methodological tunnel vision on the part of investigators. With few exceptions, proponents of multiple memory systems have worked with amnesic patients, while proponents of transfer-appropriate processing and process dissociations have worked with intact subjects. Accordingly, the two groups have tended to work in parallel, with relatively little intersection between their methods, findings, and theories. It is obviously easier to do research with college sophomores, but it also seems plausible to argue that studies of implicit memory in normal subjects, which rely on functional dissociations between the effects of various experimental variables, may well miss features of the phenomenon that are revealed in studies of patients, which rely on population dissociations between patients and non-patients, or between different kinds of patients. This is not to give neuropsychological studies any privileged status in studies of implicit memory; it is just to say that the two camps should talk to each other more, and past each other less.

To make things worse, both groups of theorists tend to ignore the potential contributions of a third methodology, involving mathematical and computational modeling of psychological processes. With the advent of brain-imaging technologies such as PET and fMRI, there is some tendency among cognitive scientists to treat computer simulations as passe -- a wholly abstract exercise with silicon chips and zeroes and ones that has nothing to do with how the brain, with its neurons and synapses, actually works. On the other hand, the work of Nosofsky and Zaki, described in the lectures on Mathematical and Computational Modeling of Memory, has shown clearly that a single unitary memory system can produce precisely the same dissociations, as between recognition and categorization, that are commonly attributed to the operation of separate and independent memory systems. The sad fact is that most computational models of memory have been devised solely with the performance of college sophomores in mind, and have not been extended to take account of phenomena observed best, if not solely, in neurological patients. Perhaps that situation will soon change, so that all three methodological approaches -- behavioral studies of normal human performance, neuropsychological and brain-imaging studies, and mathematical and computational modeling will begin to interact with, and influence, each other. When that happens, we will have better, more comprehensive, more satisfactory theories of implicit memory than we have now.

 

Interactions Between Implicit and Explicit Memory

The idea that explicit and implicit expressions of memory are mediated by different memory systems in the brain is very attractive, especially among neuropsychologists and other neuroscientists, but it has not proved to be easy to support empirically. Most dissociations between explicit and implicit memory can be explained without assuming multiple memory systems, by such principles as transfer-appropriate processing. The fact that implicit memory can be selectively impaired, and explicit memory spared, is another problem, as is the fact that many explicit-implicit dissociations can be simulated in computer models of memory that assume the existence of only one memory system. Another reason to doubt at least a strong version of the multiple-memory systems view is that explicit and implicit memory interact: explicit memory can contribute to performance on implicit memory tasks, and implicit memory can contribute to performance on explicit memory tasks. These interactions are prohibited, if you will, by a theory that explicit and implicit memory are mediated by separate and independent memory systems such as the medial temporal lobe memory system (for explicit memory) and various modality-specific cortical centers (for implicit memory). Independent systems don't interact, for the simple reason that they're independent.

The claim of independence comes from two main lines of evidence. First, there is the fact that performance on tests of explicit and implicit memory is affected by different experimental variables. For example, explicit memory, but not implicit memory, is affected by level of processing at the time of encoding, and implicit memory, but not explicit memory, is affected by shifts in modality between study and test phases. If explicit and implicit memory reflected the functioning of only a single memory system, then they should both be affected, to at least some degree, by the same experimental manipulations. Second, there is the fact that implicit memory can be completely spared even when explicit memory is grossly impaired. If explicit and implicit memory reflected the functioning of a single memory system, any degradation of the one should also impair the other, and any sparing of the one should also facilitate the other.

I have accepted both kinds of dissociations as evidence for unconscious memory: that memories of the past can influence experience, thought, and action in the present, outside of conscious awareness. At the same time, it is important to appreciate that these dissociations are not complete. Explicit and implicit memory interact after all, and in a manner that challenges the multiple-systems view of memory, and supports the view that explicit and implicit memory are products of the same memory system.


The Dangers of Paradigmatic Narrowness

As a prelude to this discussion, I wish to repeat a point made earlier. From the beginning, the study of implicit memory has been dominated by a single paradigm, priming, and only a very specific form of priming, repetition priming, at that. In repetition priming, the prime and target are identical, or at the very least the target is some kind of portion of the prime. In perceptual identification, the prime is presented again, albeit very briefly or degraded by a mask or noisy background, as the target, and must be named by the subject. In word-stem or word-fragment completion, portions of the prime are presented, and must be filled out by the subject. Whether we are talking about population dissociations between amnesics and normals, or functional dissociations involving level of processing or modality shifts, dissociations between stem- or fragment-cued recall, and stem- or fragment-completion, or between perceptual identification or lexical decision, and recognition, constitute the vast bulk of research on implicit memory.

The problem is that there are other kinds of priming besides repetition priming. In particular, there are various forms of semantic or conceptual priming, where the relationship between what is presented to the subject during the study and test phases of the experiment is not at the level of perception, but rather at the level of meaning. Semantic and conceptual priming can also be dissociated from explicit memory, both in cases of amnesia caused by brain damage and in neurologically intact normal subjects. But the nature of this dissociation has not been subject to very much experimental investigation. The debate over implicit memory has been almost completely dominated by studies of repetition effects observed perceptual identification, lexical decision, stem and fragment completion, picture fragment naming, and judgments about objects, and I worry that our theories of implicit memory may have been distorted accordingly.


Levels of Processing and Implicit Memory

A major piece of evidence for the dissociation between explicit and implicit memory is that explicit memory is affected by processing activity at the time of encoding, but implicit memory is not. For example, Jacoby and Dallas had subjects perform "shallow" physical and "deep" semantic processing tasks on a list of items, and found that level of processing affected explicit recognition but not implicit perceptual identification. This is the canonical result in implicit memory experiments, and it has played a major role in theory, and it has been cited by proponents of each of the four views:

Theories aside, the first thing to note is that almost all demonstrations that implicit memory is unaffected by level of processing have employed repetition priming tasks. Given that repetition priming is mediated by the physical resemblance of the prime and the target, something which can be determined by purely perceptual processing at the time of encoding, it should surprise nobody if deep, semantic processing plays a relatively small role in performance on such a task. Roediger and his colleagues made this point somewhat differently, when they pointed out that most explicit memory tasks are conceptually driven, involving elaboration, organization, and meaning analysis, while most implicit memory tasks are data-driven, involving no more than the analysis of perceptual structure. One review of this literature found only 12 studies, involving a total of only 20 experiments, involving semantic tests of implicit memory; this contrasts with 29 studies, and 45 separate experiments, involving perceptual tests. Despite repeated demonstration by Roediger and his colleagues of dissociations between perceptual and semantic tests of implicit memory, which should have told us something about the extent to which we can generalize from the perceptual case, research on implicit memory continues to be tightly focused on repetition priming.

Even in the repetition case, the dissociation observed in levels-of-processing experiments does not appear to be secure. The typical finding in these experiments is that level of processing exerts a statistically significant effect on explicit memory, but not on implicit memory. While this comparison is fair enough, it's also the case that the ability of an experiment to detect a significant effect is a function of the power of the experimental design, and power is affected by such things as the number of subjects run, the number of observations per subject, and the strength of experimental manipulations. Level of processing may affect repetition priming after all, but the effect may be too weak to be detected by any single experiment unless it is a very powerful one. In such a situation, researchers often resort to a powerful statistical technique called meta-analysis to aggregate across the results of several different experiments. Meta-analysis effectively combines the power of several relatively weak experiments into a single, much more powerful one.

c01Chall.JPG (23368
                bytes)Following up on an earlier meta-analysis, Challis and Brodbeck reviewed a large number of different experiments on level of processing and repetition priming (mostly word-fragment completion, some perceptual identification), and found a numerical advantage for deep, semantic processing over shallow, perceptual processing in 33 out of 35 comparisons. The differences were not particularly large in most experiments, but averaged out, there was the clear trend depicted in the figure. Repetition priming is stronger when the primes are encoded under deep processing conditions that encourage semantic analysis, compared to shallow processing conditions that do not.

c02Brown.JPG (25566
                bytes)This conclusion was strengthened by a further analysis that included tests of semantic priming (e.g., on free-association and category-generation tests) as well as repetition priming. The figure shows the essential results of their study, which involved 166 separate outcomes across 38 different studies, and included tests of explicit as well as implicit memory. As can be seen, level of processing affected performance on both perceptual and conceptual tests of explicit memory. Although the levels effect was numerically greater on conceptual than on perceptual effects, the difference was not statistically significant. Level of processing may affect explicit memory more strongly than it affects implicit memory (though Brown and Mitchell did not statistically test this possibility), but it does not appear to be the case, after all, that implicit memory is independent of level of processing. Level of processing affects implicit memory, just as it does explicit memory, it affects both perceptual and conceptual tests of implicit memory, and it does so to roughly the same degree.

This is just what we would expect if explicit and implicit expressions of memory reflected the operation of a single memory system. Of course, it might be that level of processing affects implicit memory because the subjects are strategically using their intact explicit memory to enhance their performance on the implicit memory test. Once subjects realize, for example, that the words being presented on a lexical decision task are often the same ones presented earlier during the study phase of the experiment, then he can draw on his memory for the studied items to aid perceptual identification. For this reason, experimenters often go to great lengths to conceal the relations between the study and test phases, or inquire whether the subjects realized that they were related. However, there is some evidence for a levels effect on perceptual priming even in amnesic patients, who by definition have very poor explicit memory. There is also a levels effect on priming in elderly subjects, who perform more poorly on tests of explicit memory than do young controls. Of course, elderly subjects are not completely amnesic, and even most amnesic subjects retain some residual ability for conscious recollection. So, findings such as these cannot rule out the possibility that performance on implicit memory tests can be supported, to at least some extent, by explicit memory.

Of course, this is precisely the sort of issue for which Jacoby invented the process-dissociation procedure, or PDP. Jacoby's argument is simply that there are no "process pure" tasks, and the PDP is intended to estimate the role of automatic and controlled processes in both explicit and implicit memory tasks. Accordingly, even a task like repetition priming may have one component that is automatic and unconscious, and another that is consciously controlled. The automatic component may be stronger than the controlled component, and the automatic component may be invariant across experimental conditions in a way that the controlled component is not, but that does not mean that there is no controlled component. In this case, it seems likely that when explicit memory is available to subjects, as it is in the usual case, they can use it strategically to aid perceptual recognition. In some cases, explicit memory can help subjects decode specific items from the previously studied list when they appear on the perceptual identification test. Even in the absence of explicit memory for individual items, however, the subjects' recognition that there is a connection between the items being presented for perceptual recognition and the items previously presented during the study phase may lead them to adopt a more appropriate perceptual set that improves perceptual identification performance.


Priming and Recognition

The literature on levels of processing shows one way in which explicit and implicit memory interact: explicit memory can contribute to implicit memory by creating a mental set that facilitates performance on such tasks as word-completion and perceptual identification. What of the other side of the interaction: Does implicit memory make any contribution to explicit memory? It would seem so. Recall that in Mandler's two-process theory of recognition, the judgment of prior occurrence can be mediated by two quite different processes, either the conscious retrieval of a more-or-less full-fledged episodic memory or by inference from a feeling of familiarity, as when someone's face or name "rings a bell" at a party. Jacoby made a similar point when he argued that tasks such as recognition, which is nominally an expression of explicit memory, are in fact not process pure: in his analysis, recognition can be mediated by either conscious recollection or by unconscious familiarity or fluency. For both Mandler and Jacoby, familiarity (or fluency) has its source in priming: the activation of some representation of prior experience, stored in memory. Thus, if recognition can be mediated by familiarity, and the feeling of familiarity has its source in priming, then implicit memory can have an influence on explicit memory.


Priming, Fluency, and Recognition

c03Johnson.JPG
                  (26859 bytes)As an example, consider a series of experiments by Johnston, Dark, and Jacoby on recognition in normal subjects. Subjects read a list of words, followed by a recognition test that also incorporated an element of perceptual identification. In this procedure, the word was originally covered by a screen of 300 dots; these dots were then removed, at a rate of one every 20 milliseconds, until the subject was able to identify the word as old or new (this is essentially a computer-driven variant of the fragment-completion task originally employed by Warrington and Weiskrantz ). The figure shows the basic results of the experiment. Of course, perceptual identification was achieved earlier for old items than for new ones -- this is just another example of priming. More interestingly, however, items judged to be old were identified more readily than those judged to be new, regardless of their actual status. Apparently, the subjects were using perceptual fluency -- the ease with which they were able to decode the degraded words -- as a heuristic for recognizing items: if they could identify the word relatively quickly, it was likely to be old.

c04MandPrime.JPG (27515 bytes)Similar findings were obtained by Mandler, Hamson, and Dorfman in a study that added conceptual fluency to perceptual fluency. In their experiments, subjects studied target words such as coat, accompanied either by a set of four rhyming words (e.g., moat, gloat, bloat, and oat) or five conceptually associated words (e.g., jacket, blouse, pants, and skirt). In one version of the experiment, the subjects rated how much they liked each word (a "deep" orienting task); in the other version, they counted the number of vowels in each word (a "shallow" task). In either experiment, half the subjects then performed a recognition test consisting of the studied targets, rhyming words such as throat, conceptually related words such as sweater, and distractor items that were neither phonemically or conceptually related to the targets. The other half of the subjects performed a stem-completion test with these same words. The figure shows the subjects' response latencies to target items, rhyming lures, and semantic lures. The important point about the graph is that the two lines are precisely parallel. This is what we would expect if priming and recognition were mediated by the same underlying process of activation and integration. This study makes the further point that it is not just perceptual fluency, the ease of perceiving items, that is a cue to recognition. It is also what we might call conceptual fluency -- the ease of bringing them to mind by an act of thought.

c05GrafRec.JPG
                (27205 bytes)An earlier study by Graf and Mandler illustrated another way in which recognition parallels priming. In their Experiment 2, these investigators asked subjects to study a word list and then tested recognition and word-stem completion immediately and after retention intervals of 20 and 90 minutes. As the figure shows, both priming and recognition performance showed similar patterns of decay over the course of the experiment. Another study showed a similar parallelism when the retention interval was extended to one week. This is of course just what we would expect to occur, if the subjects employed the priming-based experience of fluency to make their recognition judgments



The Fluency Heuristic

Of course, fluency -- whether perceptual or conceptual -- is not a wholly reliable guide to recognition. This is a point made by Kahneman and Tversky in their discussion of the availability heuristic used to estimate the frequency or probability of events. In an experiment that is now a classic in the field of judgment and decision-making, these investigators presented subjects with what is known as the fame problem. In the first phase of the experiment, the subjects studied a list of 39 names, either 19 famous women and 20 non-famous men, or the reverse. When they were asked to estimate the frequency with which names appeared on the study list, they generally believed that there were more famous than non-famous names. Similarly, when asked to remember the list items, they recalled more famous than non-famous names. Kahneman and Tversky concluded that people base judgments of frequency or probability on the ease with which relevant instances can be brought to mind. This is not a bad strategy: frequent or likely events should be easy to bring to mind. But Kahneman and Tversky's point is that people can make mistakes in judgment when they ignore other factors besides frequency that can also affect availability -- such as the prior level of familiarity that comes with having a famous name.

Recognition by familiarity or fluency is a variant of the availability heuristic -- it is sometimes even called the fluency heuristic -- and as such it is subject to the same sorts of liabilities. In fact, it is from people's mistakes in judgment that we can infer the heuristic principles on which their judgments are based. In the case of the fluency heuristic for recognition, people can mistakenly attribute to the present what are in reality effects of past experience, as in cases of cryptomnesia or unconscious plagiarism. People can also misattribute fluency to past experience, as in the experience of deja vu. In a clever variant on the fame problem, Jacoby and his colleagues presented subjects with equal numbers of famous and non-famous names, such as Satchel Paige and Sebastian Weisdorf. Later, the subjects were presented with a new list of names, some of which had appeared on the prior study list, and asked to indicate which were the names of famous people. The general finding of these experiments was that people tend to falsely identify previously presented non-famous names as famous. Jacoby interprets this false fame effect in terms of priming, which increases familiarity, which in turn is incorrectly interpreted as evidence of fame.

c06JacWhie.JPG
                (26986 bytes)Another clever experiment also illustrates the influence of priming-based familiarity on recognition. In this study, subjects studied a long list of words, followed by a recognition test. On this recognition test, however, two-thirds of the items were preceded by "context" words that either matched or did not match the word being tested (there were no context words for the remaining test items). More important, half the context words were rendered "subliminal" by masking stimuli that prevented subjects from being aware that the context word had been presented. The figure shows the results of the study for both true and false recognition -- that is, for instances where subjects incorrectly judged new lures to be old targets, and correctly judged old targets to be old targets. Compared to the control condition where no context item was provided at all, subliminal presentation of a matching context word increased false recognition of new, unstudied lures, while subliminal presentation of a non-matching context word reduced this error. Parallel, though weaker, effects were observed for true recognition of old, studied targets. Apparently, subliminal presentation of a matching context word increased perceptual fluency when the subjects saw the supraliminal presentation of the same word, creating a false feeling of familiarity leading to mistaken recognition. It may even have been the case that the subliminal presentation of the non-matching context word made it harder to identify the supraliminal version. When the context words were presented supraliminally, the subjects were able to discount the influence of the context words, and thus reduce their influence on recognition judgments.

Still, precisely because past experiences should feel familiar, unless nature -- or the experimenter -- is being particularly tricky that day, perceptual or conceptual fluency should be a reliable heuristic for making a judgment of prior occurrence -- which is all Mandler's theory requires it to be. In Jacoby's experiments, when subjects are warned that familiarity is an unreliable guide to fame, the false fame effect disappears, or at least diminishes. Nevertheless, when someone's name or face "rings a bell" at a party, it is a good bet that you have met them before. In fact, under ordinary circumstances, recognition-by-familiarity is such a reliable guide to recognition that teachers instruct students taking multiple-choice tests that, when all else fails, they should choose the alternative that seems most familiar to them.


The Puzzle of Recognition in Amnesia

The role that priming plays in recognition may help us to understand one of the persisting puzzles in cognitive neuropsychology, which is that amnesic patients can recognize items that they cannot recall. 

In one early study, Piercy and Huppert (1972) tested two amnesic subjects who were so amnesic that they did not even remember the study phase of the experiment. Nevertheless, they showed almost perfect recognition performance of the unrecalled items.

c07Huppert.JPG
                (26908 bytes)Subsequent studies with larger groups of amnesic patients and controls essentially confirmed these findings, at least for recognition of pictures as opposed to words. The figure shows the results of a representative study (Huppert & Piercy, 1976), in which subjects studied a set of 80 pictures, followed shortly by tests of recall and recognition. Admittedly, picture recognition by amnesics was usually (though not always) inferior to that of controls, but it was far superior to their recall performance. These investigators obtained a similar pattern of results when they tested high-frequency words.

In the second experiment, Huppert & Piercy (1976) had their subjects study pictures on two different days, and were asked to recognize only those pictures that they had seen on the second day. The amnesic patients had a great deal of difficulty with this task, indicating that the patients' recognition judgments were based on a general sense of familiarity rather than retrieval of the context in which the item had been initially presented. Huppert & Piercy concluded that amnesic patients "remember the attributes of an item far more efficiently than information concerning the context in which the item was presented" (p.. 18). They further proposed that amnesia entails "a dissociation between memory for items and memory for context" (p. 18).

Subsequent experiments by Huppert & Piercy (1977, 1978, 1979) confirmed that recognition memory was relatively spared in amnesic patients.

c08Hirst.JPG (27277
                bytes)In a more extensive study, Hirst, Johnson, and their colleagues tested 16 alcoholic and nonalcoholic amnesic patients, and controls, on recognition of categorized and uncategorized word-lists. Some lists were presented at a very fast rate, in an attempt to impair the performance of the controls; other lists were presented at a very slow rate, in an attempt to enhance the performance of the amnesics. The figure shows the basic results. Because the two kinds of amnesic patients performed somewhat differently on the memory tasks, their results are presented separately. In each case, however, the recall and recognition scores are aggregated across presentation rates. Although recall was severely impaired among the amnesic patients, recognition was considerably better. In fact, in those conditions where the presentation rate was showed for amnesics (or, if you will, speeded for controls), levels of recognition were indistinguishable between amnesics and controls, and the only difference between the groups showed up on free recall. Similar findings were obtained in later studies.

Along these same lines, Aggleton and Shaw, reviewing the literature on a standard test of recognition for both words and faces commonly administered to amnesics and other neuropsychological patients, noted that patients with focal lesions of the hippocampus and anterior thalamus performed relatively well on recognition tests, in some cases indistinguishably from normals. Those amnesic patients who perform especially poorly on recognition were those with "collateral damage" -- such as patients with Korsakoff's syndrome, who typically have damage in the frontal lobe as well as the limbic system. Patients with damage to the perirhinal cortex of the temporal lobe and the medial dorsal nucleus of the thalamus also do poorly on recognition. I will return to the implications of this argument later.

The observation that recognition is better than recall in amnesia is interesting, because under many theories of amnesia, spared recognition should not occur.


The Problem of Proportionality

Perhaps for this reason, the claim that recognition is relatively superior to recall in amnesia has not gone unchallenged. The claim that recognition is relatively spared in amnesia, compared to recall, has been disputed by some investigators who have suggested that the patients in the Hirst et al. study may not have been severely amnesic. In fact, however, both their Korsakoff's syndrome patients and their nonalcoholic patients were severely amnesic by any standard. 

c09SquireRec.JPG (24732 bytes) More to the point, the patients tested by Squire and Shimamura, all of whom met a rigorous criterion for amnesia, also showed better recognition than recall. The figure shows the performance of amnesic patients and control subjects on the first trial of the Rey Auditory Verbal Learning Test, in which subjects studied 15 words followed by tests of recognition. Although the amnesic patients performed worse than the controls on both recall and recognition, the amnesic patients performed much better on the recognition test than they had on the recall test -- and that is the essential point. Similar findings were reported by Haist et al., and by Shimamura and Squire in a study that compared recognition with cued rather than free recall.

However, the comparison of recall and recognition may not be that simple. Amnesic patients recognize more than they can recall, but so do neurologically intact individuals who are not amnesic. Squire and Shimamura argued that the proper question is not whether amnesics can recognize items that they cannot recall. Instead, the proper question is whether the relationship between recall and recognition is any different in amnesics than it is in normals -- that is, whether "recognition memory is disproportionately spared in amnesia" (p. 874, emphasis added). In fact, the two studies by Shimamura and Squire found that recognition was no less impaired than free recall or cued recall. In both cases, amnesic recognition was impaired compared to that of controls, and amnesics' performance on the recognition tests paralleled their performance on the recall tests.

c10RecRec.JPG
                (29357 bytes)In a later study, amnesics and controls received recall and recognition tests at varying intervals ranging from 15 seconds to eight weeks. The main results are shown in the figure. As we might expect, that recognition was superior to recall in both groups, and that the controls performed better than the amnesics on both tasks. However, the interaction was not statistically significant. Because the curves for recognition and for recall are essentially parallel, there did not appear to be any differential deficit in recall, compared to recognition, among the amnesic patients.

This point can be made in another way. As can be seen from the figure, after only 15 seconds amnesics showed levels of recall and recognition that were only reached by controls after about 1 day. However, if the lines representing amnesic performance are shifted rightward until they meet, it can be seen that the slopes of the recall and recognition forgetting curves are continuous, and the intercepts for the curves are the same, regardless of whether the subjects are amnesic. Similar findings were obtained for confidence ratings obtained during the recognition test. As Haist et al. concluded that "The striking similarity of recall, recognition, and confidence ratings in both amnesic patients and normal subjects during the course of forgetting suggests that these measures of memory function are equivalently affected in amnesia. Recognition memory is "closely linked" (p. 691) to recall, because recall and recognition are both products of a single same brain system supporting explicit (or, in Squire's terms, declarative) memory.

c11RecAmn.JPG
                (28077 bytes)Hirst et al. had actually anticipated this proposal in their study, which is why they tested memory after both short and long retention intervals. In effect, they hoped to simulate amnesia among neurologically intact subjects by increasing the rate of presentation during the study phase to a point (to a mere half a second per item) where the items are poorly encoded; similarly, by slowing the rate of presentation (to an excruciating 8 seconds per item), they hoped to maximize the likelihood that, despite their poor encoding abilities, amnesic patients would nevertheless be able to remember something of their experience. The figure shows the results. With the patients given 8 seconds to encode each item and the controls only half a second, recognition performance was essentially equated between both groups of amnesics and their respective controls. Working backward, Hirst et al. found that, with levels of recognition equated, recall was even more impaired among the amnesic patients than it was in the nonamnesics. Put another way: compared to recall, recognition is relatively spared in amnesia. Hirst et al. confirmed these results in subsequent research that equated amnesic and nonamnesic recognition by extending the retention interval for controls.

So we are left with a dilemma: Hirst et al. found that amnesia impaired recall more severely than recognition, but Squire and his colleagues did not. This conflict is not likely to be resolved by further studies of the sort performed by Hirst, Squire, and their colleagues, not least because different methods of equating amnesic and control performance give different results. For example, Sullivan Giovanello and Verfaellie found that recall was disproportionately impaired when amnesics were given more study time, as in the Hirst studies, but not when controls were given longer retention intervals, as in the Squire studies. There is no a priori reason to prefer one method over the other, leaving the question in a sort of limbo.


Sensitivity and Bias

Setting aside the question of proportionality, the relative superiority of recognition to recall may be illusory in another way. Recall tests require subjects to actively retrieve items from memory, but subjects can appear to recognize old, target items simply by saying "yes" to every item that is presented to them. In this case, of course, they would also show an equal number of false alarms -- "recognizing" items that had not, in fact, been presented during the study phase. This possibility can be addressed through signal-detection theory, which attempts to separate the sensitivity of an information-processing system from the decision criteria employed by the user of that system. Signal-detection theory was invented to study how people (and, for that matter, other organisms) detected signals (hence its name) presented against a background of noise. Imagine, for example, a radiologist who must decide whether a spot on a lung X-ray is just random noise (which can be safely ignored) or a tumor (which cannot). Whereas earlier approaches to psychophysics implied that the detection of such signals was merely a function of their intensity, and the corresponding sensitivity of the sensory system that processes them. Signal detection theory, by contrast, argued for a less mechanical view, that whether a signal will be detected will depend on the receiver's expectations (in this case, whether the patient is a long-time smoker) and motives (in this case, the consequences for letting a tumor go untreated). In signal-detection experiments, a signal is embedded against a background of noise on some trials, and omitted on others, and the subject is asked to decide each trial whether the signal was present or not -- a difficult judgment usually accompanied by some degree of uncertainty. On any trial, there are four possible outcomes, indicated in the table below:


Response

Signal Status

Present

Absent

"Yes"

Hit

False Alarm

"No"

Miss

Correct Rejection

In signal-detection experiments, the subjects' expectations may be experimentally manipulated by varying the proportion of trials on which the signal is present. If the signal is present 75% of the time, for example, a guess of "signal present" will be right more often than it is wrong, and so will induce a bias for subjects to say "Yes". Similarly, if a signal is present only 25% of the time, the same guess will be wrong more often than it is right, inducing a bias toward saying "No". In practice, signal-detection experiments are usually arranged so that the signal is present about 50% of the time. By the same token, subjects' motivations may be manipulated through a payoff matrix that varies the consequences for hits, false alarms, and other decisions. For example, a scheme in which subjects receive 25 cents for every hit, and are penalized only 5 cents for every false alarm, will induce a bias toward saying "Yes"; the reverse, will induce a bias toward saying "No". Even in the absence of formal experimental manipulations, the pattern of hits and false alarms can be used to derive estimates of two parameters: the discrimination index d' ("d-prime), measuring the true sensitivity of the "operating system" (in this case, the observer or group of observers); and the bias index B ("beta"), assessing the impact of the observer's motives and expectations. Depending on the precise circumstances of the experiment, there are also other indices of sensitivity and bias available for use.

The standard discrimination index d' varies upwards from 0, indicating a total lack of sensitivity -- 50% hits and 50% false alarms. Because of the way it is calculated, the highest possible value of d' is infinite, but a performance record of 99% hits and 1% false alarms, indicating almost perfect discriminative sensitivity, yields a value of 4.65. Negative values, sometimes observed in experiments on subliminal perception indicate that the operating system is performing at less than chance levels. The standard bias index B varies around a value of 1.0, indicating no bias, with lower values indicating a liberal response bias (i.e., toward saying "yes"), and higher values indicating a conservative bias (i.e., toward saying "no"). For example, in the example above, where d' = 4.65, B = 1.0. A pattern of 92% hits and 46% false alarms, where the increase in false alarms reflects a bias toward saying "yes" even when the signal is absent, yields a value for d' of 1.51 and a value for B of 0.37. The opposite pattern of 54% hits and 8% false alarms, reflecting a bias to say "no" even when the signal is present, yields the same value of d', 1.51, but a value for B of 2.67.

As these last two examples illustrate, in principle d' is independent of B, meaning that either index can vary without affecting the other one. That is to say, two operators may have different sensitivities, but perform under the same response biases; alternatively, two operators may have different biases, but the same underlying sensitivity. However, the independence of d' and B should not be overstated. In the first place, the former appears in the formula for calculating the latter. More important, as will become clearer shortly, d' can shift dramatically when subjects employ different criteria for recognition.

Signal-detection theory had its origins in the study of sensory processes, but it has also been employed in the study of memory, as well as in medical diagnosis. In memory experiments, studied items are the signals, so hits occur when subjects say "old" to targets and false alarms occur when subjects say "old" to lures. Values of d', B, and similar parameters can proceed for an event-recognition task just as they do for a signal-detection task. 

c12HitFalse.JPG
                (27503 bytes) Snodrass and Corwin, analyzed recognition data collected by Butters and his colleagues, where patients with amnesia and Huntington's disease (a degenerative disorder which affects the motor system, and, in its later stages, produces a dementia similar to Alzheimer's disease), and their controls, completed a memory test similar to the one employed by Squire and Shimamura. The figure shows the pattern of hits and false alarms observed in each of these groups. Compared to controls, the amnesic patients had fewer hits and more false alarms -- not a pattern that could occur if subjects were saying "yes" to everything. While their performance was worse than that of the controls, with respect to both hits (fewer) and false alarms (more) the amnesics still achieved a positive value for d' -- 1.58 (compared to 3.35 for the controls and 2.26 for the demented patients). Equally important, Snodgrass and Corwin calculated values for Br, a variant on B, and found that amnesics, like the controls, displayed a somewhat conservative response bias that contrasted to the somewhat liberal response bias of the demented subjects. So, amnesic patients do not achieve their recognition performance by saying "yes" to everything. Instead, they are making valid, if imperfect, discriminations between old and new items, and if anything, they are quite conservative in their decision-making processes.


The Process(es) of Recognition

So assuming that recognition is relatively spared in amnesia, compared to recall, how do amnesic patients accomplish it? To answer this question requires a closer examination of the cognitive processes involved in these two expressions of explicit memory. Early theories of memory generally assumed only a quantitative difference between the two expressions of memory: recall is more difficult than recognition because it requires a stronger memory trace. More recently, Tulving and others have offered an alternative single-process view: recognition is easier than recall because there is more overlap between the information contained in cues presented at the time of retrieval and information encoded as part of the original memory trace. Both these views, like Squires', predict that amnesia should have the same (proportional) impact on recall and recognition.

On the other hand, there are also "two-process" theories that suggest that there are qualitative differences between recall and recognition. The classic two-process view is the generate-decide theory of Anderson and Bower, which has its roots in the writings of William James. According to this viewpoint, memory retrieval involves two processes: after the person generates candidate memories, he or she then decides which of these candidates fits the requirements of the retrieval task at hand. Recall entails both the generation and decision stages; but in recognition, the candidate items are supplied by the test itself, and the person need only decide which ones are correct. If amnesia affected the decision phase of retrieval, it would affect recall and recognition alike. But if amnesia affected only the generation phase, it would selectively impair recall, and leave recognition unimpaired.

To complicate things further, George Mandler has proposed that there are two processes involved in recognition as well as recall. As described earlier, Mandler's theory begins with the premise that recognition is a judgment of prior occurrence -- one that can be based on two quite different processes, retrieval and familiarity. While retrieval reflects the conscious recollection of a previous event in all its episodic glory, familiarity can be affected by the priming-based feeling of familiarity. The implication of Mandler's theory is that if priming is spared in amnesia, then we ought to see preserved recognition as well, provided that the amnesics are encouraged to capitalize on the phenomenal salience which is part and parcel of what priming is all about. Under such circumstances, recognition will be relatively spared, compared to recall. But if amnesics are required to base their judgments on retrieval, they will show the same deficit in recognition as they do in recall.


Shifting Criteria for Recognition

The importance of considering the basis of the subject's recognition judgment is illustrated by the phenomenon of posthypnotic amnesia. Like the amnesic syndrome, posthypnotic amnesia also involves an inability to remember past experiences; it differs, however, in that it is a functional amnesia, not caused by brain insult, injury, or disease. Posthypnotic amnesia spares priming and other aspects of implicit memory, just as the amnesic syndrome does. It also differentially impairs recall and recognition memory. 

Consider a study (as yet unpublished) in which hypnotized subjects memorized a list of 16 items, and then received a suggestion for posthypnotic amnesia. After the termination of hypnosis, recognition was tested with a four-point confidence scale:

Subjects responded on a 4-point confidence-rating scale:

"1" if they were certain that the word had not been in the study list;
"2" if they thought it was not, but were not completely sure;
"3" if they thought it was on the study list, but were not completely sure;
and "4" if they were certain that the item was on the study list indicated.

Such a scale yields three criteria for recognition:

Strict, counting only the 4s;
Moderate
, counting the 3s as well;
and
Liberal, counting even 2s.

c13RecPHA.JPG
                (23576 bytes)The figure shows the subjects' performance, in terms of both hits (old, studied items identified as such) and false alarms (new, unstudied lures incorrectly identified as old). It is apparent that hits increase appreciably as the criterion for recognition is loosened. Of course, this could happen if subjects just said "yes" to everything, regardless of whether it was old or not. In fact, false alarms also increased, but not at nearly the same rate, so that there is a genuine improvement in recognition memory. Speaking the language of signal-detection theory, there is a genuine increase in d', representing the sensitivity of recognition, from 1.73 to 2.92 to 3.54. What seems to have happened is that these subjects added a criterion of familiarity to that of retrieval, capitalizing on feelings of perceptual and conceptual fluency as they made judgments about the past.

c14ExpImpECT.JPG (27919 bytes)The contribution of priming to recognition is more directly illustrated by an experiment conducted by Jennifer Dorfman and her colleagues, on the retrograde amnesia associated with electroconvulsive therapy. ECT induces both an anterograde and a retrograde amnesia in the patients who receive it. Previous research showed that the anterograde amnesia affected explicit memory, but spared implicit memory, so Dorfman's study focused on the retrograde component. A group of patients receiving ECT for depression studied a word list immediately before one of their treatments. Once they regained consciousness in the recovery room, they were presented with three-letter stems of list items and matched controls. For the explicit test, they were asked to recall an item from the previously studied word list that began with the stem. For the implicit test, they were asked simply to report the first word that came to mind that began with that stem. As expected, the patients showed a dissociation between explicit and implicit memory: a profound deficit cued recall (compared to those who did not receive ECT), but strong priming on a test of stem completion. Dorfman's finding extended to the retrograde component of ECT-induced amnesia the dissociation between explicit and implicit memory found in the anterograde component by Squire et al. (1985).

c15RecECT.JPG
                (24513 bytes)In DorfmanDPrime.JPG (35427 bytes) the next phase of Dorfman's experiments the patients received a test of recognition: for half the items, they were instructed to adopt a very conservative criterion, saying yes only if they were fairly sure that the item was on the study list; for the other half, they were encouraged to adopt a more liberal criterion, saying yes if the item rang a bell, even if they were uncertain. The figure shows the results. When the subjects shifted criterion, hits went up but false alarms stayed constant, resulting in another significant increase in d'. This genuine improvement in recognition memory occurred, Dorfman et al. concluded, because the subjects were encouraged to capitalize on the experience of perceptual and conceptual fluency that is the phenomenal accompaniment to priming. 

The notion that amnesic patients can strategically capitalize on the experience of priming to aid their performance on a recognition test is in line with Mandler's (1980) two-process theory of recognition. Further, it appears to challenge the strong view that recognition (an expression of explicit memory) and priming (an expression of implicit memory) are independent of each other because they are mediated by separate and independent memory systems in the brain.


The Denial of Priming in Recognition

If priming can contribute to recognition performance, then the memory systems underlying explicit and implicit memory are not independent after all. Implicit memory may be independent of explicit memory, at least in the limited sense that implicit memory is spared in amnesia, where explicit memory is grossly impaired. But if amnesic patients can make use of priming-based fluency to enhance their performance on recognition tasks, it would seem that explicit memory is not independent of implicit memory. It is no surprise, then, that Dorfman's experiment, and its implications, came under challenge from some proponents of the multiple-systems theory of implicit memory.

c16CritShift.JPG (26550 bytes)For example, Reber and Squire reported a series of three experiments on patients with the amnesic syndrome that sought to cast doubt on Dorfman et al.'s ECT findings and conclusions.

 

 

Accordingly, Reber and Squire concluded that "the phenomenon reported by Dorfman et al. does not generalize well" and that "nondeclarative memory (priming) is independent of recognition memory and does not contribute to recognition memory scores" (p. 510).

Despite the fact that Reber and Squire worded their conclusions rather strongly, the actual empirical findings on which their conclusions are based are anomalous at best. In the first place, their first two experiments did not replicate Dorfman's procedures. In her study, Dorfman employed a specific test order so as to encourage a familiarity-based mode of response on the low-criterion recognition test and a retrieval-based mode of response on the high-criterion test. Specifically, the stem-completion test of implicit memory preceded the low-criterion recognition test, and the stem-cued-recall test of explicit memory preceded the high-criterion recognition. In contrast, Reber and Squire conducted both recognition tests first, before the priming test. In fact, the two tests were conducted in separate experiments conducted almost a year and a half apart (p. 505)! Reber and Squire's failure to replicate Dorfman's procedures makes it impossible to interpret their failure to replicate her results. Reber and Squire's third experiment did replicate Dorfman's procedures, but unaccountably failed to produce priming on the stem completion test. Because the hypothesis of recognition by familiarity assumes that priming will be at least relatively spared, the failure to observe priming in Reber and Squire's third experiment vitiates that experiment as a test of Dorfman et al.'s hypothesis.

Reber and Squire also included nonamnesic controls in their experiments. Relaxing the criterion failed to improve recognition for these subjects as well. For several reasons, however, these negative results do not tell us much. In their Experiment 1, Reber and Squire did not demonstrate that priming occurred at the delays of 1-2 days or 1 week employed as the critical control conditions of Experiment 1. It is not that priming did not occur -- they just failed to test for it, just as they failed to test for priming in the amnesic group. Priming simply cannot be assumed to occur over such long delays. Presumably priming would have occurred with immediate testing, but there is certainly no reason for nonamnesic control subjects to differentially draw on familiarity information in the low-criterion test. This is because, at such short retention intervals, nonamnesic subjects remember the items of the study list perfectly well. In Experiment 2, which was focused on demonstrating intact priming in the amnesic patients, there were no nonamnesic controls. And in Experiment 3, as Reber & Squire acknowledge, ceiling effects are again at issue. There is no reason to expect an effect of relaxing the criterion for recognition when recognition is already almost perfect under the strict criterion! "Relaxing the criterion" had no effect on recognition performance in Dorfman et al.'s control subjects either. Nor should it. When recognition is already good, as it was for her controls, there is no reason, or need, for subjects to strategically rely on familiarity to back up recollection.

Setting aside these empirical issues, Reber and Squire implied (p. 509) that the level of priming obtained by Dorfman et al., 26% in the amnesic patients, following a retention interval of 60-90 minutes, is atypical. They cited several studies that observed priming levels of only 3.3-17%, and suggest that Dorfman's higher levels of priming may have been contaminated by residual declarative memory. In fact, however, Dorfman's findings were hardly atypical. For example, a study by McBride and Dosher found priming of 20% in normal subjects at a 60-minute delay. Reber and Squire refer to this paper, but cite only lower priming values obtained over longer intervals than Dorfman used in her study. McBride and Dosher further concluded, based on quantitative modeling of the data, that stem-completion priming decays slowly over intervals of 15-90 minutes, and does not approach baseline levels until delays of 24-48 hours. In fact, stem completion priming may persist even longer than this. Similarly, Craik and his colleagues found 13% priming at a 24-hour delay, and Squire himself has found levels of priming that were similar to Dorfman's in both amnesics and controls.

Reber and Squire's suggestion that Dorfman et al.'s patients may have strategically relied on residual explicit memory to perform the stem completion task -- precisely the opposite of what Dorfman et al. concluded -- is hardly warranted. The cued recall performance of Dorfman et al.'s patients was only 5% correct -- far worse performance than in the Reber and Squire paper itself. So, there was not much residual memory for the amnesic patients to use. Moreover, the control subjects, who were not amnesic, would be just as likely to use residual memory, if not more so: yet they showed weaker priming effects than the amnesics. Dorfman's subjects did not strategically rely on explicit memory to aid their performance on the implicit memory task. But they did strategically rely on implicit memory to aid their performance on the explicit memory task, when encouraged to loosen their criterion for recognition.

c17RecEP.JPG (22830
                bytes)Or did they? Consider the amnesic patient, E.P., who shows intact levels of priming but only chance levels of recognition. The case is interesting, because Squire has pointed out that most amnesic patients are not completely amnesic, and that residual explicit memory may be responsible for the apparent sparing of recognition. E.P. doesn't just have relatively impaired recognition, compared to controls. He doesn't have any recognition at all -- though he does have spared implicit memory. In a new study, Stark and Squire presented a series of five experiments suggesting that shifting criteria failed to improve E.P.'s recognition memory: it remained at chance levels throughout. Therefore, they concluded that priming makes no contribution to recognition performance. However, like the earlier studies of Reber and Squire, this new evidence is ambiguous at best.

For example, in their Experiment 1, speeded recognition judgments did not improve E.P.'s recognition performance (or that of other amnesic subjects). The rationale for this experiment was that priming is automatic, and automatic judgments should be made rapidly -- "off the top of the head", as it were. But this ignores Mandler's point that familiarity-based recognition is an inferential judgment -- a judgment that, presumably, requires time for reflection. Priming is automatic, but making an inference about the past based on feelings of familiarity takes time, as our studies showed using the tripartite remember/know/feel distinction.

In Experiment 2, the recognition test was modified so as to remove any reference to the study phase; instead, E.P. and other amnesic subjects were simply asked which of two words was "more familiar". Again, E.P.'s recognition remained at chance levels. While on the surface such an instruction would seem to enhance familiarity-based recognition, in fact it changes the task entirely. Rather than making judgments about prior occurrence during a study phase, as would be essential for an episodic memory test, the subjects are simply asked to make judgments about the subjective frequency of the word in spoken and written English -- a semantic memory task.

Experiment 3 combined the speeded judgment of Experiment 1 with the generic familiarity judgment of Experiment 2. Again E.P.'s recognition performance failed to improve -- but as the old saying goes, two wrongs don't make a right. The fact that E.P. performed poorly on this test says nothing about whether priming can contribute to recognition judgments in episodic memory, under circumstances where amnesic patients are encouraged to rely on intuitive feelings of familiarity.

In Experiment 4, a trial of stem-completion priming was immediately followed by a test of recognition memory for the targets and lures produced by the stem-completion test -- thus following the spirit of Dorfman's procedure in a way that Reber and Squire had not. Again E.P. showed poor recognition despite intact recognition -- but then again, contrary to Dorfman's procedures, he was not encouraged to make strategic use of the feeling of familiarity when making recognition judgments.

Experiment 5A repeated the procedure of Experiment 4 with a yes/no recognition task as opposed to a forced-choice task. The results were the same, but again E.P. was not encouraged to strategically use the feeling of familiarity to make his recognition judgments. (Experiment 5B was concerned with controls, not amnesics.)

Despite these anomalous findings, Stark and Squire concluded (p 459) that "although recognition memory judgments may be made on the basis of familiarity, repetition priming is not the source of this feeling of familiarity". On the one hand, then, Stark and Squire seem to be conceding that Dorfman et al. were right all along, and that a feeling of familiarity can contribute to recognition performance even among amnesic patients. On the other hand, they continue to insist on a strict separation between priming and recognition, reasserting that the feeling of familiarity has nothing to do with priming. All of which leads to the following questions:

If recognition can be mediated by familiarity, as Stark and Squire concede, and if the feeling of familiarity is unrelated to repetition priming, as they insist, where does the feeling of familiarity come from?

Are there now three memory systems in the brain? -- one system that mediates repetition priming, a second system that mediates retrieval-based explicit memory, and a third system, separate from the first two, that mediates familiarity-based recognition?

Perhaps. The danger here is one that cognitive neuroscience has slipped into before -- the danger of postulating new memory systems to account for every task dissociation revealed by behavioral experiments. Considerations of parsimony may cast doubt on the idea that the brain has so many memory systems, each separate and independent from all of the others. It seems more reasonable to conclude that priming does make a contribution to recognition after all, by creating the feeling of familiarity.


Priming, Familiarity, and Recognition in Amnesia

In arguing to the contrary, Squire and his colleagues ignored a sizable literature from studies of normal subjects, such as those discussed earlier, indicating that priming, giving rise to perceptual and conceptual fluency, can influence recognition judgments. Even more important, they failed to take account of other neuropsychological studies, besides Dorfman's, indicating that priming does in fact contribute to recognition in amnesic patients. 

c18CompRec.JPG
                (29193 bytes) For example, Verfaellie and Treadwell studied the recognition performance of amnesic patients with Jacoby's process dissociation procedure, or PDP.  PDP attempts to separate task performance into its automatic and controlled components -- in this case, recognition by familiarity or fluency and recognition by retrieval or recollection. This experiment employed a levels-of-processing manipulation in which study items were presented either as words to be read (a relatively shallow processing task) or as anagrams to be solved (a relatively deep one). Later, they completed a recognition task under Inclusion or Exclusion instructions. On the inclusion test, which corresponds to a standard recognition task, the amnesics showed some degree of recognition, with roughly 47% hits compared to 32% false alarms. Comparable figures for the control subjects were clearly better, approximately 56% and 17%, respectively, so the amnesic patients were impaired on recognition, but that is not the important point. Based on the data from the exclusion condition, the investigators were able to identify the contributions of fluency and recollection to recognition. As the figure indicates, amnesics were impaired only on recognition by the conscious recollection process. With respect to recognition by the automatic familiarity process, they were indistinguishable from nonamnesic controls. Put another way, recognition-by-recollection differed greatly between levels of processing, and between amnesics and controls, as it should if recollection is closely tied to explicit memory. But recognition-by-familiarity was unaffected by levels of processing, and did not differ between amnesics and controls, as we would expect if familiarity were closely tied to priming and implicit memory. Note that almost all of the correct recognition judgments by amnesics were mediated by priming-based familiarity -- exactly the pattern of performance predicted by Dorfman et al.. Amnesics do not remember much, but because of their spared implicit memory, they can strategically rely on priming-based feelings of familiarity to make relatively accurate judgments about the past, and thus improve their performance on explicit memory tasks.

c19PercepFlu.JPG (32535 bytes)In another study, Verfaellie and Cermak examined the contribution of perceptual fluency, an effect of priming, to amnesics' recognition judgments tasks. During the study phase of their first experiment, the subjects were told that words were being flashed too quickly to permit conscious perception; in fact, however, no words were presented at all -- only a pattern mask. The test phase employed a variant on perceptual identification, in which items were presented within a mask that was gradually removed until the subject identified the word. At this point, the subject was asked whether the word had been presented during the initial study phase. Because no words had actually been presented during the study phase, "yes" responses represented instances of false recognition. For both amnesics and controls, items that were identified more readily during the perceptual identification test were more likely to be (falsely) judged to be old. A second study, in which some words were actually presented during the study phase, without a mask, yielded similar findings. Regardless of whether they were old or new, test items were more likely to be judged to be old if they were identified more quickly. This was especially the case for the amnesic subjects, with the result that amnesics showed higher levels of false recognition than the controls.


The Process-Dissociation Procedure Redux

For a final piece of evidence that recognition can be mediated by priming-based familiarity, let us return to the demonstration of the process-dissociation procedure by Jacoby et al. (1997).  

SchonfeldRobertson.JPG (44539 bytes)The story properly begins with a study of aging memory by Schonfield and Robertson (1966) found the usual age-related impairment in recall, but no impairment -- not a whit -- in recognition.  Now, this in and of itself isn't necessarily a puzzle.  According to two-process theories of retrieval (e.g., Anderson & Bower, 1973), recall requires search and decision, while recognition requires only a decision process.  Therefore, the pattern of recall and recognition performance could be explained in terms of a specific age-related deficit in the search mechanism.  But, as we've seen from the literature on, there's only one retrieval process: cue dependency, qualified by encoding specificity.  So, it might be that recognition is just easier than recall after all, because it supplies so many appropriate retrieval cues, and the elderly fall down on the harder cognitive task.  But that would be boring.  

Fortunately, a less boring possibility is offered by Mandler's (1980) two-process theory of recognition, discussed earlier in the context of the activation view of implicit memory.  Recall that the two processes are retrieval and familiarity, and familiarity is mediated by the same activation of a pre-existing representation that gives rise (in the activation view) to priming effects.  So, the more interesting idea is that aging memory entails an impairment in the retrieval process that leaves the familiarity process unimpaired.  

Evidence in favor of this view comes from the study by Jacoby and his colleagues described earlier, using the process-dissociation procedure (PDP).  Employing the method of opposition, they estimated that the automatic component of recognition (essentially, Mandler's recognition familiarity) was the same for young and old subjects, but that the controlled component (essentially, Mandler's recognition by retrieval) was significantly diminished.  Other investigators, offering alternate concepts of the PDP, which assume that the two processes are redundant instead of independent, confirmed these general trends, if not the precise values assigned to the two components of processing.  

So, it seems that, so long as priming is preserved, amnesic and other memory-impaired individuals can rely on the familiarity component to perform relatively successfully on recognition tasks.  Provided that they are allowed (or even encouraged) to do so, and the experimenter is not fiendishly tricky, subjects who do so will be right more often than they're wrong.

The role of priming-based familiarity in recognition underscores the point that recognition, like signal detection in general, is a matter of judgment, and whether something is recognized will depend to a great degree on the criterion set by the person doing the remembering.  


Recollective Experience

Another perspective on priming and recognition is provided by studies that examine recollective experience, or the phenomenal experience of remembering. 

In a groundbreaking paper, Tulving (1985) distinguished between two quite different recollective experiences:

Whenever subjects endorsed an item as old on a recognition test, they were asked to indicate the nature of their recollective experience. If recognition of a particular item was accompanied by conscious recollection of its occurrence on the study list, they were to rate the item as "remembered"; if not, they were to rate it as "known". In his study, Tulving found that about 88% of all recognition judgments were accompanied by conscious recollection of the study experience. For the remaining 12% of endorsements, the subjects simply "knew" that the item had appeared on the list, much the way they knew that Ontario was a province of Canada (the study was conducted in Canada). More important, Tulving discovered that "remember" judgments were affected by level of processing, but "know" judgments were not. 

Similar findings were obtained by Gardiner (1988), who offered a somewhat different characterization of recollective experience following Mandler's two-process theory of recognition. 

Gardiner1988.JPG (46528 bytes)Gardiner (1988) showed that judgments of remembering and knowing were dissociated by two experimental manipulations: depth of processing (making acoustic vs. semantic judgments) and a read-generate paradigm.  Note the similarity between these dissociations and those affecting explicit and implicit memory.  Gardiner's take on the remember-know distinction has since been extremely influential -- though, as we will see, that does not mean that it is correct. 

 

Yonelinas2001a.JPG (47210 bytes)In Yonelinas2001b.JPG (50787 bytes)an important series of experiments, Andrew Yonelinas (1994, 2001, 2002) has mapped the remember-know distinction onto the distinction between automatic and controlled components of processing.  In an experimental tour de force, Yonelinas (2001) sought converging evidence on this score from three quite different experimental paradigms: (1) remember-know judgments, (2) the process-dissociation procedure (Yonelinas was a student of Larry Jacoby), and (3) signal-detection theory.  The general finding was that a divided-attention manipulation affected all three dependent variables similarly.  Yonelinas concluded that "remembering" and "knowing" were distinguished from each other by three features: 

Yonelinas (1994, 2010) has further offered a dual-process signal-detection model of recollective experience.  

Yonelinas2010.JPG (53753 bytes)A brain-imaging study by Yonelinas et al. (2010) mapped these two processes into different brain systems.




Note that whether it's the distinction between explicit and implicit memory, or the difference between recollection and familiarity, everybody always points to the hippocampus as the mediator of the "big, complex, conscious" thing, and relegates the rest to "the cortex".   

Shimamura2010.JPG (78627 bytes)A more specific  proposal about the role of the cortex in memory is the cortical binding of relational activity (CoBRA) theory proposed by UCB's Arthur Shimamura (2010, 2011).  Shimamura proposes that the various episodic features that characterize a unique event, including information about spatiotemporal context underlying source memory and information about the individual's internal mental state at the time of the event, are stored in various locations around the cortex.  These features are then bound together by mechanisms that are localized in the ventral posterior parietal cortex (vPPC), which acts as a kind of convergence zone connecting the medial temporal lobe with executive structures localized in the prefrontal cortex.  Shimamura's proposal is doubly interesting, because (1) it represents more than "hand-waving" toward the neocortex; and (2) in bring an entirely new set of structures, those associated with the vPPC.

RecollectiveExperienceRevised.JPG (64219 bytes)Although Gardiner's development of what has come to be called the remember-know paradigm was inspired by Tulving's 1985 paper, his conception and Tulving's seem to differ in important ways. For Tulving, the remember-know distinction maps onto his earlier distinction between episodic memory, or the person's awareness that an event "is a veridical part of his own past existence" (1985, p. 3), and semantic memory, or the person's "symbolic knowledge of the world" (1985, p. 3). But for Gardiner, the remember-know distinction maps more closely onto Mandler's distinction between recognition by retrieval and recognition by familiarity. Recognition by retrieval involves remembering an event as an event, including the personal and spatiotemporal context in which the event occurred; by contrast, recognition by familiarity involves a feeling or intuition that some event occurred in the past, in the absence of conscious recollection of that event. An alternative framework for Gardiner's distinction, of course, is provided by the distinction between explicit and implicit memory, which was just emerging at the time. Explicit memory involves the conscious recollection of an experience from the past, while implicit memory is a memory-based change in behavior that occurs independent of, and in the classic case in the absence of, conscious recollection. For Gardiner, then, remembering reflects explicit memory, while knowing reflects implicit memory.

These alternative interpretations of remembering and knowing are connected, from Mandler's point of view, because both recognition by familiarity and implicit memory are based on the activation by an event of previously stored knowledge. Nevertheless, as I have argued elsewhere, it may be a mistake to conflate semantic memory with implicit memory, and knowing with intuiting. Accordingly, I have suggested that there are at least three varieties of recollective experience -- or, if you will, three different memory qualia:

In fact, there may be a fourth memory quale (that's the singular of qualia):

This form of remembering is more like a judgment than a memory, though along with Mandler and Bartlett, it is important to understand that, in the final analysis, all acts of remembering are acts of judgment.

These memory qualia are familiar to anyone who has ever taken a multiple-choice test. Sometimes, we choose a response because we remember the circumstances under which we learned it -- the particular lecture, or, as sometimes happens, the location on the textbook page where the information appeared. On other occasions, we choose a response because we just know the answer -- it's part of our knowledge about the world, and we don't remember the circumstances under which we learned it. On still other occasions, we choose a response because we feel that it is the right one. We don't actually know the answer, and we certainly don't remember where we learned it, but we choose a response because it strikes us intuitively as familiar, and we infer from this feeling of familiarity that, of all the choices available, this is the one that is most likely to be correct. We're guessing, in a way that we're not guessing when we're remembering or knowing, but it's not random guessing; it's guessing informed by the feeling of familiarity -- a feeling of familiarity that comes from priming.

One methodological problem with studies in the Tulving-Gardiner tradition is that "knowing" is, essentially, a wastebasket category.  That is, subjects are asked to identify those items associated with conscious recollection of the original episode (what both Tulving and Gardiner mean by "remembering"); everything else is dumped into the category of "knowing".  But if there is a substantive difference between knowing and feeling, this will be obscured by such instructions.  Accordingly, it is important to give "knowing" a substantive definition -- and that means giving "feeling" a substantive definition as well.

There is now a considerable amount of experimental evidence that remembering, knowing, and feeling, at least, are distinct recollective experiences.  In the first place, subjects can make the distinctions when they are asked to. They understand intuitively the distinctions among remembering, knowing, and feeling something about the past.

LoP-GKnow.JPG (51672 bytes) An experiment by Kihlstrom, Kim, and Dabady (1996, Exp. 1) defined "knowing" as Gardiner does, as essentially whatever is left over after "remembered" items are accounted for.  And we got essentially what Tulving and Gardiner got: depth of processing affected remembering but not knowing (or, as we prefer to label it, G-knowing, or knowing according to Gardiner).

LoP-RKF.JPG (49929 bytes)

But when we gave knowing a substantive definition that distinguishes it from feeling, we got quite different results.  Now, depth of processing affects remembering and knowing in the same way, and the opposite of its effects on feeling.  
RKF-Confidence.JPG (43763 bytes) A subsequent experiment by Kihlstrom and Kim (1997, Exp. 1) found that remember and know judgments were both associated with high confidence levels, compared to feel judgments.
RKF-Latencies.JPG (48074 bytes) And across four experiments, feeling judgments were associated with longer response latencies than remember or know judgments.  This is because remembering-by-feeling really is intuitive, and the subject must take time to make a judgment as to whether to recognize the probe.
RKF-Criterion.JPG (58314 bytes) Shifting subjects from a strict to a liberal criterion of recognition had no effect on remember or know judgments, but did have an effect on feeling judgments.
RKF-Overlearning.JPG (63616 bytes) In all the previous studies, subjects went through only a single study-test trial.  But under conditions of multi-trial free recall, recollective experience shows a clear trend.  On initial trials, most item-recognitions are accompanied by feelings of remembering or feeling.  Over time, however, both of these experiences drop out, and are replaced by experiences of knowing.  By the time 20 learning trials are finished, subjects aren't really remembering list items.  Rather, they know what's on the list, in the same way that they know that Bill Clinton was President of the United States.  

Recollection by remembering and by knowing is affected by level of processing, while recollection by feeling is not. Recollection by feeling is also associated with longer response latencies, and lower confidence levels, than either of the other two modes. And repeated study trials increase the frequency of recollection by knowing, while decreasing the frequency of recollection by remembering or by feeling. However, the vast bulk of the literature on recollective experience employs only the dichotomous distinction between remembering and knowing. As it happens, however, in this literature "knowing" is usually defined in terms closer to Gardiner's construal than Tulving's. Because "remembering" is tantamount to Mandler's "retrieval" process, while "knowing" is tantamount to "familiarity", studies of memory in amnesia that use the remember-know paradigm can be interpreted as bearing on the role of priming-based feeling of familiarity in recognition. In an effort to make things clearer, in what follows I will use the phrase familiarity-based "knowing", to indicate that it is really a euphemism for the feeling of familiarity.

c20RemKnow1.JPG
                (24448 bytes)For example, Knowlton and Squire employed the conventional remember/know paradigm (in which "knowing" really means "feeling") to study recollective experience in amnesia. Although at first blush the very concept of recollective experience in amnesia might strike us as a contradiction in terms, the fact that the question can be raised suggests that amnesic patients can indeed recognize past events with some accuracy: the question then is how they do it. In their study, Knowlton and Squire asked amnesic patients and control subjects to study a list of words, and then administered a test of recognition 10 minutes later. For each item judged to be "old", the subjects were further asked whether their recollective experience was one of "remembering" or "knowing" (Knowlton and Squire checked to make sure that the amnesic subjects understood and carried out these instructions). The figure shows the basic results of the experiment. As might be expected, the amnesic patients recognized fewer targets than the controls. The important point, however, is how they recognized the targets they did. Although Knowlton and Squire claimed that both retrieval-based "remembering" and familiarity-based "knowing" were impaired in amnesic patients, compared to controls, their study also (and critically) yielded a significant interaction. For amnesics, recognition by remembering was more impaired than recognition by feeling-based knowing. This is exactly what we would expect if amnesic patients based their recognition judgments on priming-based feelings of familiarity. But even after one week, control subjects were less reliant on familiarity than the amnesics were after only 10 minutes.

c21RemKnow2.JPG
                (27749 bytes)Similarly, two studies by Schacter, Verfaellie, and their colleagues indicated that amnesics have severely impaired "remember" processes and fairly normal "familiarity" processes. Their studies employed a procedure known to produce the associative memory illusion, or AMI. In a typical AMI experiment, subjects study a list of words, such as thread, pin, eye, sewing, sharp, and point, which are all relatively high-frequency associates of the word needle, which is omitted from the list. The associative memory illusion occurs when subjects who have studied the items of the inducing list falsely remember having studied the critical lure, needle, as well. As in the study by Knowlton and Squire, the amnesic subjects showed a nontrivial level of recognition memory, and they also falsely recognized the critical lures at a substantial level. The figure portrays the basic results. Interestingly, the amnesic subjects were less susceptible to the AMI than were normal controls -- which might be about the only benefit of being amnesic! Nevertheless, whether they correctly recognized studied items, or falsely recognized critical lures, most of these recognition judgments were accompanied by a recollective experience of "knowing" (i.e., feeling). By contrast, among the control subjects both true and false recognition judgments were overwhelmingly accompanied by remembering. Note that in this study, as in the preceding study by Knowlton and Squire, recognition-by-knowing (feeling) was approximately equivalent in the two groups. This is what we would expect if the feeling of familiarity was mediated by priming, which is equivalent in amnesics and normals.

c22CritShift.JPG (29438 bytes)A later study by Verfaellie and colleagues examined the contributions of both criterion shifts, as in Dorfman's experiment, and recollective experience, as in Knowlton and Squire's experiment, to amnesic recognition. In their first experiment, instructions to adopt a more liberal criterion for recognition failed to improve recognition performance in the amnesic subjects, contrary to Dorfman's prediction. However, a second experiment, using a different manipulation, succeeded. In this experiment, subjects were misinformed that either 30% or 70% of the items presented on the recognition test were old, when in fact the test items were half old and half new, following standard procedures for this kind of test. As can be seen in the figure, when the amnesic subjects shifted criterion they showed a substantial increase in hits, but not in false alarms, resulting in a genuine increase in d'. Moreover, an analysis of recollective experience showed that, across the two experiments, most (approximately 71%) of the amnesics' hits were associated with knowing (i.e., feeling), whereas most of the controls' hits (approximately 70%) were associated with remembering.

If recognition by amnesics were mediated by priming-based feelings of familiarity, we would also expect that the levels of confidence attached to these judgments would be relatively low, compared to those of normals whose recognition is mediated by retrieval or remembering. As I have already noted, studies of normal memory that make the tripartite distinction between remembering, knowing, and feeling find that recognition by feeling is attached to relatively low confidence levels, compared to recognition by remembering or by genuine knowing. And this is also the case for recognition by amnesics. 

c23ConfRat.JPG
                (24660 bytes) In the study by Haist et al., for example, after only a 15-second retention interval amnesic patients gave confidence ratings averaging about 4.0 on a five-point scale to correct judgments, compared to an average of about 5.0 for controls. As the figure indicates, it took more than a day for controls' confidence ratings to fall to the levels reported by amnesics after only 15 seconds. Similar findings were obtained by Shimamura and Squire. When based on feelings of familiarity, rather than conscious recollection of the past, recognition has more of the character of an intuitive judgment -- a judgment which is admittedly fallible, and in which the person, whether amnesic patient or forgetful normal, has relatively little confidence.


Putting it All Together

In their extremely thorough study, Yonelinas and his colleagues (2001) combined Jacoby's process dissociation procedure and the Tulving-Gardiner remember-know paradigm with confidence ratings and signal detection theory to uncover the components of recognition in amnesia. Again, as in the work discussed earlier, what Yonelinas et al. call "knowing" is really the intuitive "feeling" of familiarity. Their re-analysis of four previous studies confirmed the original conclusion of Verfaellie and Treadwell that amnesia severely impairs recognition by remembering (recollection, in Jacoby's terms), but leaves recognition by familiarity (Jacoby's fluency) relatively spared. In a new experiment, Yonelinas et al. tested a dual-process theory of recognition similar to those proposed by Mandler and Jacoby. As its name suggests, the model holds that recognition can be mediated by two processes.

c24ROC.JPG (32686
                bytes)Yonelinas' dual-process model is a formal, mathematical model that generates quantitative predictions represented by a set of "receiver operating characteristic" (ROC) curves generated according to the principles of signal-detection theory. The figure shows theoretical ROC curves of the sort generated by the model. Chance performance, with hits equaling false alarms, is represented by a straight line running along the diagonal. Performance above chance, with hits exceeding false alarms, is represented by lines above the diagonal -- the better the performance, the further the curve is from the diagonal. Note that the theoretical ROC curve for recognition by familiarity (fluency or feeling) alone is lower than that for familiarity plus recollection. This is because fluency alone is somewhat less reliable than recollection, for reasons discussed earlier. Moreover, the familiarity curve is symmetrical, while the curve for familiarity plus recollection (retrieval or remembering) is asymmetrical. This follows from the assumption that familiarity is a continuous process, while recollection is a threshold process. At the high levels of confidence represented by the left-hand side of the graph, where subjects have relatively few hits but also very few false alarms, recollection makes memory more accurate than it would be if the subject relied on familiarity alone.

The figure also shows empirical ROC curves generated from the actual performance of amnesics and controls in a recognition experiment in which the subjects studied a list of words under a levels-of-processing manipulation, and then made recognition judgments using a six-point confidence scale. Yonelinas found, as would be expected, that the curve for amnesics was closer to the diagonal than that of normal controls -- meaning simply that amnesics did not perform as well as the controls did. More important, however, the ROC curve for the amnesics was symmetrical, while that taken from the controls was asymmetrical. It is apparent from the traumatic interocular test that the match between theoretical predictions and actual performance is quite remarkable. The overall pattern of findings strongly suggests that recognition by amnesics is primarily mediated by familiarity, while recognition by controls is mediated by recollection as well.

c25CompRec.JPG
                (28083 bytes)Applying the dual-process model to actual performance data, Yonelinas was able to estimate the contributions of recollection and familiarity, to recognition. They found that amnesia impaired both processes, compared to controls. But the important point was that there was some recognition spared in the amnesics, as we find so often to be the case. And, as would be predicted from the argument of Dorfman et al, almost all of it was based on a feeling of familiarity -- a feeling of familiarity that is a product of priming. A later study added evidence supporting the conclusion that the familiarity component in the process dissociation procedure is closely related, if not identical, to the recollective experience of "knowing" -- that is, of feeling". Both, in turn, are reflected in the familiarity-based ROC curve derived from confidence ratings. Priming underlies all three aspects of recognition.

The point of all of this is that recognition by familiarity is mediated by an experience of perceptual and conceptual fluency that in turn reflects the priming effects of past experience. And it is familiarity -- feeling, not knowing, which seems to underlie successful recognition by amnesics. Heuristic reliance on the feeling of familiarity, in turn based on the perceptual and conceptual fluency that comes with priming, permits amnesic patients to recognize items that would not be recognized if they relied on conscious recollection. Because familiarity does not help much when it comes to free recall, amnesics perform better on tests of recognition tests than they do on tests of recall. Because the feeling of familiarity is based on the perceptual and conceptual fluency that comes with priming, recognition in amnesia constitutes a case where implicit memory interacts with explicit memory.


Where We Are Now

There is now abundant evidence that a priming-based feeling of familiarity can mediate recognition judgments of both normal subjects and amnesic patients, and that the recognition judgments of amnesic patients, in contrast to those of normal subjects, are mediated more by familiarity than by recognition. Nevertheless, recognition is not always improved by relaxing decision criteria. Why the difference in outcome? In the first place, it should be understood that the process proposed here is not simply one of "relaxing" decision criteria. It is not necessarily enough for subjects just to lower their standards for what they are willing to judge to be "old" rather than "new". Rather, they may need to apply a qualitatively different set of standards. It is one thing to judge that someone is familiar because you can remember the circumstances under which you initially, or most recently, encountered them. This is what Mandler means by retrieval, Jacoby by recollection, and both Tulving and Gardiner mean by remembering. It is quite another thing to make the same judgment based on the fact that the person's face or name "rings a bell". That is what Mandler means by familiarity, Jacoby by familiarity or fluency, Gardiner means by knowing, and I mean by feeling. To shift from one to the other as the standard for making the "judgment of previous occurrence", is not to raise or lower the threshold on a single continuum. Rather, it is to shift from one standard to another. Instructions to subjects must make this clear.

The difference between quantitative and qualitative shifts in recognition criteria helps resolve a paradox in the application of signal-detection theory to recognition. In both posthypnotic amnesia and ECT-induced amnesia, d' (or some cognate measure) increased as the criterion for recognition was liberalized. Although Yonelinas et al. did not calculate d' for the familiarity-plus-recollection process in the dual-process model, it is apparent that familiarity and recollection-plus-familiarity are associated with different values of d'. But signal-detection theory holds that d' is independent of B, implying that d' should remain constant despite shifts of criterion. The resolution of the paradox, if that is what it is, is that recollection and familiarity entail different continua. Items can "ring a bell" more or less loudly, and retrieval can be more or less successful (here I depart from Yonelinas, who construes recollection as an all-or-none process). If subjects base their recognition judgments on one standard or the other, presumably d' will remain constant as the criterion for recognition is loosened or tightened. But if subjects shift from one standard to the other, d' will change. The fact that d' changes with shifts in criterion constitutes evidence in support of theories like Mandler's, which hold that there is more than one basis for making recognition judgments.

But there is more to shifting criteria than simply giving subjects carefully worded instructions. Subjects have to believe that they can comply with them -- a problem known in personality and social psychology as self-efficacy. If subjects don't think they'll succeed, they may not be motivated to try, and instead respond randomly. Reber and Squire suggested as much when they noted (p. 509), "ECT patients are unaccustomed to their memory dysfunction...", while "Amnesic patients... are accustomed to their condition...." This is certainly true of college students experiencing posthypnotic amnesia as well. Perhaps Dorfman's patients, for whom amnesia is an acute problem rather than a chronic condition, were simply trying harder to remember, and were thus more motivated to capitalize on priming-based familiarity. We know that, despite their impairments in explicit episodic memory, amnesic patients can incorporate new "semantic" self-knowledge into their self-concepts, and amnesia may be one of those facts. Although it might seem paradoxical to say that amnesic patients "remember" that they can't remember, even H.M. is aware that he is amnesic. Although some amnesic patients may be unaware of their memory deficits, a condition generally called anosognosia, many patients for whom amnesia is part of their identity may just not try to remember anymore.

It may also be that, in severe cases, the amnesic syndrome itself prevents amnesic patients from strategically capitalizing on the experience of fluency that gives rise to the feeling of familiarity. This is because fluency is not absolute. It is not the case, for example, that items identified in less than 750 milliseconds are experienced as more familiar than those which require more time. Rather, fluency is relative: the fluency to which priming gives rise is only experienced in comparison to the relative difficulty of processing other items. This comparison itself requires memory -- at least the ability to remember, over a relatively short term, how easy or difficult it was to process one item when processing another item. Amnesia normally spares such primary or "short-term" memory, of course, so under ordinary circumstances amnesic patients retain the capacity to make such a comparison -- provided that they are encouraged to do so. At the same time, it is possible to imagine experimental procedures that would tax this capacity severely: blocking the presentation of primed and unprimed items at the time of the test, for example, or increasing the interval between items, or inserting a distracting task. Anything that compromises primary or "short-term" memory would impair the ability of anyone -- amnesic or not -- to appreciate the relative fluency, created by priming, that gives rise to the feeling of familiarity that can mediate recognition.

Recognition based on fluency or familiarity is sometimes referred to as automatic or unconscious, but it is important to be careful here. The priming effect that causes fluency, and in turn gives rise to the feeling of familiarity, is an unconscious influence of memory, in that the person who displays priming need not consciously remember the priming episode. And the activation and integration which causes priming may well occur quickly and automatically, in the technical sense of the term.  However, the person who experiences fluency and familiarity is conscious of something -- to wit, the very fluency and familiarity that he or she subsequently attributes to an event that occurred in the past. And this attributional judgment, far from being fast and automatic, is deliberate and time-consuming. As Bartlett argued long ago, and others have more recently reminded us, the remembering subject is engaged in deliberate effort after meaning (Bartlett's phrase) -- trying to make sense of the present in light of the past, trying to reconstruct the past in terms of the present. Such a process takes both time and effort.

As was the case with explicit and implicit expressions of memory, the two modes of recognition memory -- by retrieval (recollection) and familiarity (fluency) -- are sometimes ascribed to the operation of different brain systems. For example, Aggleton and Brown have suggested that retrieval-based recognition is mediated by the hippocampus and the anterior thalamus, while familiarity-based recognition is mediated by the perirhinal cortex of the temporal lobe and the medial dorsal nucleus of the thalamus. Similarly, Yonelinas has suggested that recognition-by-retrieval is mediated by hippocampal structures (particularly CA1, CA3, and the dentate gyrus), while recognition-by-familiarity is mediated by structures in the parahippocampal region. If priming itself is mediated by one or more other systems, Stark and Squire may be right to imply that recognition-by-retrieval, recognition-by-familiarity, and priming are mediated by three separate memory systems. But these systems cannot be strictly independent. The feeling of familiarity has to come from somewhere, and the best candidate for that "somewhere" is priming itself. Priming may well be independent of recognition, in the sense that priming occurs whether recognition occurs or not. But recognition cannot be independent of priming, because priming creates the feeling of familiarity that is one mode of recognition. Recognition is independent of priming only in those limited circumstances where, whether by virtue of brain damage (perhaps to regions surrounding the hippocampus), experimental instructions, or other factors, subjects do not (or cannot) take strategic advantage of the feeling of familiarity that comes from priming.


Independence and Interaction of Explicit and Implicit Memory

An enduring theme in the study of implicit memory is that of independence. For many neuropsychologists, explicit and implicit expressions of memory are the product of separate brain systems that somehow function independently -- a multiple-systems viewpoint which is consistent with the dominance of modularity in cognitive neuroscience as a whole. However, there are some problems attached to this dominant view.

It may well be that multiple-memory theories can surmount these challenges, and we will eventually come to agree that explicit expressions of memory, such as recall and recognition, and implicit expressions of memory, such as priming, are mediated by different brain modules or systems. Here, however, I have been concerned with the specific proposition that explicit and implicit memory do not interact. This simply does not seem to be the case.

Explicit and implicit memory may well prove to be mediated by a number of separate systems, as dogma in contemporary cognitive neuroscience insists. But at the same time, theory has to find a way for these systems to interact -- to combine and communicate with each other to support the individual's performance of the task at hand. Similar considerations apply to single-system theories, all of which assume that different processes play different roles in explicit and implicit memory. Because there are no process-pure tasks, these processes must be able to interact to support the subject's performance of each and every one of them. In the final analysis, it's a whole, sentient, intelligent, conscious person who performs the task, using all the information that is available to him or her, whether those influences arise from conscious or unconscious memory.

 

This page last modified 07/22/2014.