Chapter Six Memory & Connectionist Models Memory is central to any cognitive system. Consider what it would be like to get into your car tomorrow and realize your body no longer knew the correct motor movements for driving. As this realization hits you, you realize that you do not remember the steps it would take to start the car (e.g., the fact that to start a car, you need to put the key in the ignition switch and turn it), and then it hits you that you do not have a recollection of any events from the past where you have driven a car. It is as if you have never been in a car before. In fact, you are no longer sure the car you are now sitting in is actually yours. As this example suggests, memory is at the heart of any intelligent cognitive system. Without memory the cognitive system is stuck in the here and now; there is no repository of knowledge to guide intelligent action or make sense of the world. Without memory, every action requires figuring out the motor movements necessary to achieve said desired action. Without memory the world has no continuity, it is as if one lives only in the moment, and there is no possibility of a sense of self. Without memory how could you answer the question, “Am I the sort of person who would get into someone else’s car and try to steal it?” In addition to pointing to the centrality of memory to cognition, this example also makes it apparent that many types of information must be stored in memory. Motor skills, facts, and events from our lives, for example, are important types of information. One major question for scientists who study human memory is the extent to which evolution has given us distinct memory systems for different types of information. This chapter: • • • • • • •

Defines a memory system. Discusses the distinction between the systems and process approach to studying memory. Explores an early information-processing model of memory: The early 3-store modal memory model. Discusses how the STM conceptualization of immediate memory in the 3-store modal memory model was upgraded to the working memory (WM) conceptualization. Discusses the major long-term memory systems of the brain. Briefly presents 2 memory process distinctions as examples of the process approach to memory. Discusses connectionist network models and how these might be used in models of memory.

183

Outline of Topics I. How Do Cognitive Scientists Study Memory? II. The Systems Approach: The Landmark Example of H.M. III. The 3-store Modal Model of Memory A. Sensory Stores B. Short-term Memory (STM) C. Long-term Memory (LTM) D. Are STM and LTM Separable Memory Subsystems? IV. Immediate Memory Reconsidered as Working Memory (WM) A. Original Baddeley & Hitch (1974) model B. Summary of the WM Approach C. Future of the WM Approach V. Long-term Memory (LTM) Systems A. Declarative (Explicit) Memory B. Nondeclarative (Implicit) Memory VI. The Process Perspective A. Recollection & Familiarity VII. Connectionist Models References Glossary

I. How Do Cognitive Scientists Study Memory? A memory system has 3 important tasks to perform: Encoding, storage, and retrieval. Information from the world must be acquired and represented. Representations must be consolidated into a durable enough form to fit the desired retention period (e.g., fractions of a second, seconds, minutes, or years). Once encoded, information must then be stored and maintained so that it may be retrieved. At the time of retrieval, efficient processes for effective retrieval will be important. Moreover, it is important to understand how different retrieval problems will emphasize different sets of retrieval processes, and how that information is encoded will also have important implications for what sort of retrieval processes will be most effective. Central to the cognitive approach is the development of models of memory. One influential theory is the 3-store modal model (Neath & Surprenant, 2003) that consists of a sequential series of locations dedicated to storage of representations for increasing lengths of time as information passes from one store to the next. More recent models involve representation of information distributed across numerous neural-like computational units that both process and store information. This family of models is referred to as neural network or connectionist models. We discuss some of the

184

influential traditional memory models, and we present some of the basic concepts of connectionist models. Historically, memory researchers have chosen to adopt one of two major research strategies, one with a focus on the processes of memory and the other with a focus on memory systems of the brain. The process approach concentrates on the complex processes used during encoding and retrieval. For example, the processes of encoding and retrieval are likely to be quite different if participants are to perform a free recall task, where they are given a blank sheet of paper and must recall all of the studied items without any help, versus a recognition task, where they are presented with an item and must simply recognize it as having been part of the study set. Another way of approaching memory is to study the memory systems of the brain under the assumption that evolution has yielded multiple memory systems to deal with multiple types of information in the environment. We explore each of these perspectives as well as briefly discuss some of the major theoretical approaches to memory developed by cognitive scientists.

“Chimp vs Human – Memory Test”

II. The Systems Approach: The Landmark Example of H.M. The modern era in research on memory systems was initiated in 1953 when a 27 year-old epilepsy patient known as H.M. underwent a surgical procedure that involved bilateral (i.e., both sides) removal of the medial temporal lobes, including most of each hippocampus, the surrounding anterior medial temporal cortex, and each amygdala (see Figure 1, also consult the Cognitive Neuroscience chapter). This radical step had become necessary as his epileptic seizures had stopped responding to drug treatment, and had become life threatening in severity. While the surgery greatly improved the epileptic seizures experienced by H.M., it became almost immediately apparent during his postoperative recovery in the hospital that he was now suffering from a severe amnesia (Milner, 2005). However, even though he lost some of his ability to recall events and facts that he learned during the month prior to his surgery, he did not have the sort of retrograde or backward looking amnesia depicted in television soap operas and other popular media where a character might lose the ability to remember events or facts (e.g., their

185

name) from the period prior to the injury. By contrast, H.M.’s memory for the events of his life, for example, childhood events, or the car he drove and the house he lived in prior to the surgery, were basically intact. What he had lost was the ability to form new long-term memories. He could not recall the names or faces of hospital staff, even if they had just been talking to him and stepped out of the room for a few minutes. He had forgotten the way to the bathroom and could not relearn this information. From this point on HM was forced to live a sheltered life in the hospital, and later in an assisted care setting. He watched TV regularly, but was unable to learn all but a small collection of new facts and new words that came into use in the English language since his operation. His ability to remember life events since his operation was almost nonexistent. In other words, H.M. had a severe anterograde, or forward looking, amnesia.

“H.M.” In the years that followed, H.M. was to become the most tested neuropsychology patient in history, beginning with Brenda Milner’s groundbreaking work to bring this surprising case to the attention of the scientific world (Milner, 2005). While H.M. was not the first patient to be reported with memory problems following removal of brain tissue, his case marks the modern study of memory systems of the brain because his impairment stood in such stark contrast to his preserved function in several major cognitive areas. While his memory impairment was surprisingly severe in terms of the loss of ability to encode new events and most new facts in long-term memory, extensive testing demonstrated preserved performance on intelligence tests, as well as tests of perception and reasoning. In fact, H.M. was able to converse normally with others in his daily life. H.M. was also able to hold small amounts of information in memory for a brief period; for example, he was able to remember a 3-digit number for 15 minutes at one point. It was also the case that H.M. could remember many facts and events from his life prior to the surgery, indicating that the hippocampus and surrounding medial temporal lobe was probably not the storage system for longterm representations, but rather involved in acquisition and encoding. Finally, when tested for the ability to learn motor skills he showed a normal ability. When presented with these tasks by the same researcher at a later date, even though he could not remember having learned the task previously, or having previously met the psychologist who was now testing him again, he showed a normal shortened time to relearn the task indicating preserved motor learning from the previous session. Because of this distinct

186

pattern of deficit and preserved function, H.M. became the standard which other amnesic patients were to be compared against. To learn more about H.M., and to hear researchers who interviewed him, go the National Public Radio website and click on the “Listen” button. As H.M.’s own words attest, he was sometimes concerned about the feeling he had that he was simply living in the moment, but then, his mental world would move on and he would process a new set of events: Right now, I’m wondering. Have I done or said anything amiss? You see, at this moment everything looks clear to me, but what happened just before? That’s what worries me. It’s like waking from a dream; I just don’t remember. (Milner, 1966)

Figure 1. Horizontal T2-weighted MRI section from H. M. The bright signal areas indicate the extent of the anterior medial temporal lobe removal.

III. The 3-store Modal Memory Model One contribution of the cognitive approach to memory is the proposal of distinct memory systems that operate at different levels of time scale. Figure 2 presents a combined information processing view of memory that sums up research from the ‘60s and ‘70s and clearly distinguishes 3 subsystems of memory that operate at increasingly long time scales (for more information consult Healy & McNamara, 1996). The term modal is used to indicate that this model is representative of a number of theories of the time. There is an emphasis on the use of verbal items (i.e., words, letters, digits) and memory is viewed somewhat passively as a location to store items. Sensory input to the

187

cognitive system is briefly held in a sensory form (i.e., no perceptual identification or semantic meaning analysis). Information that is attended is moved into short-term memory (STM), which can be thought of as containing the contents of current conscious awareness. Finally, portions of STM may be moved to more durable storage in longterm memory (LTM).

A. Sensory Stores. Sperling (1960) studied the operation of the visual sensory store. He had participants view briefly flashed arrays of letters (0.05 s) and used a partial report procedure where participants listened to a tone to tell them which part of the letter array to report. This partial report procedure was necessary because when participants would attempt to report all the letters in the array, they would complain that they had the sense that they perceived all of the letters momentarily but that the letters would fade from awareness very quickly. Figure 3 presents the results of this groundbreaking study, indicating a high percent of letters in the display were available when the letter array was presented at the same time as the tone. However, as the delay between the letter array and the tone is increased, the percent of items available for report decays to the point that it is nearly the same as when participants are asked to report all items in the letter array. Successive research studies have confirmed this finding, and have demonstrated that the rate of decay of information in the visual sensory store depends on viewing conditions (e.g., size and font of letters, spacing of letters, contrast of letters with background of display), but the primary finding is the visual sensory store holds visual information available for further processing for fractions of a second to a couple of seconds. Moreover, the representations in visual sensory memory appear to be coded in a pre-categorical sensory form. For example, participants can perform the partial report task easily, if they select which letters in the array to report based on visual aspects of the display such as location (e.g. row), or color (e.g., just the red items). However, performance is greatly impaired if they are asked to partially report based on category (e.g., just the letters, or just the digits). Other researchers have provided evidence for other sensory stores (e.g., auditory, touch). The visual and auditory sensory registers are by far the most extensively studied.

B. Short-term Memory (STM). STM is a limited capacity system for holding information in an active state on the order of seconds. Most theorists during the modal model’s formative period held the view that items attended to both entered consciousness and were selected for encoding

188

in STM (Smith & Kosslyn, 2007). A time-dependent decay of that activity level of representation was thought to lead to forgetting unless active rehearsal of items in STM was employed to keep representations from decaying before they could be moved to a more stable representational form in a longer-term store. The prototypical real-world task that motivates understanding of the usefulness of STM is the example of a friend telling you to call a certain phone number. Most people faced with this task will find themselves mentally rehearsing the number as they pull out their cell phone, power it up, and dial. STM is proposed as an immediate memory system for holding information briefly for a few seconds for immediate use, or alternatively, as a system for extending the contents of consciousness. Most of the research used verbal memory items, and so it is no surprise that STM was thought of as representing information in a verbal form. Other representations (such as visual and semantic) were discussed from time to time, but STM was mostly discussed as a system using verbal representations. Moreover, STM was seen as a unitary system even in those cases where multiple types of representation were proposed. For example, much was made of the finding that STM memory tasks would yield far more rhyme errors (confusions based on how a word sounds) than synonym errors (confusions based on the meaning of a word).

STM is viewed as a limited capacity system that can reliably hold a certain number of items, with capacity varying for different types of items and for different individuals (Neath & Surprenant, 2003). To save space, items can be combined into functional chunks, depending upon personal experience. For example, you might combine the ordered list of 3 letters, F-B-I, into the single informational chunk FBI (Federal Bureau of Investigation) based on your personal experience as an American, whereas someone from another country might not do so and would have to deal with the list as 3 separate verbal items. The way that psychologists determine the capacity of STM is by testing

189

with an STM span test where items are presented one-by-one and immediately repeated back in order. The most commonly used STM span test is the digit span test. A varying number of randomly ordered digits are auditorily presented, and the person being tested immediately repeats the digits back in order. What tends to happen is participants will perform at a very high level, virtually error free, until the number of digits presented crosses some threshold and they will suddenly produce frequent errors. This threshold number is the digit span of the individual, or the number of random digits that can be reliably repeated without error. A variety of STM span tests have been devised using different materials (e.g., spoken words, visual words, visual shapes, spoken digits repeated backwards), with the same basic testing procedure. The normal range for single syllable items that do not encourage chunking (e.g., digit span in English), is typically about 7±2 items, but the increased difficulty of non-verbal or multi-syllable materials typically yield lower spans.

C. Long-term Memory (LTM). By contrast, LTM is a large capacity system for holding information in a more stable, but less active state, than STM. STM is proposed as the gateway to LTM (Neath & Surprenant, 2003). Items in STM are encoded in LTM using 2 types of rehearsal processes, rote rehearsal (repetition) and elaborative rehearsal where the meaning of the item is associated with meaningful information already in LTM. The classical research on LTM focused on semantic representations for verbal materials. Studies of LTM using word lists find less sensitivity to rhyme confusions than synonym confusions, the opposite of STM with the same materials. This suggested to researchers of the time that, at least for verbal items, we tend to represent information according to the meaningfulness of the items presented. We say the capacity of LTM is large, but we really do not know how large, and it will be extremely difficult to find out. We do know that information can be forgotten, but this may be due to a problem with retrieval processes, or the information may not have been durably encoded in the first place, rather than any capacity limitation. It is important to note that psychologists have developed span tasks to measure the capacity of STM, but no widely accepted tests exist for the capacity of LTM.

D. Are STM and LTM Separable Subsystems? The standard story, based on many early studies that at first appeared to yield clearly interpretable results, was that STM and LTM form what is known as a double-

190

dissociation pattern indicating separable systems for STM and LTM. One important set of evidence comes from studies of the serial position curve during free recall of lists of randomly ordered words. In this procedure, relatively long list of words (in the 15-25 range is common) are verbally presented at a set rate, and upon completion, participants are signaled to either recall as many words as they can immediately, or after a brief period of a simple distraction task (e.g. counting backwards from 100 by 3’s). When recall for each word as a function of the word’s position in the list (e.g., percent recall for the 1st versus the 10th versus the 20th word in the list). Figure 4 presents stylized results that summarize typical results.

“STM and LTM” The standard serial position curve has a region of increased recall for the first few items of the list, called the primacy effect, which is thought to reflect improved encoding of the first few words into LTM. The standard serial position curve also has a region of improved recall at the end of the list, called the recency effect (these are the items most recently presented), which is thought to reflect direct retrieval from the words active in STM. Manipulation of the rate of presentation of the words in the list usually affects the primacy portion (LTM) of the serial position curve, leaving the recency portion (STM) of the curve relatively unaffected. The opposite holds for the addition of a distracting verbal task where the recency portion (STM) of the curve is typically far more affected than the primacy portion (LTM). The fact that both experimental manipulations primarily affect one portion of the curve leaving the other relatively unaffected is a sort of doubledissociation pattern. A more traditional double-dissociation involving neurological patients with mirror image patterns of memory deficits has also been reported. Patient H.M., with an impaired ability to encode new events and most new facts into LTM due to medial temporal lobe surgery, shows a digit span in the normal range (i.e., normal STM) but greatly impaired LTM for word lists. Converse patients, E.E. and K.F., have been reported with inferior parietal and superior temporal lobe damage who produce LTM test scores in the normal range compared to control subjects, but greatly impaired digit and word spans (STM). These and similar results reported in the literature made the STM/LTM distinction a strong one for many early researchers, but newer research (Neath & Surprenant, 2003) and a new wave of theoretical models (Cowan, 1995) are calling the strict separation of STM and LTM into question. One promising proposal (Ranganath & Blumenfeld, 2005) is that the same neural networks that hold information active are involved in storage in LTM. If this proposal is correct, then STM and LTM are more connected than previously believed. In the future, this will most likely be an increasingly active area of research.

191

Sensory Input Sensory Stores Visual Auditory Touch ...

Attention

Short-Term Memory (STM)

Long-Term Memory (LTM) Figure 2. 3-store modal model of memory. Sensory stores briefly hold sensory information for further processing. Information attended to is moved to a limited capacity short-term memory (STM) and can remain available as long as it is maintained in an active state. Some of the information in STM may be moved to a more durable storage system, long-term memory (LTM).

192

Percent Reported Letters

100 80 60 40 20 0 -0.2

0

0.2 0.4 0.6 0.8

1

1.2

Delay of tone (seconds) Figure 3. Percent of letters reported from a letter array using a partial report procedure. A tone indicates which row in the letter array to report. The tone is offset in time, falling either slightly before, simultaneous with, or delayed from the onset of the letter array on the screen. The height of the bar on the right indicates the percent report when participants must report all letters (i.e., whole report instead of partial report). Adapted from Sperling (1960). [Sperling, G. (1960). The information available in brief visual presentations. Psychological Monographs: General and Applied, 74, 1-28.

193

Percent Recall

100

Standard Presentation Distraction Faster Presentation

Primacy-LTM

Recency-STM

0 1

20

Word Position in List Figure 4. Stylized results of studies of the serial position curve, free recall as a function of word position in a list of items presented one-byone. In the standard presentation condition, there is greater recall for the first few words presented, called the recency effect, this is thought to be due to greater encoding into LTM in relation to words in the middle of the list. Also, there is greater recall for the last few items, presumably due to direct retrieval of active words in STM. Manipulation of presentation rate and distraction (e.g., counting out loud) between the last word and commencement of recall selectively reduce the primacy (LTM) and recency (STM) portions of the curve, respectively. This suggests separable STM and LTM systems. Serial position curve online demonstration. Dr. Timothy Bender at Missouri State University has posted an online demonstration of the serial position curve, along with manipulations intended to reduce the primacy and recency effects (see Figure 4). The reader is encouraged to go to the Web page and click on and run the Serial Position Curve demo.

IV. Immediate Memory Reconsidered as Working Memory (WM) Besides extending the contents of consciousness, what do we need from our immediate memory system? One major purpose of such a system might be to support complex cognitive processing involved in such tasks as reading, reasoning, problem solving, and decision making. Consider the following example. You have just finished a meal at a restaurant and you are presented with the bill. You want to leave a 15% tip, and so you close your eyes and compute the tip in your head. You begin by figuring

194

10% of the bill and temporarily storing that amount. Then, you halve the 10% subtotal to get 5% of the original bill. Finally, you add the current 5% result to the 10% value you had been holding in memory to get the 15% tip. This example highlights immediate memory as having an active role in providing a mental workspace in support of ongoing cognitive processes. We can contrast this example with the example we used to motivate STM, the task of holding a phone number active in immediate memory until we are ready to dial the number on our cell phone. We now see that STM proposes an immediate memory system that is a location for storing information for a brief period until the information is to be used. STM does not stress the need to hold information while performing concurrent processing is performed, and is therefore thought of as a passive memory system. An alternative approach, called the working memory approach, instead views the immediate memory system as providing storage concurrent with ongoing active cognitive processing. Over the years, researchers have identified 2 more important weaknesses of the STM model of immediate memory (Smith & Kosslyn, 2007). One problem is that the 3store modal model is a serial processing model, and STM is proposed as the gateway to LTM. As we discussed earlier during the discussion of the double-dissociation of STM and LTM systems, patients have been identified who show quite impaired digit and word spans on traditional STM span tests, while showing normal performance on tests of LTM. It is difficult to see how this could be the case if STM is the gateway to LTM. Another problem with the STM approach is its reliance on verbal codes. While alternate representation schemes for STM have been proposed, such proposals were relatively rare and failed to propose separable subsystems for different representational codes. Evidence from studies of dual-task performance has accumulated over the years supporting, for example, separable immediate memory systems for verbal and visual information. Logie et al. (1990, as reported in Baddeley et al., 2009) provide an influential example. The logic behind the dual-task methodology is that if two tasks use the same immediate memory system, then performance on one of the tasks should be interfered with when the other task is also performed. Logie et al. had participants perform either a verbal (spoken consonant) span task or a visual (checkerboard pattern location) span task. While performing one of

195

these span tasks to determine the capacity of immediate memory for verbal or visual information, participants either performed a second visual or verbal task. Figure 5 presents the visual and verbal span scores when these span tasks were performed concurrently with a second interfering task. Scores are presented as a percent of the score they earned on these span tasks when performed in isolation. As can be seen, immediate memory spans were high (about 80% of the span for each span task in isolation) when performed with a second interfering task that was from the opposite modality (e.g., visual span with verbal secondary). However, memory spans were much lower (30-40%) when the span tests were performed in conjunction with a secondary task from the same modality (e.g., visual span with visual secondary). This pattern of results has been observed in a range of studies and has been taken as strong evidence for separable verbal and visual immediate memory subsystems.

A. Original Baddeley & Hitch (1974) WM Model Because of evidence such as that presented in Figure 5 suggesting a need for separate visual and verbal subsystems, and because they believed that the true purpose of the immediate memory system was to provide a mental workspace in support of complex cognition, Baddeley and colleagues (Baddeley et al., 2009; Baddeley & Hitch, 1974) have proposed a modified theory called working memory. They envisioned a working memory system with independent modality-specific modules for holding verbal and visual information in an active state: phonological loop and visuospatial sketchpad (see Figure 6). A central executive subsystem provides control and coordination functions. Phonological loop. The phonological loop is made up of 2 parts. The phonological store can hold memory representations (in a phonological form that represents what a word sounds like) in an active state for a few seconds before they fade, and the articulatory rehearsal process that is analogous to the silent inner speech that most people experience when they read. To be refreshed, phonological representations are retrieved from the phonological store and re-articulated using silent inner speech. The working memory model predicts, correctly, that items that take longer to speak aloud (overt articulation) are more prone to being lost from the phonological loop (Baddeley, 2003). The phonological loop can be seen as a more detailed and flexible version of the verbal STM proposed in the modal model. Visuospatial sketchpad. Imagine that you are playing a game of checkers with your young niece who suddenly knocks the board over. In attempting to recreate the game, you close your eyes and imagine the game board with the pieces arranged as they were just before being upended. The ability to create and manipulate a visual image is thought to be a central function of the visuospatial sketchpad. Baddeley and colleagues

196

(Baddeley et al., 2009; see also Neath & Surprenant, 2003) have proposed that the visuospatial sketchpad has a refreshing mechanism (called the visual scribe). Recent evidence suggests that the same neural systems that allow us to move the focus of visual attention around the external visual world allow us to covertly move our mental focus of attention around the imagined visual world, acting to refresh locations and shapes in the visuospatial sketchpad (Smith & Kosslyn, 2007). There is also accumulating evidence for 2 separable working memory systems, one for visual shape information, and the other for spatial location information. For example, Kohler et al. (1995) had participants view displays with pictures of common everyday objects in particular positions in the visual field. After a brief delay, a second display appeared which might contain the same objects in the same locations, or some objects might be shifted in location, or the locations may be as before, but some of the object shapes are changed. Participants had to either detect changes in location or changes in object identities while a functional PET scan was conducted. The results indicated that shape change detection following a brief delay depended on temporal lobe areas responsible for object identification, and that object location change detection following a brief delay depended on parietal areas responsible for keeping track of the location of objects. Whether these subsystems for shape and spatial location use the same neural areas for refreshing visual and spatial information is a matter of much current debate. Central executive. The central executive can be thought of as a cognitive controller that (a) selects which modality-specific subsystem is used for storage, (b) controls the timing of input to each subsystem, (c) integrates and coordinates between subsystems, and (d) provides the cognitive control mechanisms for access to the information in the subsystems for use in complex cognitive processing (Smith & Kosslyn, 2007).

“Central Executive” Working memory span tests. Just as in the STM model it was designed to supplant, the WM model is comprised of limited capacity subsystems. There are 2 major differences, however, as the WM model allows for the phonological loop and visuospatial sketchpad to work somewhat independently, with only partially interfering processes (see Figure 5). As discussed earlier, researchers using dual-task methodology have found that span (capacity) of each subsystem is only modestly affected by an interference task from the opposite modality (e.g., visual secondary task and verbal span test). A second major difference with the STM model is

197

the WM attempts to explain immediate memory performance in the presence of concurrent cognitive processing. It is therefore a major requirement of WM span tasks, as opposed to STM span tasks, that the newer WM span task require storage for brief intervals during concurrent performance of a processing task. Figure 7 presents a typical WM span task. Participants must hold a growing list of words in WM as they are presented with a concurrent processing task of verifying the correctness of simple arithmetic problems. Their WM span is the number of words they can reliably recall in order without error. By contrast, the common digit span (STM span) task only requires storing digits for immediate recall, there is no concurrent processing load. It should be of little surprise that WM span tasks are more difficult, and the normal range of spans on a typical WM span task is less than that for the typical STM span task. Also, in keeping with the fact that the WM model is designed to provide a more realistic explanation of how we use our immediate memory abilities to provide temporary storage to support complex cognitive processing tasks (e.g., reasoning, decision making, reading) a common finding has been that scores on tests of complex cognition (e.g., reading comprehension) are typically more strongly related to WM spans than they are to STM spans. That is differences in scores for 2 individuals on a difficult reading comprehension test are better predicted by their differences on WM span tasks (e.g., Figure 7) than by their scores on a STM span task (e.g., digit span). This is presumably because the WM span tasks better capture how WM is used when we actually use it during a complex cognitive task such as reading. STM and WM span test demonstrations online. Dr. Timothy Bender at Missouri State University has posted versions of the digit span and operations span tests commonly used to assess immediate memory capacity from an STM (digit span) or a WM (operations span) perspective. The reader is encouraged to go to the Web page with these demos and to click on and run the digit span and operations span demos.

B. Summary of the Working Memory Approach. The working memory model proposed by Baddeley and colleagues (Baddeley, 2003; Baddeley et al., 2009) handles the 3 shortcomings of the STM model discussed earlier. The great strength of the WM approach is that it better captures our need to use our immediate memory system to support ongoing complex cognition. WM is a fundamentally active system that must store information in the presence of concurrent processing, in comparison to the more passive STM view that only requires passive storage of information for a brief interval until it is needed. The WM approach also provides for a wider variety of codes (e.g., visual, spatial, verbal) that are implemented in modality-specific subsystems that explain the results of dual-task studies indicating minimal interference across code modalities (e.g., visual vs. verbal). Moreover, the WM model does not propose a strict serial order of processing as does the STM model of the 3-store modal memory model. In the modal model, STM is seen as the gateway to memory, but the discovery of patients with neural damage leading to drastically impaired STM span (e.g., trouble repeating back more than 2 digits in the digit span task), but relatively preserved ability to encode information into LTM have called into question the assumption of a short-term store that acts as a gateway to LTM. The WM

198

approach leads to a more flexible approach to interactions between the immediate memory system and LTM than does the older STM approach.

C. The Future of the Working Memory Approach. Baddeley (see Baddeley, 2003, for a readable review) has recently made some serious upgrades to his WM model, which, in its original form, was published in Baddeley and Hitch (1974). Recent neuropsychological evidence from patients with neural deficits, and from new brain imaging research on typical individuals, has led to an accumulation of evidence for a working memory space for conscious awareness (Dehaene & Naccache, 2001). For example, patients have been identified who have a severe amnesia that keeps them from encoding new facts and events into LTM, but with an ability to immediately repeat extended sections of text back that far exceed the phonological loop or visuospatial sketchpad. Another weakness has been that the original version of the WM model did not contain explicit mechanisms for chunking (e.g., one person might combine F-B-I into a single integrated representation FBI to save space in a limited capacity WM system, whereas another person might represent this information as a sequence of 3 distinct representations). Chunking in immediate memory is well-documented and an important source of individual differences in memory performance. Accordingly, Baddeley (see Baddeley, 2003) has recently proposed the addition of a module, the episodic buffer that acts as a mental workspace for conscious awareness and uses complex integrated multimodal representations that can support chunking processes. As depicted in Figure 7, the expanded WM model is more explicit about the ways that the WM subsystems interact with LTM. Note the bidirectional arrows between sections of LTM and the phonological loop, episodic buffer, and visuospatial sketchpad, denote strong bidirectional communication between WM and LTM. Baddeley (2003) notes that one important pathway to the development of the working memory approach to memory is to explore emotional and motivational control over the processing goals of WM. Moreover, researchers interested in identifying the networks of brain regions that support cognitive control processes have recently proposed that lateral frontal brain areas long thought to be involved in the active maintenance of representations in WM might also be involved in representing the current processing goals of WM (e.g., Braver et al., 2002). This new focus on the factors that drive, form, and represent the goals of the immediate memory system is a promising area for research activity in the near future.

“Working Memory Approach”

199

Percent Single Task Span

100

75

50

25

Verbal Span Visual Span

0 Verbal

Visual

Secondary Interference Task Figure 5. Performance on visual and verbal span tasks while performing a secondary visual or verbal interference task. Performance presented as a percent of span task performance when performed in isolation. Greatest interference when the secondary interference task is in same modality (e.g., visual span with visual secondary). Supports proposal of a working memory system for immediate memory that has separable subsystems for visual and verbal storage. Adapted from Logie et al. (1990). [Logie, R.H., Zucco, G.M., & Baddelely, A.D. (1990). Interference with visual short-term memory. Acta Psychologica, 75, 55-74.]

200

Phonological Loop

Central Executive

Visuospatial Sketchpad

Figure 6. Working memory, as proposed by Baddeley & Hitch (1974). Is (2 × 6) + 3 = 14? Is (8 × 3) - 9 = 15? Is (3 × 4) + 7 = 19? Is (5 × 3) - 2 = 16?

Table Car Tree Rock

Figure 7. Example working memory span task. Participants must process simple arithmetic problems while holding a list of words in memory. Each line of the test is presented in isolation, and the participant makes a yes/no response, before moving to the next item on a new page (or computer screen). The list of to-be-remembered words grows as each arithmetic problem is verified. The test ends with immediate ordered recall of the words.

201

Figure 8. The expanded WM model. Addition of an episodic buffer that supports complex integrated multimodal representations allows for chunking processes. The episodic buffer is proposed to act as a mental workspace for conscious awareness. Also, the nature of WM-LTM interactions have been made more specific. From Baddeley (2003), Figure 5.

V. Long-term Memory (LTM) Systems Researchers who study the memory systems of the brain tend to take an evolutionary perspective. On this view, the brain has evolved memory systems to deal with different types of information. Consider the question of what you had for breakfast this morning. That deep fried pressed potato patty may have been particularly flavorful and recalling the event of biting into it, how it tasted and smelled, is making you hungry. Consider the factual knowledge you have of what that deep-fried potato patty from this well-known fast food restaurant is called. You can easily recall the name of the patty, but you have absolutely no recollection of when or where you first learned this factoid. It is as if the

202

representation of the factual knowledge has somehow been separated from the events you experienced when you first learned this information. Finally, consider teaching your young niece to tie her shoes. At first you try to explain it verbally, but you get confused, and suddenly you realize that words cannot capture your true knowledge of the procedures used to tie a shoe. You choose instead to demonstrate for her, first by tying one of the shoes yourself, and then working on copying the sequence of movements you have made by holding her hands and guiding her movements. It is as if you do not have full conscious access to the representations of the motor movements required to tie a shoe. As these examples suggest, memory researchers have identified memory systems in the brain for encoding, storing, and retrieving major types of information, most notably, events, facts, and procedures. Figure 9 presents the taxonomy of long-term memory systems in the brain. Two important types of learned information, classical conditioning and non-associative learning, have been excluded to improve clarity and to concentrate on those systems of greatest interest to cognitive scientists. The LTM taxonomy proposes 2 major subdivisions in LTM, declarative and nondeclarative memory (Smith & Kosslyn, 2007; Squire, 2004).

A. Declarative (Explicit) Memory Declarative memory refers to those memory systems that store facts and knowledge about events in our lives in a form that is explicitly available for conscious retrieval, hence the alternative term, explicit memory. By contrast, nondeclarative (implicit) memories are stored in such a way as to not be directly accessible to consciousness via explicit volitional retrieval processes. Rather, such memories are only indirectly available. The fact that we have such memories is apparent in how the knowledge representations influence our behaviors (e.g., being able to tie a shoe without conscious access directly to the representation of this knowledge). It is as if nondeclarative (implicit) knowledge is coded in a way that does not lend itself to verbal, visual, or semantic (or other) codes used by the conscious mind. Patients, such as H.M. who was discussed in an earlier subsection of the present chapter, with damage to the medial temporal lobes, including the hippocampi, will typically present with an amnesic syndrome that is marked by a decreased ability to encode new facts and events into LTM. In the case of H.M., there was a near total loss of the ability to recall meeting and having a conversation with people, and a major loss in the ability to learn new facts (e.g., the location of the bathroom) following

203

his surgery. However, he could learn motor movement skills at a normal rate (e.g., using a pointer to follow a moving object). Moreover, even though he would deny having met an experimenter who was making a return visit, and he would deny having learned the task previously, when tested on a previously learned motor movement task he would learn the task faster a second time, just as a nonamnesic individual would. H.M. could not remember the objects, places, or people he experienced and would rapidly forget having a conversation with a person once they left the room and he shifted his attention to something else. It was as if he was living totally in the moment. This form of amnesia is known as anterograde amnesia, meaning a forward-looking amnesia for learning new information. In fact, H.M. also showed preservation of another type of nondeclarative memory effect, priming (see Figure 9, see also nondeclarative subsection below). Eventually, a picture emerged where patients with damage to the hippocampus and associated medial temporal lobe cortex exhibit impaired ability to encode new facts and events, but have a relatively preserved ability to learn and relearn motor skills, and to exhibit priming effects (see section on nondeclarative memory for an explanation of priming). This has been taken as strong evidence of separable declarative (explicit) and nondeclarative (implicit) memory systems. The declarative system employs the hippocampus and medial temporal lobe as the gateway to memory representations that are explicitly available to volitional retrieval into conscious awareness. Nondeclarative representations, by contrast, are only indirectly available. We observe the influence of nondeclarative learning as an implicit influence on behaviors (e.g., a child learns to tie her shoes or ride a bike). Episodic (event) memory. Declarative memory can be further fractionated into an episodic system for representing the events of life in an integrated multiomodal (i.e., visual, auditory, verbal, semantic, touch, smell) code. This leads to the question of where episodic representations are actually stored. Even though we have strong evidence that the hippocampus and

204

associated medial temporal lobe is the gateway to episodic memory, it is now believed that actual storage is distributed throughout the neocortex. We can better understand this principle by considering how encoding and retrieval might be performed. As depicted in Figure 10, encoding of an event commences with a pattern of cortical activity during perception of the event. Visual areas process visual aspects of the event, auditory areas process auditory aspects of the event, and so on. The pattern of cortical activation leads to a pattern of activation in the hippocampus that is then bound together into an episode. Understanding how this binding of distributed representations is thought to occur in the hippocampus is referred to as the binding problem. The binding problem is an important focus of memory research. From this perspective, retrieval is seen as a process of reactivation of the distributed bound representations in the cortex of the various aspects of the event (e.g., visual, auditory, etc.). Retrieval begins with volitional free recall processes, or with an object in the environment called a memory cue that acts as a trigger (e.g., seeing an old friend may trigger retrieval of a past event involving that friend). The bound episode representation in the hippocampus is activated and acts as a key to unlock or reactivate the distributed representation of aspects of the event in the cortex. Many researchers now believe that with time, and with repeated retrievals, the representations in the cortex are somehow bound together in a more direct fashion such that reactivation of the cortical representations is no longer dependent upon the hippocampus. This process of forming a more durable bound representation in the cortex is referred to as memory consolidation.

“Binding Problem” Recent functional brain imaging research has confirmed the distributed nature of storage of episodic information. Wheeler et al. (2000) used functional MRI (fMRI) technology to scan for areas active during episodic retrieval. Participants learned a set of pictures and sounds. On the 3rd day of the study, they were scanned during both perceptual processing of the pictures and sounds and during episodic retrieval. The results are depicted in Figure 11, and indicate that perceptual processing of pictures activated widespread visual areas and perceptual processing of sounds activated widespread auditory areas. What is of most importance was the fact that the activations during retrieval of both visual and auditory items results in differential activation of a subset of the same areas that had been activated during perceptual processing of those

205

same items. These results suggest that episodic representations are stored in a distributed manner in the cortical areas that are activated during perception. Current research in functional brain imaging and memory is focusing on the specific role of the hippocampi and associated medial temporal areas during encoding and retrieval. Semantic (fact) memory. Much of what we have discussed above regarding episodic memory is applicable to semantic memory. Semantic and episodic memory systems are not as easily separable as are declarative (explicit) and nondeclarative (implicit) memory. For example, remembering events from a childhood birthday may also require use of meaningful concepts such as cake, balloons, and clowns. Evidence from amnesic patients points to a semantic system that depends on binding in the hippocampus and associated medial temporal lobe areas just as the episodic system. What seems unique about semantic memory is the apparent dissociation of factual knowledge from the events surrounding learning those facts. Moreover, semantic memory is clearly organized by meaning. As opposed to a dictionary, which stores items organized in alphabetical order, we store conceptual items organized by meaning, with meaningfully related items being more directly connected. Before examining an important model of semantic memory, we must first consider the contents of this system. A common approach has been to concentrate on concepts and propositions (concepts and propositions have been covered in the Knowledge Representation chapter) as representation types. A concept can be thought of as a representation of some thing, event, or idea that includes all the information necessary to allow categorization. For example, the concept dog would include a representation of the knowledge that drives our ability to categorize animals as dogs versus non-dogs. A proposition is a statement regarding concepts that has a certain truth-value (e.g., my dog does not bite). It is worth noting that the vast majority of research on concepts has concentrated on concrete object categories (e.g., dog, chair), and far less research has examined abstract concepts (e.g., war, love), and we shall accordingly use concrete concepts to motivate our discussion.

206

An important class of semantic memory models (e.g., Collins & Loftus, 1975), the spreading activation semantic network, concentrates on how concepts are related in semantic memory (see Figure 12). In this class of models, concepts are represented by nodes and relationships between concepts by connections in a graphical network. Concepts can be semantically related (e.g., dogs and cats are both mammals), or they can be associatively related (e.g., salt-pepper) by virtue of co-occurrence in the environment. Concepts can be more closely or more distantly related. Processing a concept, for example, seeing a picture of a dog, leads to that concept being activated. A common assumption is that activation of a concept above some threshold level brings it into conscious awareness. Moreover, as a concept becomes increasingly activated, activation can automatically spread to related concepts along the connections between the conceptual nodes in the network. This partial activation of related concepts, typically below the threshold for conscious awareness, may help prepare us for meaningful information that may be needed in the immediate future. For example, a friend may be talking about a visit to the doctor, and then jump to a statement about something a nurse did. Having activation spread automatically from doctor, when that concept is processed, to the related concept nurse, may help you process the word nurse more efficiently, supporting language comprehension. This type of effect, where prior processing of a related concept facilitates (what psychologists call semantic priming) processing of a current concept is quite common. Literally hundreds of studies have demonstrated that semantically and associatively related words will prime each other, an experimental effect referred to as semantic priming. For example, if you are given the task of naming words one at a time as they flash on a computer screen, you are quite likely to name the word nurse more quickly following the related word doctor, than some unrelated word (e.g., rock). A spreading activation network provides a straightforward explanation for the fact that semantic priming effects are so common. We begin by assuming the representation for a word contains conceptual knowledge regarding the meaning of the word, in addition to spelling and pronunciation information. When the first word is presented and identified, the underlying concept node is activated above threshold and the word is named. Spreading activation also automatically spreads to related concepts. If the next word is associatively or semantically related, there is a good chance that the concept node was partially activated and this partial activation facilitates (or primes the pump of word recognition) naming of the second word. Spreading activation network models also explain the false recognition effect. The false recognition effect is best considered by a concrete example. False recall and recognition demonstration. Dr. Timothy Bender at Missouri State University has posted a version of the false memory task explained below. Before reading further, the reader is encouraged to go to the Web page with the demo, click on and run the False Recall and False Recognition demo.

207

Here is a list of study items: sugar, sour, bitter, candy, tooth, taste, nice, chocolate, cake, eat pie, honey, soda. Now, without looking back at the study list, was the word taste candy on the list? How about the word sweet? Researchers (e.g., Roediger & McDermott, 1995) have found that participants will often incorrectly recognize strongly related items that were not on the study list (e.g., sweet) at about the same rate as they will correctly recognize items that were on the list (e.g., candy). This strong false recognition effect is predicted by spreading activation network models. The idea is the strongly associated item not on the study list (e.g., sweet) is partially activated during study of the related items on the list and when this item is later presented during the recognition memory test, the partial activation of the concept node for that word is mistaken for a recent presentation of the item. Spreading activation network models are limited in how they can represent complex factual knowledge of the type that can be represented in a verbal sentence. Some models allow for different types of links between concepts. For example, the concepts dog and animal might be connected with an IS-A link to represent the fact that a dog is an animal. Or we might use a HAS link to represent the fact that a dog has fur. However, such modeling schemes have difficulty representing a complex pattern of factual knowledge. To deal with this problem, researchers use a type of network called a propositional semantic network that combines properties of networks and propositions (see the Knowledge Representation chapter for more on propositions). A proposition is the smallest factual statement that has a truth-value. The concepts dog and animal do not have an associated truth-value, they are not true or false in and of themselves, but the propositional statement a dog is an animal can be seen to be either true or false.

“Propositional Semantic Network” In a propositional semantic network, a proposition is represented as a node that is linked to concepts. Perhaps the simplest proposition is one that contains an agent that performs some action, a relation that defines the action, and an object to which the action is directed. Consider the proposition Lassie is a dog, depicted in Figure 13a. The proposition is represented as a node with the agent Lassie, the object dog, and the

208

relationship IS-A. The true value of the propositional semantic network is its ability to handle complex inter-relationships between propositions. Figure 13b presents a sample network representing a common plot from the classic television show that followed the adventures of Lassie the dog and her owner Timmy. The network connects inter-related propositions that point to the same concept. For example, in Figure 13b, the propositions Lassie finds Timmy and Lassie runs home are connected by both sharing Lassie as the agent. The network as a whole represents the plotline where Timmy falls into a mineshaft and Lassie finds him and runs home to “tell” his parents resulting in Lassie becoming a hero. With the addition of some additional forms for the proposition nodes, a propositional semantic network can be constructed to capture the meaning of just about any ideas that can be expressed verbally as a sentence. Exactly how conceptual knowledge is stored in the brain is controversial. Patients with damage to lateral prefrontal cortex often have trouble retrieving words and semantic information, suggesting a general role for this region in retrieval from semantic memory (Martin & Chao, 2001). Patients with damage to the temporal lobes often have trouble with object identification and categorization, and in answering questions about properties and categories of objects, suggesting that this region is important in representing object-specific information. Beyond these generalities lie many different theoretical proposals and conflicting empirical evidence. However, a recent string of functional brain imaging studies can explain a lot of the evidence. What is emerging is evidence supporting the view that a wide range of cortical systems involved in modality-specific representation of objects and events during perception are reactivated when we think about categories. For example, Chao and Martin (2000) had participants view pictures of tools, places, animals, and faces and found that tools differentially activated motor planning areas associated with grasping, and also motor-visual integration areas (see Figure 14). Subsequent studies (Simmons et al., 2005) have also found that tools also activate visual shape processing areas (ventral occipitotemporal cortex), and motion processing areas that represent the resultant motion when the tool is used (middle temporal gyrus). Simmons et al. (2005) compared viewing of pictures of food to pictures of locations and found food-selective activation in gustatory processing areas (right insula/operculum and the left orbitofrontal cortex), as well as visual shape processing areas. Taken together, these, and many other recent, findings suggest that conceptual knowledge is widely distributed and retrieval involves reactivation of many modality-specific areas responsible for perception of objects and events. From an evolutionary perspective, it may be that humans have recruited sensorimotor processing areas for representation of conceptual knowledge, grounding such knowledge in sensorimotor experience with the world.

209

Figure 9. Long-term memory systems and associated neural areas. See text for explanation. Nondeclarative memory systems associated with classical conditioning and sensory habituation are excluded as they are out of the scope of the present chapter.

210

(a) Encoding

(b) Retrieval Hippocampus

Hippocampus

Hippocampus Binding Figure 10. The role of the hippocampus as the gateway to episodic memory. (a) During encoding a pattern of cortical areas are activated during perceptual processing of an event, this results in a pattern of activation in the hippocampus, a binding process packages the pattern of activation in the hippocampus into a functional unit. Understanding the binding problem is a central focus of memory systems research. (b) Retrieval works in reverse with volitional retrieval processes, or a cue from the environment, leading to reactivation of the bound episodic pattern in the hippocampus, which acts as a key to reactivate the cortical areas originally involved in perceiving the event that now store the representations of the event.

211

Figure 11. Participants studied a set of 20 pictures and 20 sounds across 2 days. They were scanned using fMRI during perceptual presentation, and during a final recall test on the 3rd day. Functional activations are displayed against a group structural scan (horizontal sections at various levels). Left panels (a, c and e) depict areas of differential activity for pictures (green in c) and for sounds (orange in e) during perceptual processing. Right panels (b, d and f) depict differential activations for pictures (green in d) and sounds (orange in f) during retrieval. Black arrows point to ventral temporal lobe (fusiform gyrus), and also areas of occipital and parietal cortex (d) associated with retrieval of visual items. Black arrow points to superior temporal lobe associated with retrieval of auditory items (f). Retrieval activates a subset of the cortex active during perceptual processing. From Wheeler et al., 2000.

212

Figure 12. An example of a portion of a hypothetical spreading activation semantic network that connects semantically and associatively associated concepts. Concepts that are semantically interrelated are presented in a common color. Associative connections, e.g. red and fire truck, based on co-occurrence of the concepts, use concepts of differing colors. The distance between concepts indicates strength of relationship. Processing a concept increases the activation level of the concept. Activation is proposed to spread along connections to other concepts.

213

(a) IS-A Relation

Lassie

Agent

Dog

Object

(b) Home

Agent

Relation

Lassie

Object

Agent

Relation

Object

RUNS IS-A

FINDS

Timmy

Agent Relation

Agent Relation Object

FALLS INTO

Object

Hero

Mine Shaft Figure 13. An example of a simple propositional semantic network using agent-relation-object proposition nodes to represent related factual statements. (a) Lassie is a dog, (b) Timmy fell into a mineshaft, Lassie finds him, Lassie runs home, and Lassie is a hero.

214

Figure 14. Participants viewed pictures of objects from the categories of tools, places, animals, and faces. Differential tool-related activity was observed in a grasp planning area in left ventral premotor cortex, and a motion integration area in left posterior parietal cortex (colored pixels superimposed over transverse MRI structural sections). Charts on left depict percent MRI signal change as a function of item category in the 2 tool-selective active regions. From Chao, L., & Martin, A. (2000). NeuroImage, 12, 478–484.

B. Nondeclarative (Implicit) Memory Nondeclarative memory systems operate outside of awareness. We are not aware of the influence of nondeclarative representations on our behavior, and we cannot consciously inspect the contents of these systems. Nondeclarative systems are qualitatively distinct from the declarative systems of episodic and semantic memory that we have discussed thus far. They support motor skill learning, behavioral habits, and a perceptual memory system that facilitates repeated perceptual processing of an object. There are also additional systems that support classical conditioning and nonassociative sensory learning that we shall not cover here as they are outside the scope of interest for a memory chapter from a cognitive science perspective.

“Nondeclarative Memory Systems”

215

Priming effects in nondeclarative (implicit) memory. Dr. Timothy Bender at Missouri State University has posted demonstration versions of implicit priming tasks online. The reader is encouraged to go to the Web page with these demos and to click on and run the (a) Implicit Memory Priming and Word Stems, (b) the Implicit Memory Priming and Anagrams, and (c) the Implicit Memory and Word Fragments demos. Perceptual Priming. Perceptual priming is a nondeclarative memory effect that demonstrates the important distinction between nondeclarative and declarative memory systems (Baddeley et al., 2009; Neath & Surprenant, 2003; Smith & Kosslyn, 2007). It allows our perceptual system to be unconsciously influenced by our previous experiences by directly facilitating identification of objects and events that repeat or are structurally similar to prior stimuli. Perceptual priming depends on the amount of perceptual overlap, within a particular modality, between successive perceptual events. Prior visual presentation of a word will prime (facilitate) processing of the same word repeated in visual form, but does not do much for a spoken presentation due to low perceptual overlap across visual and auditory presentations (e.g., Jacoby & Dallas, 1981). One task used to measure perceptual priming is the perceptual identification task where participants view words on a study list, then view extremely briefly flashed test words presented so briefly that they can only identify a few of the test words. However, when a briefly flashed test word was previously studied on the study list, the probability of correct identification is boosted. This facilitation effect is a form of perceptual priming, and participants are typically unaware that it is happening. Somehow neural representations of the prior experience with the same written words are available to influence current perceptual processing of the same visual shape. Another typical task is the word fragment completion task. During this procedure participants read some words. They are often given some meaningless task to do with the words, such as decide as quickly as they can if the word is an animal or a tool. They will then be given a distractor task, followed by the word fragment completion task. By ordering the tasks this way participants are typically unaware of any connection whatsoever between the first and the last tasks. Consider the situation where the word horse was on the previous word list for half the participants but not for the other half. Then imagine that participants are to fill in the blanks with the first words that come into mind: H _ _ _ _. Perceptual priming is the observed boost in the percent completion of the word with horse for the group that was presented with horse on the earlier word list, as opposed to the percent for those without this prior perceptual experience.

216

“Perceptual Priming” Amnesic patients with damage to the hippocampus and associated medial temporal lobes, e.g., H.M. (discussed in the Introduction to the present chapter) have impaired declarative memory such that they have trouble encoding new episodic and semantic representations. But, they have relatively preserved nondeclarative memory performance such that they produce perceptual priming effects on the perceptual identification and word stem completion tasks in the normal range. Amnesics such as H.M. also perform in the normal range of skill learning tasks. Perceptual priming is thought to be directly supported by most areas of the neocortex that perform perceptual processing. Procedural (Skill) Memory. Procedural memory refers to representations of the motor movements needed for skilled performance in a task. It also refers to learned behavioral patterns (habits). Humans have a remarkable ability to become experts at a wide variety of motor skills such as golf, typing, or even walking and talking at the same time. Laboratory skill learning tasks typically require a continuous movement sequence that can be measured for accuracy. For example, using a joystick to move a mouse on a computer screen and to keep the mouse on top of a moving dot. Learning occurs as the total amount of measured error declines with practice. Another example task would be the serial reaction time task. Imagine that you have 4 lights in front of you, when each light is lit up you are to press the associated button. Lights are lit in quick succession and you must struggle to keep up. The sequence of lights is long and appears to be randomly ordered. Unbeknownst to you there is a hidden sequence of 8 lights, 13412312, that repeats periodically. Typical skill learning for the sequence would result in improved speed of performance of the hidden embedded brief sequence with repetition. Medial temporal lobe amnesics will typically perform in the normal range in terms of a similar improvement with practice on hidden sequences (Gazzaniga et al., 2009). Another related nondeclarative memory effect is the stimulus-response habits (also referred to as operant conditioning) that reflect gradual development of motor movement sequences triggered by specific stimulus events (e.g., catching a ball that is unexpectedly thrown at you). Skill learning and habit formation are thought to depend on the function of the basal ganglia. It is worth noting that some skills also involve the cerebellum, and early on in skill learning motor control processes supported by frontal cortex are also important.

217

“Procedural (Skill) Memory”

VI. Process Perspective The process perspective on human memory focuses on the processes of encoding and retrieval, and how these processes depend on the information being encoded, the type of retrieval task, and the match between processes at time of encoding and retrieval.

A. Recollection & Familiarity Imagine that a friend is curious about a party you went to last night that she or he missed. Your friend asks if you met anyone new, and after a moment you tell them about the events surrounding meeting and talking with a new person. You tell your friend what the person looked like, was wearing, what their voice sounded like, and what they said to you. In doing so you are exhibiting a sort of memory retrieval process referred to as recollection, an effortful search for the consciously available aspects of episodic long-term memory. The memory task you faced was a form of recall. You were asked about a particular party, and you performed a strategic search for all information in event memory related to this party. Now imagine that instead your friend shows you a picture of a person on their cell phone taken at some other event and asks you if this person was also at last night’s party. Again you perform a conscious recollection of events from the night before. You do not retrieve any explicit recollection of seeing the person in the picture, but you have a strong feeling of familiarity when you look at the person’s face. This feeling is so strong that you feel that you must have seen the person at the party and you tell your friend that you believe the person pictured was at the party. However, just then you realize that you saw the person in the picture in line at the local coffee shop yesterday. You now have some doubt as to whether this person was at last night’s party. Familiarity is a retrieval process that is sensitive to having recently been perceptually exposed to an item, but the aspects of the memory context are not accessible by the familiarity process, leading to the possibility of bias in recognition memory (Neath & Surprenant, 2003). This second memory task was a form of recognition memory, where you are presented with an item and must indicate if it was part of a studied set of items or not. Many researchers have proposed that recognition depends on 2 types of retrieval processes, recollection and familiarity.

218

In a classic study, Brown et al. (1977) had participants view 10 individuals in-person. Participants were told that each of these 10 individuals had just committed a crime. After a 90-minute delay, participants viewed 15 mug shots and identified those they thought were in the previous criminal group. Five of the mug shots were of individuals in the criminal group viewed earlier, and 10 were new innocent individuals.

After a week delay, participants then viewed lineups of individuals. Most of the individuals were new, i.e., not included in the criminal group, or the innocent mug shot group. During this lineup participants correctly identified 65% of the criminals who had been in the mug shots as being criminals. However, they also identified 20% of the innocent mug shot individuals as being criminals they had seen in-person prior to the mug shots. But only 8% of brand new individuals only appearing in the line-up were identified as criminals. This is an example of the strength of bias that familiarity can have on recognition memory. The line-up is a recognition task, and presumably innocent individuals from the mug shot seemed familiar enough at time of line-up to be falsely recognized as being in criminal group. One reason for this may be that familiarity is not sensitive to context. That is, familiarity processes are sensitive to having seen the individual recently, but the source information of exactly in which context, criminal group or mug shots, is not accessible by the familiarity processes. One of the great successes of cognitive science in terms of influencing public policy was the U.S. Justice Department’s changing of the guidelines for appropriate in-person and mug shot line-up procedures in accordance with many of the recommendations of cognitive science researchers studying recognition accuracy and bias in eyewitnesses (see Wells et al., 2000, for a discussion of this successful effort on the part of cognitive scientists). Much of the procedural changes have to do with the fact that recognition memory depends on the 2 retrieval processes of recollection and familiarity.

B. Transfer Appropriate Processing Another example of the process approach to the study of human memory is the idea of transfer appropriate processing, a theory that encoding processes are most effective when matched to the type of processing at time of retrieval (Neath & Surprenant, 2003). In a classic study Morris et al. (1977) had participants judge whether a test word fit meaningfully into a sentence (e.g., the train had a silver engine), or whether the test word rhymed with another word (e.g., eagle rhymes with legal). Participants were then given a surprise memory test. They were given either a standard recognition test (e.g., was train presented earlier?), or a rhyme recognition test (e.g.,

219

was there a word that rhymed with beagle? Correct answer, “yes”: eagle). The interesting result was that correct memory performance was 20-25% greater when the encoding (rhyme versus identification) task and the retrieval task (rhyme versus identification) matched.

VII. Connectionist Models

Learn More About Connectionist Models Online Connectionist models are network models that attempt to model cognitive processes using interconnected networks of simplified neuron-like computational units (Dawson, 2005). The units (also called nodes) send and receive signals analogous to the neural communication in the brain. Inputs model the post-synaptic potentials (PSPs, see the Cognitive Neuroscience chapter). The strength of an input depends on the strength of the output signal from the sending unit combined multiplicatively with the strength of the connection between the units (called a connection weight). Connection weights can be positive to model an excitatory PSP in a real neuron (i.e., makes the unit more likely to send an output signal), or negative to model an inhibitory PSP (i.e., makes the unit less likely to send an output signal). The weights vary in magnitude (usually between -1 and +1) to model synaptic plasticity in real neurons where learning due to experience can result in a strengthening or weakening of synapses between neurons. Figure 15 presents an example neuron called a threshold logic unit (TLU, also called a McCulloch-Pitts unit after early researchers who proposed such units in the 1940’s). On the left of Figure 15 are the inputs, meant to model dendrites in neurons, the inputs are summed in the middle, and on the right is the threshold portion of the unit. If the sum of the weighted inputs, called the activation level of the unit, equals or exceeds the threshold setting (T in diagram), an output signal is sent. For the TLU the output is a +1 if the threshold is met by the activation level, and 0 otherwise. Operation of a TLU. Example inputs and outputs of an example TLU with 2 input and 1 output connection(s) is presented in Figure 16. In panel A only one input signal is received, but it is weighted at +1 (excitatory connection) and the unit’s activation level reaches threshold and an output signal is sent. By contrast, panel B depicts what

220

happens when the other input connection, with an inhibitory weight (-1), becomes active. The additional weighted inhibitory input drives down the activation level of the unit so that threshold is no longer met, and the output is withheld (output signal of 0).

Learn More About TLUs (also called McCulloch-Pitts units) Online Simple Associative Memory Network. We can build a simple memory system using TLUs of the type depicted in Figure 16. Figure 17 depicts a simple pattern associator that consists of 2 layers of TLUs. The cue layer accepts input signals and sends a pattern of signals to the output layer, which, in turn sends a pattern of output signals. Units producing an output (+1 output) are depicted as filled circles, and units that are not active enough to meet the threshold for sending a signal (i.e., an off neuron with an output of 0) are depicted as open circles. By adjusting the weights of the connections between the cue layer and the output layer, the network learns to associate specific cue patterns with specific output patterns. The system acts as an associative memory that takes a memory cue or hint and retrieves an associated pattern. For example, we could define the pattern of on and off units in the cue layer depicted in Figure 17 as representing the concept SALT, and we could define the output pattern of on and off units depicted in Figure 17 as representing the concept PEPPER. The pattern associator could then be seen to have stored the association between SALT and PEPPER in the pattern of weights of the connections between cue and output layers. Every time the cue pattern representing SALT is presented, the output representation of PEPPER will be output.

“Simple Associative Memory Network” Learning Algorithms. One of the great strengths of connectionist modeling is the connection weights that determine the memory storage of the network do not have to be adjusted by hand. Algorithms have been devised that allow the network to learn via incremental adjustments to the connection weights as the

221

network is exposed to a set of stimulus input patterns. One way to classify learning algorithms is the distinction between supervised and unsupervised learning algorithms. Supervised learning algorithms that work during a special learning phase where the network is taken offline and the operation of the network is in a learning mode only. This allows repeated cycling through the set of training input patterns, evaluation of the deviation of the observed output pattern in comparison to the correct output pattern, and the output error is used to incrementally change the connection weights to minimize the error for that input pattern the next time it is presented. With repeated cycling through the input and output patterns, the connection weight eventually approach an optimal setting to produce as many of the desired output patterns for each input pattern. In some cases, the network architecture may not allow complete learning of the inputoutput pairs and the network will need to be revised. An unsupervised learning algorithm, works online. Connection weights are incrementally changed as the network experiences as set of inputs. However, the network is not explicitly given a set of correct output patterns to produce for a given set of input patterns. Instead, the system must learn the informational structure of a set of inputs and learn to respond differentially to inputs of different types in much the same way that a human child learns to classify dogs and cats into different categories by direct experience with the world. Distributed Representation. Note that by defining a particular pattern of on and off units to stand in for a concept, this pattern is acting as a symbol in much the same way the word salt stands in for the concept SALT. Use of a pattern of activity across a set of units to represent information is known as a distributed representation scheme. A local representation scheme could be used by having each individual unit represent a concept as in the semantic network models discussed in the semantic memory subsection of the present chapter. For example, we could have defined the first unit on the left of each layer of the network depicted in Figure 17 as representing the concept SALT, and the second unit as representing the concept PEPPER. We then would have taught the network to turn on only the second-from-left output unit representing PEPPER when the cue layer has only the left unit representing SALT turned on via input. This would be a local representation scheme. Each unit is the equivalent of a symbolic representation of a concept. Subsymbolic Representation. One of the strengths of the connectionist approach to modeling is the flexibility that distributed representation schemes allow. Distributed representations are often defined as patterns of signaling across a layer of units. We can be restrictive in our modeling of cognition and require that for a particular concept, e.g., SALT, to be said to be active in the system, the exact pattern of on and off units must be observed. However, it is often the case that learning algorithms that attempt to find the optimal pattern of connection weights that will maximize learning result in input patterns that lead to output patterns that only partially correct. For example, imagine that the network depicted in Figure 17 learns to pair the input pattern for the concept

222

SALT with an output pattern that is partially correct for PEPPER in that only 3 of the 4 output units have the correct on-off output signaling values for the concept PEPPER. Use of distributed representation schemes that allow patterns of activity across units that partially match the defined pattern for representing information is known as a subsymbolic representation scheme. A symbolic representation is traditionally an allor-none proposition. For example the word salt represents the concept SALT, but the partial match sald, is a non-words that does symbolically represent the concept SALT even though it partially matches the symbol salt. From this analysis we see that written words are all-or-none symbols representing conceptual information. However, perhaps the same exact pattern of neurons in the brain do not activate each and every time we see the word salt. Perhaps the brain allows patterns of neural activity that are close partial matches to some prototypical activity to represent the concept SALT. In other words, perhaps the brain has a subsymbolic level of representation based on groups of patterns of neural activation. There has been much argument in cognitive science regarding whether subsymbolic distributed representation in connectionist models is truly different enough from the traditional symbolic representations to result in a truly distinct approach to cognitive science. Regardless of where one stands on this issue, subsymbolic distributed representation schemes that allow partial pattern matches across sets of units to have a representational role play an important part in connectionist models of cognition. Weakness of Simple Pattern Associator. Simple pattern associators such as the one depicted in Figure 17 are quite limited in their ability to learn sets of associated patterns. There are certain types of associations that just cannot be learned by such a simple network. That is, optimal sets of connection weights do not exist for many sets of input patterns the modeler would like the network to learn. Because of these limitations, a more powerful network architecture is required. Connectionist Units as Logic Gates. Traditional symbolic computation is based on the idea of formal operations on symbolic representations. The ability of a cognitive system to implement the formal operations of logic is an important aspect of this traditional form of computation. One question of interest regarding the individual TLUs is how computationally powerful they are in the traditional sense of computation. Of critical interest is the ability of the TLU to produce transformations of the input signals to output signals that are equivalent to the operations of formal logic. For example, it is easy to create a TLU that performs the formal operation of OR on 2 inputs. The logical operation of OR, with 2 inputs, returns a value of True if either of the inputs is True, and also if both inputs are True, but returns a value of False if both inputs are False. To create a 2-input TLU equivalent to OR, we simple define 0 as indicating a False input or output, and 1 indicating a True input or output. We quickly find that by setting the weights of both the input connections to +1, and setting the threshold of the single output unit to 1, the result is an OR-equivalent TLU. Table 1 presents the truth table for this TLU. As depicted in Table 1, the output is +1 except when both the inputs are 0, and then the output is 0. This pattern is equivalent to the truth table for the logical operation of OR.

223

This analysis quickly runs into a problem once we get to the logical operation of XOR, exclusive OR, where the output should be True if either of the inputs is True, but not if both are inputs are True. If both inputs are True, or if both inputs are False, the output should be False for the XOR operation. It turns out our simple TLU is not powerful enough to implement this critical logical operation. Historically, when researchers realized that networks that have the basic structure depicted in Figure 17, with just 2 layers of units, could not learn the structure of many sets of input patterns, and could not implement the XOR operation of logic, they realized that the basic network architecture for connectionist models of cognition had to change. The result was the 3-layer feedforward architecture of Figure 18. The 3-Layer Feedforward Network and Beyond. As depicted in Figure the standard network architecture that is powerful enough to learn the structure of most sets of input patterns, can solve the XOR problem, and for which a powerful learning algorithm has been discovered is the 3-layer feedforward network depicted in Figure 18. Signals flow only from lower layers to upper layers, hence the term feedforward, and there are no lateral connections between units within the same layer, and there are no feedback connections of a higher layer to a lower layer. This simplicity of design allows for application of a powerful learning algorithm for adjusting the connection weights. Common enhancements are to add lateral and feedback connections to the architecture. Another common enhancement is to upgrade the TLU so that it produces a continuous graded output between a maximum and a minimum. This models the idea of frequency coding in neurons (see the Cognitive Neuroscience chapter) where neurons send different signals to each other by varying the frequency or rate of firing action potentials between a minimum baseline rate of firing and a maximum physically possible. Added computational flexibility is gained by allowing units to have a continuous graded output rather than the binary 0 or +1 of the TLU.

224

Σ

T

Figure 15. A standard threshold logic unit (TLU). The unit accepts 2 input signals on the inputs on the left. Each input is weighted (often -1 to +1), and the activation level of the unit is determined by the weighted sum (i.e., each input 0 or +1 multiplied by the weight of its connection) as symbolized by the summation sign in the center. The activation level is compared to the threshold (T), and if the activation level meets or exceeds the threshold, a signal (+1) is sent out to all other units this unit is connected to.

A 0

-1

+1

+1

Σ = (-1⋅ 0) + (1⋅1) = 1

1

+1

Σ = (-1⋅ 1) + (1⋅1) = 0

1

0

B +1

-1

+1

+1

Figure 16. Example inputs and outputs for an example TLU with 2 input and 1 output connection(s). Panel A: Only one of the input connections has an input signal (+1), the other input connection has a non-input signal (0). Each signal is multiplied by the weight of its connection, and the products are summed (middle of panel A). The resultant activation level (+1) meets the threshold (+1) and an output signal is sent on the single output connection. Panel B: Both inputs signals are now at +1, and the inhibitory weight (-1) now comes into play and drives the activation level (weighted sum of inputs) down to 0, which, does not meet the threshold, and no signal is sent on the output connection.

225

Output Layer

Cue Layer Inputs Figure 17. A simple pattern associator built from TLUs. Darkened units indicate units active above threshold that are currently producing an output of +1, open units have an output of 0. The associator acts as a cued memory system by applying an input pattern to the bottom (cue) layer, and signals are then sent to activate a previously-learned associated pattern on the upper (output) layer. Weights of connections determine the learned associations between cues and outputs. Note that weights of connections and thresholds of units are not depicted to keep figure simple.

Figure 18. The improved network architecture that results from adding a middle layer to the 2-layer simple pattern associator depicted in Figure 17. All signals flow from bottom to top along unidirectional output connections. Inputs are applied to the input layer and signals are sent to the middle layers, referred to as the hidden layer, and from there, to the output layer. Note that there are now 2 sets of connection weights, for the middle and output layers, that determine learning. This architecture is referred to as a 3-layer feedforward only network because the only connections are from lower layers to higher layers in the network. There are no connections between units within a layer and no feedback connections to lower layers. Addition of lateral and feedback connections is commonly used to create a more complex network.

226

Table 1 Truth Table for an OR-equivalent 2-input TLU with Weights of +1 on Both Inputs and a Threshold of +1. Input 1 0 0 1 1

Input 2 0 1 0 1

Activation 0 1 1 2

Output 0 1 1 1

References Baddeley, A. (2003). Working memory: Looking back and looking forward. Nature Reviews Neuroscience, 4, 829-839. Baddeley, A., & Hitch, G.J. (1974). Working memory. In G.H. Bower (Ed.), The psychology of learning and motivation (Vol. 8). New York: Academic Press. Baddeley, A., Eysenck, M.W., & Anderson, M.A. (2009). Memory. Psychology Press. Braver, T.S., Cohen, J.D., & Barch, D.M. (2002). The role of the prefrontal cortex in normal and disordered cognitive control: A cognitive neuroscience perspective. In D.T. Stuss and R.T. Knight (Eds.), Principles of frontal lobe function (pp. 428-448), Oxford University Press. Brown, E., Deffenbacher, K., & Sturgill, W. (1977). Memory for Faces and the Circumstances of Encounter. Journal of Applied Psychology, 62, 311-318. Collins, A.M., & Loftus, E.F. (1975). A spreading-activation theory of semantic processing. Psychological Review, 82, 407-428. Corkin, S., Amaral D.G., Gonzalez, R.G., Johnson, K.A. & Hyman, B.T. (1997). H. M.’s medial temporal lobe lesion: Findings from magnetic resonance imaging. The Journal of Neuroscience, 17, 3964–3979. Cowan, N. (1995). Attention and memory: An integrated framework. New York: Oxford University Press. Dawson, M.R.W. (2005). Connectionism: A hands on approach. Oxford, UK: Blackwell. Dehaene, S. & Naccache, L. (2001). Towards a cognitive neuroscience of consciousness: basic evidence and a workspace framework. Cognition 79, 1–37. From Chao, L., & Martin, A. (2000). Representation of manipulable man-made objects in the dorsal stream. NeuroImage, 12, 478–484. Gazzaniga, M.S., Ivry, R.B., & Mangun, G.R. (2009). Cognitive neuroscience: The biology of the mind 3rd Ed., New York: Norton. Healy, A.F., & McNamara, D.S. (1996). Does the modal model still work? Annual Review of Psychology, 47, 143-172. Jacoby, L.L., & Dallas, M. (1981). On the relationship between autobiographical memory and perceptual learning. Journal of Experimental Psychology: General, 110, 306-340. Kohler, S., Kapur, S., Moscovitch, M., Winocur, G., & Houle, S. (19995). Dissociation of pathways for object and spatial vision: A PET study in humans. Neuroreport, 6, 1865-1868.

227

Logie, R.H., Zucco, G.M., & Baddelely, A.D. (1990). Interference with visual short-term memory. Acta Psychologica, 75, 55-74. Martin A, Chao LL (2001) Semantic memory and the brain: structure and processes. Current Opinion in Neurobiology, 11, 194--201. Milner, B. (1966). Amnesia following operation on the temporal lobes. In C.W.M. Whitty & O. L. Zangwill (Eds.) Amnesia. London: Butterworths. Milner, B. (2005). The medial temporal-lobe amnesic syndrome. Psychiatric Clinics of North America, 28, 599–611. Morris, C.D., Bransford, J.D., & Franks, J.J. (1977). Levels of processing versus transfer appropriate processing. Journal of Verbal Learning and Verbal Behavior, 16, 519533. Neath, I., & Surprenant, A.M. (2003). Human memory, 2nd Ed., Belmont, CA: Wadsworth. Ranganath, C., & Blumenfeld, R.S. (2005). Doubts about double dissociations between short- and long-term memory. TRENDS in Cognitive Sciences, 9, 374-380. Roediger, H.L., III, & McDermott, K.B. (1995). Creating false memories: Remembering words not presented in lists. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 803-814. Simmons, W.K., Martin, A., & Barsalou, L. W. (2005). Pictures of appetizing foods activate gustatory cortices for taste and reward. Cerebral Cortex, 15, 1602—1608. Smith, E.E., & Kosslyn, S.M. (2007). Cognitive psychology: Mind and brain. Upper Saddle River, NJ: Pearson. Sperling, G. (1960). The information available in brief visual presentations. Psychological Monographs, 74, 1-29. Squire, L.R. (2004). Memory systems of the brain: A brief history and current perspective. Neurobiology of Learning and Memory, 82, 171-177. Wells, G.L., Malpass, R.S., Lindsay, R.C.L., Fisher, R.P., Turtle, J.W., & Fulero, S.M. (June, 2000). From the lab to the police station: A successful application of eyewitness research. American Psychologist, 55, 581-598. Wheeler, M.E., Petersen, S.E., & Buckner, R.L. (2000). Memory’s echo: Vivid remembering reactivates sensory-specific cortex. Proceedings of the National academy of Sciences, USA, 97, 11125-11129.

Further Reading Recommended Memory Textbooks Baddeley, A., Eysenck, M.W., & Anderson, M.A. (2009). Memory. Psychology Press. Neath, I., & Surprenant, A.M. (2003). Human memory, 2nd Ed., Belmont, CA: Wadsworth.

228

Scientists (including a Nobel laureate) Write About Memory Kandel, E. (2006). The search for memory: The emergence of a new science of mind. Norton. Schacter, D.L. (2001). The seven sins of memory: How the mind forgets and remembers. Houghton Mifflin Harcourt. Squire, L., & Kandel, E. (2008). Memory: From mind to molecules. Roberts and Company Publishers. Recommended Memory Review Articles Baddeley, A. (2003). Working memory: Looking back and looking forward. Nature Reviews Neuroscience, 4, 829-839. Squire, L.R. (2004). Memory systems of the brain: A brief history and current perspective. Neurobiology of Learning and Memory, 82, 171-177. General Introduction to Connectionist Modeling Dawson, M.R.W. (2005). Connectionism: A hands on approach. Oxford, UK: Blackwell. Connectionist Modeling Applied to Semantic Memory Rogers, T.T., & McClelland, J.S. (2004). Semantic cognition: A parallel distributed processing approach. Cambridge, MA: MIT Press.

Glossary 3-Layer Feedforward Network A 3-layer network formed by taking the simple pattern associator and adding a third hidden layer of units between the input and output layers. A powerful architecture that can learn a wide range of pattern associations. Activation Level A value for a connectionist network unit that is based on the sum of its weighted inputs and that is compared to a threshold to determine if the unit will produce an output signal. Amnesia Loss of memory function. Binding Problem The problem of binding distributed representations in the brain into a packaged memory representation. Central Executive The cognitive control module in Baddeley’s WM theory.

229

Chunking (STM) A STM memory phenomena where verbal items (e.g., the letter sequence I-B-M) are combined into a single representation chunk (e.g., IBM). Connection Weight The strength of a connection between 2 units in a connectionist network. Connectionist Model A network model that uses neural-like interconnected units that send and receive signals to model cognitive processes. Consolidation A process of making memory representations increasingly stable. Thought to take place primarily during sleep. Declarative (Explicit) Memory A memory system that allows conscious access to its contents via volitional (explicit) retrieval processes. Digit Span Test A common STM span task that involves immediate ordered recall of randomly presented digits. Encoding, Storage, and Retrieval The 3 cognitive processes that define a memory system. A memory system acquires information and forms an appropriate code (encoding), the code is extended in time (storage), and the code must be accessible to retrieval processes. Episodic Memory An explicit (declarative) memory system for events. False Recognition A memory effect where a list of related words is presenting that is missing a word that is related to all of the words on the list. Later, during a word recognition test, participants will often indicate the missing word was presented. Familiarity versus Recollection An idea from the process approach to memory that retrieval during recognition memory tasks for events involves 2 processes. Recollection is similar to the retrieval used for recall tasks, an effortful strategic search results in retrieval of details of the event, e.g., what a person’s voice sounded like when you met them. Familiarity is a feeling of having recently seen the item on the recognition memory test that has become separated from information regarding the source of the familiarity, e.g., thinking you have met a person before when you have seen them at the local coffee shop but not met them.

230

Free Recall A type of memory test where no clues (called memory cues) regarding the correct answer are provided. Immediate Memory A generic term for a memory system that allows people to immediately repeat information back. Local versus Distributed Representation Connectionist models that represent words or concepts using single units are said to use a local representation scheme, and models that use a pattern of activation across a set of units to represent a word or concept are said to use a distributed representation scheme. Long-term Memory (LTM) A virtually unlimited capacity system for storage of information on timescales of minutes to years. Part of the modal memory model. Modal Memory Model The 3-store memory model that is a combination of the most common aspects of the dominant information processing memory theories of the 60’s and 70’s. Nondeclarative (Implicit) Memory A memory system that does not allow conscious access to its contents. Retrieval is only implicitly apparent in its effects on behavior. Operations Span Test A common WM span task that involves immediate ordered recall of sequentially presented words while concurrently solving simple arithmetic problems. Perceptual Priming System An implicit (nondeclarative) memory system for maintaining perceptually processed shapes in a partially active state to facilitate processing the same shape if it repeats. Perceptual Priming An implicit (nondeclarative) memory effect where a part or all of a shape is repeated and identification is facilitated (primed) for the second presentation. Phonological Loop The mental workspace module for temporary storage and processing of verbal information in Baddeley’s WM theory. Procedural Memory An implicit (nondeclarative) memory system for motor skills, behavioral habits, and cognitive skills.

231

Propositional Network A complex formulation of a semantic network that represents propositions (see Knowledge Representation chapter) as networks. Recognition A type of memory test where correct and incorrect answers are provided and the correct answers are to be chosen and the incorrect answers avoided. Retrograde & Anterograde Amnesia An amnesic syndrome that involves selective loss of memory for information learned before the onset of amnesia (retrograde, backward looking), or after the onset of the amnesia (anterograde, forward looking). Semantic Memory An explicit (declarative) memory system for facts. Semantic Network A network model of semantic memory that represents concepts as nodes or units, and relationships between concepts as connections. Semantic Priming A memory testing effect where a word is processed faster if presented immediately following a semantically (e.g., dog-cat) or associatively (e.g., salt-pepper) related word. Thought to be due to the operation of automatic spreading activation in semantic memory. Sensory Store Part of the modal memory model that extends sensory information for brief intervals. Serial Position Curve A curve obtained by graphing the percent correct free recall of words from a long word lists graphed as a function of each word’s position of presentation in the word list. Short-term Memory (STM) A limited capacity system for storing information for several seconds until it is needed. Part of the modal memory model. Simple Pattern Associator A 2-layer network that can associate patterns across its input and output layers. This architecture is limited in terms of the associations it can learn. Spreading Activation Theory The idea that concepts in semantic memory are activated when processed, and that some activation will spread to related concepts resulting in their partial activation thus facilitating subsequent processing of related concepts (e.g., semantic priming).

232

STM Span Test A test for measuring the capacity of STM to passively store information. Subsymbolic Representation Distributed representation schemes have to deal with how to represent close, but not perfect, representation patterns across a set of units. The use of similar patterns in representation is known as subsymbolic representation. Supervised and Unsupervised Learning Algorithms Learning algorithms adjust the connection weights to reflect connectionist network learning. A supervised learning algorithm trains the network using pairs of input signal and correct/desired output signal patterns. An unsupervised learning algorithm does not have access to correct/desired output patterns, but rather, trains the network learn the structure of a set of input signal patterns. Threshold Logic Unit (TLU) A simple unit that can only send binary (0, 1) signals, and which typically has input connection weights limited to the continuous range (-1 to +1). Threshold Criteria met by the activation level of a connectionist network unit in order to produce an output signal. Transfer Appropriate Processing An idea from the processing approach to memory that encoding processes work best when matched to the type of retrieval that will be performed later. Unit A neural-like unit that sums input signals to determine its activation level and sends signals to other connected units in a connectionist network if its activation level meets a threshold. Visuospatial Scratchpad The mental workspace module for temporary storage and processing of visual and spatial information in Baddeley’s WM theory. WM Span Test A test for measuring the capacity of the WM system to store information while performing concurrent cognitive processing. Working Memory (WM) A limited capacity system that acts as a mental workspace for both temporary storage and concurrent cognitive processing.

233

234