The sine qua non of skilled cognitive performance is the ability to access large amounts of domain specific information. For example, it is estimated that chess masters have access to as many as 100,000 familiar configurations of chess pieces (Chase & Simon, 1973). As another example, in order to make sense of what he or she is reading, a reader must have access to information gained from previously read text. This is particularly true when reading complex technical material filled with jargon.
Working memory (WM) refers to the simultaneous and temporary storage and processing of information. This definition dictates that the activation of extant knowledge brought to bear on ongoing performance is temporary. For example, the activation of contextual information, gained through previously read text and used to make sense of what is being currently read, is only temporary. Ericsson and Kintsch challenge this view by focusing on two questions: 1) "Can mechanisms that account for subjects’ limited working memory capacity in laboratory tasks [i.e., temporary activation] also account for the greatly expanded working memory capacity of experts and skilled performers?" and 2) "How can working memory based on temporary storage account for the fact that skilled activities can be interrupted and later resumed without major effects on performance" (p. 211)? The aim in addressing these questions is to show that skilled performers can expand STM capacity through reliance on domain-specific knowledge and control processes that allow for rapid encoding and retrieval of information in LTM.
Traditional Models of Memory and LT-WM
Problem solving, decision making, and other complex activities require rapid access to information. Within traditional models of memory, STM is the cognitive locus of these activities because LTM retrieval and storage processing are thought to be slow and error prone. That is, "On the basis of a century of laboratory research on memory many theorists have concluded that LTM can meet neither the criteria of speed and reliability for storage nor those for retrieval" (p. 213). Ericsson and Kintsch challenge these assumptions given that the severe limitations of STM "might seem far too restrictive to allow for human performance levels" (p. 212).
The rejection of LTM involvement in WM is based on two findings. First, storage of information in LTM is not reliable. Second, accounting for the retrieval of information in LTM, even if it could be reliably selected and stored, is problematic within standard memory models. Ericsson and Kintsch do not challenge these limitations of LTM, but disagree "with the stronger claim that the invariant characteristics of LTM rule out an expansion of working memory by storage in LTM in all types of performance" (p. 213).
Research on the skilled memory effect suggests that experts’ superior memory for domain-specific information is LTM mediated. Chase and Simon originally explained the skilled memory effect by arguing that experts hold labels for LTM chunks in STM. But Charness (1976) showed that a STM interference task between stimulus presentation and recall had no effect on performance. This ruled out a STM locus for experts’ superior memory. Two related findings form studies of mnemonists also implicate LTM in skilled memory performance: 1) tasks designed to interfere with STM have minimal effects on performance, 2) aftersession recall is high, and 3) practice with memory tasks produces improvements only for domain specific material. Also, research indicates that subjects actually rely on LTM storage during domain-specific performance. For example, skilled subjects’ memory is as good when memory assessment is unexpected as when the memory assessment is expected. The second issue concerns the rate of LTM access. Generally, research shows that it takes longer to retrieve and item from LTM than from STM.
Skilled memory theory is based on the idea of reliable storage and encoding of information in LTM. This is predicated on the existence of a large domain-specific knowledge base and on the association of encoded information with retrieval cues. A stable structure of retrieval cues is called a retrieval structure. Retrieval structures allow for the durable and rapid encoding of information in LTM, and for the recreation of conditions of encoding at recall. LT-WM is governed by retrieval structures. That is, "At a very general level we can characterize LT-WM as being mediated by a retrieval schema in which information the subjects has encountered is encoded and stored in LTM, where it is associated with its appropriate retrieval cues . . . " (p. 216). The available evidence suggests that retrieval structures mediate exceptional memory performance by allowing for fast, rapid, and flexible LTM access.
The mechanisms of LT-WM are next discussed. First, how do experts overcome proactive and retroactive interference? That is, when the task requires updating of information so that new information must be associated with a retrieval cue associated with previously stored information, how can subjects overcome proactive and retroactive interference? Two mechanisms are proposed. The first is recency: temporal information about the time when information was encoded can be used as a retrieval cue. Additional elaboration is not required. That is, linking information to a cue within a retrieval structure is sufficient. The second is elaboration: encoded information is linked with a retrieval cue, but also additional semantic links are formed between encoded information within and between supergroups. The working memory demands of the task dictate which encoding method will be used. Recency encoding will be used for activities that require only short-term maintenance of information (e.g., abacus). Elaborative encoding will be used for activities that require longer term retention (e.g., digit span).
Ericsson and Kintsch argue that LT-WM is not a "generalizable capacity." Rather, "LT-WM is acquired in particular domains to meet specific demands imposed by a given activity on storage and retrieval" (p. 221). A model of cognition is sketched to provide a context for this view. In this model, cognitive processes are conceptualized as a sequence of states, which are end-products of lower-level processes (e.g., sensory). A cognitive state is understand in terms of 1) the current state of the environment, 2) an individual’s beliefts, knowledge, etc. stored in LTM, and 3) previous cognitive states.
What is the relevance of this model to LT-WM? The modal view of working memory posits that the activation of the intermediate states that make up a cognitive process is temporary. By contrast, the LT-WM view is that for skilled activities (but not for all activities) the intermediate states are stored in LTM. That is, "Our account of LT-WM proposes that in skilled activities a significant part of the activated information is stored in LTM and is accessible through retrieval cues in STM" (p. 222). How can the two views be distinguished? Interruption then resumption of task performance at a particular state will result in loss of information if activation is temporary. By contrast, information will be retrievable if performance is based on LTM. What is the evidence?
The tenets of LT-WM are explicated with the example of text comprehension. A LTM representation of text consists of multiple levels, including surface, semantic, and so forth. According to Ericsson and Kintsch, this representation must be accessible so that it can be expanded by reading new texts. Thus, "Our main conjecture is that the accessible portions of this structure in LTM serve as an extended working memory (LT-WM)" (p. 223). Research by Glanzer shows that interruptions during reading affect reading time but not comprehension. Thus, "the transient portion of working memory (ST-WM) is not necessary for continued comprehension of the type of texts they studied" (p. 225). Ericsson and Kintsch argue that the increased reading time is required for the reinstatement of the retrieval structure that readers have generated for the text.
Reading comprehension is predicated on accessibility of information from previously read text. According to the LT-WM model, "Retrieval cues to the hierarchical organization of the encoded text provide access to this information, but direct access is limited to recent information as well as to elements central to the structure" (p. 226). This is supported by research showing that recognition of central propositions is faster than recognition of superfluous propositions.
Just and Carpenter and Engle propose models in which WM is viewed as capacity to maintain LTM activation. The LT-WM proposal by contrast reflects superior LTM encoding skill. Consistent with the latter view, knowledge seems to be the key predictor of comprehension. In summary, the capacity view holds that good readers have more room to store information in WM during reading. By contrast, the LT-WM view holds that "the total capacity does not differ between good and poor readers, but the processing efficiency of good readers is assumed to be higher, so that their effective working memory capacity is enlarged because they can use their resources better" (p. 229). The difference between the two perspectives is thus captured by the following distinction: resources vs. acquired skill.
Construction of Retrieval Structures
A detailed explanation of the construction process is provided in the paper. Briefly, Ericsson and Kintsch argue that abstract representations of a text is created while reading. These representations become a retrieval structure. The construction process is summarized as follows:
The important point here is that the links between propositions currently in the focus of attention and propositions in the long-term episodic text memory, which are established incidentally by the very nature of the comprehension process, make available to the reader a large subset of text memory in LTM, thus generating what we call LT-WM.
Expert Performance and LT-WM
Expanded working memory demands are accommodated by acquired memory skills. The structure of LT-WM depends on the nature of working memory demands. For example, mental abacus requires the frequent updating of information. Previously stored information would be susceptible to retroactive interference because it is not needed later in the solution process. This should result in poor aftersession incidental memory, a finding supported by research. As another example, dinner order must be encoded to allow for updating because people sometimes change their mind about what they want to order. In addition, the retrieval structure must accommodate similarities among dinner orders. Postsession recall is high, as predicted. In addition, the waiter JC encoded each item from a category (e.g., entrée) with previously ordered items from that category. This strategy allows for differentiation of same items within a category. In medical diagnosis, individual facts must be encoded before diagnosis. Patel and colleagues have shown that when presented with a random series of facts, recall is structured in terms of meaningful categories. Research on chess suggests that chess players use a retrieval structure in the form of an actual chessboard to plan ahead, play chess blindfolded, etc.
Summary, Comments, and Questions
Knowledge is certainly a prerequisite for skilled performance. But just as important, the performer must be able to rapidly and reliably access this information. Hence, "Our central claim is that in addition, [experts] acquire domain-specific skills to expand working memory capacity by developing methods for storing information in LTM in accessible form" (p. 239). Retrieval structures are tailored to the working memory demands of the task. This theory provides an account for how severe limitations of WM are acquired. But, "Our proposal does not abolish constraints on working memory; it merely substitutes new constraints on rapid storage in and efficient retrieval from LT-WM for the old constraints on ST-WM" (p. 240).
The model proposed in this paper has important implications for how human cognition is conceptualized. Consider the following statement: "Individual differences in the capacity of working memory are not fundamentally fixed and unchangeable. Instead, they are deliberately acquired" (p. 240). This view is quite different from the dominant information processing perspective in which constraints on human information processing (e.g., the capacity of STM) are invariant. Ericsson and Kintsch’s advocate a more situational view of cognition in which the situation dictates processing constraints. To illustrate, a chess master has a greatly expanded WM capacity when playing chess. Otherwise, he is normal.
The following question comes to mind: Do people perform activities in which they cannot use domain-specific knowledge to expand the functional capacity of WM? According to Baltes’ selective optimization with compensation framework, people restrict the range of activities they engage in as they grow older. That is, people tend to specialize in what they are good at, and, by implication, what they are most knowledgeable about. With this in mind, Ericsson and Kintsch’s perspective seems to suggest that age-related declines in cognitive functioning, and particularly in working memory, may not have important implications for everyday functioning. Indeed, there isn’t a relationship between age and work performance, and the best predictor of work performance is job knowledge, with only indirect effects of fluid cognition.
On the other hand, one might imagine that the efficiency and rapidity of LTM access (i.e., of LT-WM) is to some extent based on WM capacity. Indeed, "even an account based on LT-WM requires sufficient capacity of ST-WM to allow retrieval from and encoding of the presented text in LTM as well as maintenance of necessary retrieval cues" (p. 227).
Additional Notes on Structure of LTM from Anderson (1995)
Ericsson and Kintsch propose that LTM knowledge structures called retrieval structures mediate expert performance by allowing for fast and reliable access to LTM. The functional capacity of ST-WM is thereby expanded. In order to understand this idea better, it is useful to consider in more detail the structure of LTM.
In a propositional representation, units of knowledge are represented as propositions—the smallest unit of knowledge than can stand as a true/false assertion. An example is "Lincoln was a president of the USA during a war" (p. 145). Typically, a proposition is expressed as follows: (president-of, Lincoln, USA, war). President-of is called the relation, and Lincoln, USA, and war are called arguments. This notion captures the meaning of the sentence "Lincoln was a president of the USA during a war" without the specific wording. Propositions can also be represented graphically as a network of nodes. Each element in the proposition (president-of, Lincoln, USA, and war) are represented as nodes. Arrows connecting the nodes are called links, and labels for the links specify the nature of the relation.
How does activation of a node in declarative memory result in action? In other words, how do declarative knowledge and procedural knowledge interact?
Declarative Memory and Procedural Memory
Anderson’s model consists of three memory systems: working memory, declarative memory, and procedural memory. Environmental input is temporarily stored in WM, and is encoded into declarative memory in propositional format. Consider the following situation. John says to Daniel "Barbara is the mother of Rebecca." This proposition (A) would enter working memory and be encoded in the following format: (mother-of, Barbara, Rebecca). Graphically, an ellipsis represents the proposition, and arrows extend to Rebecca (object), Barbara (agent), and mother-of (relation). At a later time, Bill says to Daniel "Joyce is the mother of Barbara." Three things now happen. First, the statement is encoded into declarative memory in propositional format. Second, activation spreads to related nodes--mother-of and Barbara. And third, by virtue of their activation, the proposition A becomes part of the activated subset of long-term memory—in other words, it enters working memory.
At this point, propositions A and B are simultaneously active. That is, they are both "in" working memory. Now assume that the following production rule has been established and resides in procedural memory:
IF Person 1 is the mother of Person 2
And Person 2 is the mother of Person 3
THEN Person 1 is the grandmother of Person 3
The production rule is executed because the conditions specified in the IF portion of the statement match the contents of WM. Execution results in the following product: Joyce is the grandmother of Rebecca. This product enters working memory as a thought, which can then be transferred into a vocalization such as "In that case, Joyce must be Rebecca’s granddaughter." At the same time, the proposition "Joyce is the grandmother of Rebecca" is stored in declarative memory, thereby increasing the declarative knowledge base. (See Benjafield for a diagram of this process.)
How does this apply to complex activities such as chess? In chess, the declarative knowledge base consists of spatial patterns representing familiar configurations of chess pieces. Simon and Gilmartin (1973) estimated that expert chess players have a vocabulary of as many as 100,000 of these patterns. Spatial patterns can be represented in propositional format. Procedural memory can be conceptualized as a knowledge base containing production rules specifying what actions should be taken if a particular pattern is encountered. As Larkin, McDermott, Simon and Simon (1980) explain, "recognition of a pattern often evokes from memory store information about actions and strategies that may be appropriate in contexts in which the pattern is present" (p. 1336). The pattern is recognized because it is contained in declarative memory. The symbolic, propositional code representing the pattern is activated, and hence becomes WM (recall that WM is the activated subset of LTM). The match between the pattern in WM and the IF-portion of a production rule triggers the execution of that production rule. The production of the production rule (specified in the THEN portion of the statement) enters working memory, and the appropriate move is made.
Does Anderson’s model apply to physical domains that require not only a cognitive action, or decision, but also the execution of some physical movement? Research on expertise in physical domains show that very specific adaptations occur through training. Research by Abernethy and colleagues shows that players in ball sports (e.g., baseball, cricket) rely on advance visual cues to anticipate the actions of their opponents. That is, they recognize familiar patterns of activity that allow them to prepare. It seems reasonable to argue that, just as in chess, these patterns are the IF portion of IF-THEN production rules. Unlike in chess, however, the action is a non-trivial physical movement, such as swinging a baseball bat. Some of these physical movements might be possible only if certain physical adaptations have occurred through training. For example, in baseball, strength changes probably occur through batting practice. In the absence of these changes, successful performance would seem unlikely. This example illustrates how successful performance in physical domains might be predicated on parallel cognitive and physiological adaptations.