In this chapter we review the three main types of current project for naturalising semantics – informational (or causal co-variance) semantics; teleological semantics; and functional-role semantics. There are severe problems for each, though perhaps least for the last. We then argue that the natural status of content does not, in fact, require a fully reductive semantics, but can rather be vindicated by its role in scientific psychology.

Introduction

Recall from chapter 2, that one of the main realistic commitments of folk psychology is to the existence of states with representational content or meaning. This is then the source of what is perhaps the most serious eliminativist challenge to folk psychology (which is also a challenge to any content-based scientific psychology). This comes from those who doubt whether meaning and representation have any real place in the natural world. The problem is this: how can any physical state (such as a pattern of neural firing) represent some aspect of the world (and so be true or false) in its own right, independent of our interpretation of that state? The contemporary project of naturalising semantics is best seen as a response to this problem. In various ways, people have attempted to spell out, in purely natural terms (that is, terms either drawn from, or acceptable to, the natural sciences) what it is for one state to represent, or be about, another.

FIVE CENTRAL FEATURES OF THE THEORY

Computational modeling of human perceptual-motor and cognitive performance based on a comprehensive detailed information-processing architecture leads to new insights about the components of working memory. To illustrate how such insights can be achieved, a precise production- system model that uses verbal working memory for performing a serial memory span task through a strategic phonological loop has been constructed with the Executive-Process/Interactive-Control (EPIC) architecture of Kieras and Meyer. EPIC is characterized by five central features that may be compared and contrasted with those of other theoretical frameworks in this volume. These features include:

1. (1) Formal implementation with multiple component mechanisms for perceptual, cognitive, and motor information processing (cf. Barnard, Chapter 9; Lovett, Reder, & Lebiere, Chapter 5; Young & Lewis, Chapter 7; Schneider, Chapter 10).

2. (2) Representation of procedural knowledge in terms of a production system whose condition-action rules are all applied simultaneously and repeatedly during the cyclic operation of a central cognitive processor (cf. Lovett et al., Chapter 5; Young & Lewis, Chapter 7; O'Reilly, Braver, & Cohen, Chapter 11).

3. (3) Executive control procedures that schedule task activities efficiently and coordinate the use of limited-capacity peripheral perceptualmotor processors (cf. Baddeley & Logie, Chapter 2; Cowan, Chapter 3; Engle, Tuholski, & Kane, Chapter 4).

4. (4) Explicit simulations that accurately account for quantitative behavioral data (cf. Lovett et al., Chapter 5; Young & Lewis, Chapter 7).

5. […]

Working memory plays an essential role in complex cognition. Everyday cognitive tasks – such as reading a newspaper article, calculating the appropriate amount to tip in a restaurant, mentally rearranging furniture in one's living room to create space for a new sofa, and comparing and contrasting various attributes of different apartments to decide which to rent – often involve multiple steps with intermediate results that need to be kept in mind temporarily to accomplish the task at hand successfully. “Working memory” is the theoretical construct that has come to be used in cognitive psychology to refer to the system or mechanism underlying the maintenance of task-relevant information during the performance of a cognitive task (Baddeley & Hitch, 1974; Daneman & Carpenter, 1980). As reflected by the fact that it has been labeled “the hub of cognition” (Haberlandt, 1997, p. 212) and proclaimed as “perhaps the most significant achievement of human mental evolution” (Goldman-Rakic, 1992, p. 111), it is a central construct in cognitive psychology and, more recently, cognitive neuroscience.

Despite the familiarity of the term, however, it is not easy to figure out what working memory really is. To begin with, the term working memory is used in quite different senses by different communities of researchers. In the behavioral neuroscience and animal behavior fields, for example, the term is associated with the radial arm maze paradigm.

Without doubt, working memory is one of the “hottest” topics in cognitive psychology and cognitive neuroscience. Since the publication of Baddeley's (1986) landmark book, several monographs and edited volumes that explore various aspects of working memory have been published (e.g., Gathertole, 1996; Gathercole & Baddeley, 1993; Logie, 1995; Logie & Gilhooly, 1998; Richardson, Engle, Hasher, Logie, Stoltzfus, & Zacks, 1996; Vallar & Shallice, 1990). Also, at least four journals have published a special issue on this topic (International Journal of Behavioral Development, 1994; see also de Ribaupierre & Hitch, 1994; Memory & Cognition, 1993; Neuropsychology, 1994; Quarterly Journal of Experimental Psychology, 1996), and another is also planning to publish a special section in 2000 (Journal of Experimental Psychology: General). In July of 1994 – 20 years after the publication of the seminal article on working memory that defined the field (Baddeley & Hitch, 1974) – an international conference specifically dedicated to working memory was held in Cambridge, UK, bringing together more than 200 researchers from across the world.

Working memory is also one of the most intensively studied areas in a new emerging field of study, cognitive neuroscience. Reflecting the dramatic surge of interest in neuroimaging studies of working memory, two general science magazines (Science and Scientific American) recently published articles that report the state of the art of research inquiry into the neural basis of working memory (Beardsley, 1997; Wickelgren, 1997).

FIVE CENTRAL FEATURES OF THE THEORY

In a connectionist control network, working memory is implemented via short-term activation and connection changes that support cognitive operations. The CAP2 approach is a model of skilled processing and learning. When applied to working memory the model instantiates multiple forms and mechanisms of working memory. The major features are:

1. (1) Memory and processing occur in a multilayered hierarchy of modular processors with limited interactions and a single executive modulating activity. The CAP2 micro- and macrostructural characteristics and temporal expectations show parallels in cortical architecture and activation patterns.

2. (2) Memory takes the form of activation vectors in modules, fast and slowly changing connection weights within and between modules with different activation, interference, and decay effects.

3. (3) The control and regulation of working memory is performed by a hierarchical control structure of an executive using activity and priority reports from the network of modules and input of messages on the inner loop to monitor and modulate message traffic.

4. (4) The executive is a limited sequential processing network that can execute the production system–like sequential operations, which are particularly critical in learning new tasks and maintaining temporary variable information that is not coded in consistent association patterns in the modular network. The executive has local memory to maintain variable bindings and sequential procedures to control the network to perform cognitive tasks.

5. (5) Skilled performance involves automatic module-to-module transmissions that can perform consistent associative mappings with little loading of the executive.

This final chapter starts where the previous chapter left off (Kintsch, Healy, Hegarty, Pennington, & Salthouse, Chapter 12). The main goal of the current chapter is to offer some thoughts we have about the future directions of working memory research. In particular, we present our own view of where the field stands and where it may be going in the belief that such reflection on the “big picture” is something this field needs.

The organization of the chapter is as follows. We will first present six points of general theoretical consensus that appear to be emerging among the models of working memory included in this volume. Despite this globallevel agreement about the nature of working memory, there are some important disagreements among different models. Thus, we will next point out some unresolved theoretical issues for each of the eight designated questions. In the last section, we will outline several issues that have not yet received much attention in the current models of working memory, but we believe will become increasingly important for future empirical and theoretical investigations.

General Theoretical Consensus About the Nature of Working Memory

At the beginning of Chapter 1, we quoted H. J. Eysenck's (1986) rather pessimistic remark about psychometric theories of intelligence1 and pointed out that some people would probably feel the same way about working memory: There are many different models of working memory out there, but they all seem so different that it is difficult to see how they relate to one another.

FIVE CENTRAL FEATURES OF THE APPROACH

The embedded-processes model of working memory relies upon the following five principles, which emphasize links between memory and attention.

1. (1) Working memory information comes from hierarchically arranged faculties comprising: (a) long-term memory, (b) the subset of longterm memory that is currently activated, and (c) the subset of activated memory that is in the focus of attention and awareness.

2. (2) Different processing limits apply to different faculties. The focus of attention is basically capacity limited, whereas activation is time limited. The various limits are especially important under nonoptimal conditions, such as interference between items with similar features.

3. (3) The focus of attention is controlled conjointly by voluntary processes (a central executive system) and involuntary processes (the attentional orienting system).

4. (4) Stimuli with physical features that have remained relatively unchanged over time and are of no key importance to the individual still activate some features in memory, but they do not elicit awareness (i.e., there is habituation of orienting).

5. (5) Awareness influences processing. In perception it increases the number of features encoded, and in memory it allows new episodic representations to be available for explicit recall.

Two prior integrative reviews of information processing, an article (Cowan, 1988) and a book (Cowan, 1995), describe a view that will serve as my basis for discussing working memory.

FIVE CENTRAL FEATURES OF THE MODEL

We define working memory as controlled processing involving active maintenance and/or rapid learning, where controlled processing is an emergent property of the dynamic interactions of multiple brain systems, but the prefrontal cortex (PFC) and hippocampus (HCMP) are especially influential owing to their specialized processing abilities and their privileged locations within the processing hierarchy (both the PFC and HCMP are well connected with a wide range of brain areas, allowing them to influence behavior at a global level). The specific features of our model include:

1. (1) A PFC specialized for active maintenance of internal contextual information that is dynamically updated and self-regulated, allowing it to bias (control) ongoing processing according to maintained information (e.g., goals, instructions, partial products).

2. (2) An HCMP specialized for rapid learning of arbitrary information, which can be recalled in the service of controlled processing, whereas the posterior perceptual and motor cortex (PMC) exhibits slow, long-term learning that can efficiently represent accumulated knowledge and skills.

3. (3) Control that emerges from interacting systems (PFC, HCMP, and PMC).

4. (4) Dimensions that define continua of specialization in different brain systems: for example, robust active maintenance, fast versus slow learning.

5. (5) Integration of biological and computational principles.

Working memory is an intuitively appealing theoretical construct – perhaps deceptively so.

FIVE CENTRAL FEATURES OF THE THEORY

From the viewpoint of the Soar cognitive architecture, the term working memory (WM) refers to the psychological mechanisms that maintain information retrieved or created during the performance of a task. The following are the five key points made in the chapter concerning Soar's treatment of human WM:

1. (1) Soar is not specifically a “model of WM,” but rather a cognitive architecture of broad scope, which focuses on the functional capabilities needed for a memory system to support performance in a range of cognitive tasks. The functions of working memory are distributed across multiple components of the architecture, including the longterm production memory.

2. (2) Even in a cognitive architecture with an unbounded dynamic memory, WM limitations can arise on functional grounds. Where such functional accounts exist, they take theoretical priority over capacity-based explanations of WM phenomena.

3. (3) Soar does not currently include any capacity limits on its dynamic memory (SDM), but is compatible with certain such limitations. In particular, a constraint that SDM can hold at most two items of the same “type” (suitably defined) yields a coherent explanation for many psycholinguistic phenomena in the comprehension of sentences. This constraint is motivated by computational efficiency concerns and embodies the general principle of similarity-based interference (Baddeley & Logie, Chapter 2; Cowan, Chapter 3; Schneider, Chapter 10; and O'Reilly, Braver, & Cohen, Chapter 11, all in this volume).

4. […]

FIVE CENTRAL FEATURES OF THE MODEL

We describe a model of working memory that is developed within the ACT-R cognitive architecture. Some of its main features are derived from the basic features of ACT-R:

1. (1) Processing depends on the current goal of the system.

2. (2) The accessibility of declarative and procedural knowledge varies with experience.

In addition, the following features are important to working memory in particular:

1. (3) There is a limited attentional resource, focused on the current goal, that increases the accessibility of goal-relevant knowledge relative to other knowledge.

2. (4) In more complex and memory-demanding tasks, this limited resource is spread more thinly thus impairing retrieval of goal-relevant items.

3. (5) The “capacity” of this attentional resource may vary from person to person, influencing the ability to access goal-relevant information across domains.

In performing almost any cognitive task, one must engage working memory to maintain and retrieve information during processing. For example, in mental arithmetic (e.g., multiplying large numbers without pencil and paper), one must hold intermediate results in memory while solving the problem. Similarly, in sentence processing, one must maintain various syntactic and semantic structures until subsequent processing reveals their roles. Because working memory is involved in so many tasks, studying its characteristics and its impact on cognitive processes is critical to gaining a deeper understanding of how people perform cognitive tasks in general.

Past research highlights two important results, each of which demonstrates that working memory modulates task performance.

FIVE CENTRAL FEATURES OF THE THEORY

1. (1) We define working memory in terms of its function, namely maintaining efficient selective access to information that is needed to complete a given task. This function can be achieved in everyday skilled performance by a wide range of different mechanisms. In contrast, traditional short-term working memory employs only a small subset of those alternatives.

2. (2) The amount of information that can be maintained in accessible form in working memory for a specific task is not limited by a fixed capacity. As part of the extended skill acquisition necessary to attain very high levels of performance, experts acquire knowledge and skills to rapidly encode information in long-term memory such that the information can be efficiently accessed with retrieval cues (longterm working memory or LT-WM) whenever it is later needed to complete the task. Similar acquired mechanisms mediate the large working memory in skilled everyday performance.

3. (3) LT-WM is mediated by associative recall from long-term memory, and to function reliably it provides different types of mechanisms for overcoming the problems of interference resulting from repeated associations to related retrieval cues.

4. (4) LT-WM reflects a complex skill acquired to meet the particular demands of future accessibility for information with tasks within a particular domain of expertise. Domain-relevant skills, knowledge, and procedures for the task are so tightly integrated into the skills for encoding of information that the traditional assumption of a strict separation between memory, knowledge, and procedures is not valid for skilled performance.

5. […]

FIVE CENTRAL FEATURES OF THE THEORY

Working memory is a system consisting of those long-term memory traces active above threshold, the procedures and skills necessary to achieve and maintain that activation, and limited-capacity, controlled attention. The specific features of our model include:

1. (1) Domain-free, limited-capacity controlled attention.

2. (2) Domain-specific codes and maintenance (phonological loop and visuospatial sketchpad are two examples but the potential number of such codes is large).

3. (3) Individual differences in both 1 and 2, but individual differences in capacity for controlled processing are general and possibly the mechanism for general fluid intelligence. Although people can, with practice and expertise, circumvent the abiding limitations of controlled attention in quite specific situations, the limitations reemerge in novel situations and even in the domain of expertise if the situation calls for controlled processing.

4. (4) Limited-capacity, controlled processing is required for maintaining temporary goals in the face of distraction and interference and for blocking, gating, and/or suppressing distracting events.

5. (5) The dorsolateral prefrontal cortex (PFC) and associated structures mediate the controlled processing functions of working memory. We also argue that individual differences in controlled processing represent differences in functioning of the PFC.

A number of intellectual influences have served to shape our thinking about working memory (WM) and its evolution as a construct separate from that of short-term memory (STM).

What have we learned about working memory from the 10 models described in this volume (Chapters 2 to 11)? What answers have these models provided for each of the eight questions raised by the editors (Shah & Miyake, Chapter 1), and how do these answers relate to one another? Do seemingly different answers given to the same question really reflect some fundamental differences between the models, or should they be regarded as a difference in emphasis? More generally, how well have these models, as a whole, addressed each of the eight designated questions? Although the answers to these questions can ultimately be obtained by a critical reading of the chapters themselves and careful reflections on these theoretical issues, we hope to help readers with this task by providing systematic, issue-by-issue comparisons among the models of working memory covered in this volume. To this end, we will first analyze and comment on the answers provided by the contributors for each of the designated questions. We will then raise and discuss some general issues that transcend the specific questions.

Analyses of the Answers Provided to the Eight Designated Theoretical Questions

Question 1: Basic Mechanisms and Representations in Working Memory

The first question asked the contributors to outline their view of three mechanisms that provide a basis for working memory processes – the encoding, maintenance, and retrieval mechanisms. The question also asked the contributors to specify their assumptions about the nature of working memory representations.

FIVE CENTRAL FEATURES OF THE THEORY

1. (1) The cognitive mechanisms underlying working memory performance involve multiple processes and types of mental representation.

2. (2) The detailed properties of performance depend on the configuration of specific processes needed to accomplish the task and the specific types of memory records they access and use in executing the task.

3. (3) There are no specific capacity limitations on what is stored at any particular level of mental representation. Capacity limitation arises out of restrictions on the interfunctioning of processes within a wider system.

4. (4) The use of memory records requires the generation or revival of a description of the content to be accessed. This can also functionally constrain performance.

5. (5) There is no unified “central executive” component; central executive functions are themselves accomplished by processing interactions among subsystems.

The dominant approach to formulating theory within experimental psychology is to develop models of restricted scope and capability. Individual models strive to predict properties of behavior in tasks that are assumed to tap specific mental faculties such as visual perception, language, problem solving, emotion, memory, or motor skills. It is taken for granted that moving toward an understanding of the complete mental mechanism is rather like solving a jigsaw puzzle. Ultimately a complete picture should emerge as local theories become validated and as segments of increasing size emerge and are themselves pieced together. One problem with this approach is the very complexity of the interrelationships between the various mental faculties.