Cognitive Systems

Human Cognitive Models in Systems Design

Chris Forsythe, Michael L. Bernard, Timothy E. Goldsmith

• Publisher: Psychology Press
• Year: 2006
• Print ISBN: 9780805852912
• eBook ISBN: 9781135605377
• Edition: 1
• Topic: Computer Science
• Subtopic: Artificial Intelligence (AI) & Semantics
• Index: Computer Science

The leading thinkers from the cognitive science tradition participated in a workshop sponsored by Sandia National Laboratories in July of 2003 to discuss progress in building their models. The goal was to summarize the theoretical and empirical bases for cognitive systems and to present exemplary developments in the field. Following the workshop, a great deal of planning went into the creation of this book. Eleven of the twenty-six presenters were asked to contribute chapters, and four chapters are the product of the breakout sessions in which critical topics were discussed among the participants. An introductory chapter provides the context for this compilation. Cognitive Systems thus presents a unique merger of cognitive modeling and intelligent systems, and attempts to overcome many of the problems inherent in current expert systems. It will be of interest to researchers and students in the fields of cognitive science, computational modeling, intelligent systems, artificial intelligence, and human-computer interaction.

PART ONE: INTRODUCTION

CHAPTER ONE: Cognitive Models to Cognitive Systems

Chris Forsythe
Patrick G. Xavier
Sandia National Laboratories

Although sometimes rooted in nature, the problems plaguing human existence, ranging from the societal to the mundane, almost universally involve a cognitive dimension. Assembly instructions may be barely decipherable. Software may behave in unexpected ways. An auto accident may follow from misinterpreting the intentions of another driver. The cure to a pervasive disease may remain just beyond reach due to the inability to grasp insights that appear obvious in hindsight.

In each of these examples, humans utilize their cognitive resources, including their unique knowledge and experience, to interpret stimuli within their environment and operate on the environment to accomplish various objectives. Throughout history, humans have created artifacts to bolster and expand the potential of their innate cognitive capacities. Spoken and written languages have allowed communication concerning objects that are not actually present or events that have not been personally experienced. Numbers have provided a means to reason about the world using commonly understood conventions. Computers have enabled the automation of basic cognitive processes and the storage of vast volumes of knowledge and experience.

Our inspiration in advancing ideas within the rubric of what we have termed cognitive systems owes to an awareness of the cognitive nature of the problems confounding human endeavors and a belief that fundamental progress in overcoming these problems requires a new artifact. We assert that cognitive systems, as described in the following sections, may be the artifact that irreversibly and positively changes how people interact with the world.

WHAT ARE COGNITIVE SYSTEMS

There have been numerous varied instances in which others have used the term cognitive systems, and there is no collective agreement with regard to a definition, scope, or best illustration. Duly conscious of this ambiguity, we have chosen to refer to the technologies described in the remainder of this chapter, and subsequent chapters of this book, as cognitive systems. Therefore, it is appropriate that our discussion begin with an explanation of how we have conceptualized a cognitive system, although our intent is not to argue for a common definition.

In defining a cognitive system, we have chosen to draw on the properties of the most common cognitive system and the one with which most of us have had the most experience: ourselves and the other humans with whom we interact. In the framework elaborated here, we assert that humans represent the prototypical cognitive system. We further assert that the objective in creating a machine-based cognitive system is to emulate the properties of human cognition that enable people to effectively engage the world and interact with other humans as fellow cognitive entities.

In our opinion, the most compelling case for the creation of cognitive systems is to enable machines to interact with humans in a knowing manner that is similar to the way in which humans interact with one another. To explain what we mean by this statement, consider the following example.

Imagine you have another human who serves as your aide. It is the objective of your aide to first follow you everywhere that you go, listen to every conversation to which you are a party, read everything that you have read, and remember everything that you have done, how you have done it, and the consequences of your actions. Second, it is the objective of your aide to do everything within his or her power to help you as you go through your day-to-day endeavors; however, this aide has the benefits of unlimited memory accompanied by perfect recall and the processing capabilities of the most powerful computing systems.

The characteristics that may be attributed to such an aide are the same characteristics that we would assert as generally descriptive of cognitive systems:

• They know what you know, including the underlying structure of your knowledge, and what you don’t know.
• They know what you do and how you do it, including the knowledge implicit in your actions.
• They know about your past experiences and can properly place events within the context of past experiences.
• They can apply your unique knowledge and experiences to interpret events in a manner consistent with how you would interpret the same events.
• They recognize when you have learned and how learning has re-shaped your knowledge of the world.
• They know the consequences of your past experiences and the resulting sensitivities, and they can anticipate how you will react to future situations.

One may rightly assert that humans are imperfect cognitive systems. Memory is fallible. People falsely attribute beliefs to others. People adopt superstitious behavior. However, one cannot disregard the remarkable effectiveness with which humans interact, particularly when placed in comparison with troublesome facets of human–machine interactions. Consider the case in which two individuals who have never met one another and speak different languages engage in mutually beneficial commerce. Second, consider the ease with which two long-time collaborators operate together mixing various levels of abstraction and drawing on shared past experiences. Finally, consider the nonintuitiveness of much commercial software, the seeming obtuseness of associated online help functions, and the misdirected behavior of automated features meant to assist the user. It is our intent with cognitive systems to emphasize those facets of human cognition that make for effective human interaction while compensating for certain shortcomings of the human cognitive apparatus through the processing power and faithful data storage and retrieval possible with computer systems.

At first blush, one may respond negatively to our description of a cognitive system on the basis of privacy concerns. It can be intimidating to imagine one’s actions faithfully observed and remembered. One can postulate numerous scenarios in which such data are used for detrimental purposes. We must acknowledge that there are no easy answers for such concerns. Of course the developers of cognitive systems may take measures to ensure the user is always in control of their personal data collection and the access of others to that data. However, as has repeatedly been the case with technological innovation, broad adoption of cognitive systems will create new possibilities for both the opportunist and, of notable importance, those who will defend against these opportunists.

In answer to the question, “What is a cognitive system?”, we have chosen the following delimitation. A cognitive system refers to a variety of software products that utilize plausible computational models of human cognitive processes as a basis for human–machine interactions. The intent is to reproduce cognitive mechanisms responsible for the effectiveness of human–human interaction so that the human–machine interaction becomes more like an interaction between a human and a human-like cognitive entity. In short, a cognitive system consists of software that helps a machine interact with people in the way people interact with one another.

PLAUSIBLE COMPUTATIONAL MODELS OF HUMAN COGNITION

In our definition of a cognitive system, we have placed a heavy emphasis on plausible computational models of human cognition. The goal is to attain the highest level of psychological plausibility that is practical for a given technology. This distinguishes cognitive systems from those that only emulate human surface features. For instance, a system may feature dialogue-based interaction with a humanlike avatar as the means to emulate a dialogue between the user and another human; however, the underlying structure for the software may be entirely rule-based, bearing little resemblance to human cognitive processes. From a different perspective, a system may utilize mainstream approaches from machine-based reasoning for problem-solving, information retrieval, user customization, or other intelligent functions. Yet these approaches are based loosely, or not at all, on human cognitive processes. Neither of these examples would clearly qualify as a cognitive system using the definition we have chosen.

Our preference is to avoid an unreasonably purist or dogmatic stance. Benefits may arise from using dialogue-based interaction and humanlike avatars as interfaces to cognitive systems that are otherwise based on plausible models of human cognition. Similarly, for a given cognitive system, all facets do not need to emulate human cognitive processes. For instance, our own approach uses the most expedient algorithmic techniques to acquire the knowledge that populates a plausible computational model of human knowledge representation and cognitive information processing, without attempting to model the mechanisms underlying human learning.

In our discussion, we have made repeated reference to plausibility as a critical attribute of the computational models utilized in cognitive systems. We acknowledge that plausibility can have many legitimate dimensions that may or may not be quantifiable. Furthermore, it is unrealistic to expect general agreement with regard to a standard for plausibility.

Our emphasis on plausibility is largely meant to distinguish cognitive systems from alternative approaches that either make no claim to emulate human cognitive processes, or cite human cognitive processes, but cannot show a clear correspondence between computational representations and scientific understanding of the mechanisms underlying human cognition. Given the latter statement, one may correctly conclude that plausibility represents a changing standard that evolves with associated scientific understanding. Thus, computational models that would have seemed plausible when knowledge of human cognition was almost exclusively the product of research in experimental psychology may now be suspect given the proliferation of knowledge concerning the mechanisms underlying human cognition that has become available from brain imaging and EEG studies.

Our own claims to plausibility derive from the methodology we have employed in designing and verifying our computational model. This methodology begins with a survey of the experimental literature concerning human cognitive processes and corresponding neurophysiological processes. From this survey, we identify functional characteristics. For instance, from the work of Klimesch and colleagues (e.g., Klimesch et al., 1997), we establish that:

Generally, there should be an increased desynchronization of activation in the upper alpha bandwidth (10–13 Hz) during periods in which semantic memory is engaged.

These characteristics then serve as functional design specifications. In designing the computational model, the intent is to engineer a design that in operation will satisfy a collection of functional specifications similar to the example given earlier. Once implemented, simulation scenarios may be developed that are consistent with the experimental procedures used in the studies on which specifications were based, and these simulation scenarios may be presented to the computational model to verify that the model responds consistent with specifications (i.e., the performance of the model in simulated scenarios corresponds to findings of the original studies). In full acknowledgment that the scientific literature does not provide a complete specification for human cognition, we contend that the path to plausibility lies in a process for model development that integrates scientific understanding with structured engineering practices.

It may be asked, “Is plausibility a necessary attribute for cognitive systems?” With the various manifestations of cognitive systems described in subsequent sections, there are many cases in which nothing precludes the use of models with weak plausibility or even models for which plausibility was never the intent. However, we would assert that plausibility is an essential attribute to attain a knowing and responsive human–machine interaction of the nature proposed here.

For a human–machine interaction that resembles the interaction between two cognitive entities, we first assert that the machine should possess an accurate model of the knowledge of a user(s). This assertion is based on a desire for the machine to adjust its interactions with the user to accommodate specific facets of the user’s knowledge of and experience with a task, system, or domain. Additionally, cognitive systems will be discussed that enable a user to compare his or her own knowledge with that of other individuals (i.e., perhaps experts in a given domain). In either case, we believe the most direct and parsimonious approach to accomplish an accurate model of the knowledge of a user is to base that model on a plausible representation of the underlying structural properties of human knowledge.

Second, we assert that the machine should possess an accurate model of the cognitive processes of a user. With certain cognitive systems, the machine will interact with the user on the basis of inferred cognitive operations. For example, the machine may adjust the salience of perceptual cues to enhance the situation awareness of a user and facilitate recovery from situation mode errors. Where inferences regarding a user’s ongoing cognitive processes are an essential facet of the cognitive system and erroneous inferences may have deleterious effects, we believe the correctness of such inferences will be a direct product of the accuracy with which models represent the corresponding cognitive operations of the user.

Finally, cognitive systems concepts are discussed that provide the user with tools for simulating the behavior of individuals, groups, or populations. As an aide or synthetic mentor, the simulation may be based on a specific individual and the user allowed to create scenarios and observe the response of the simulated individual(s). Other concepts augment the hypothetical reasoning of users by allowing them to create simulations involving groups or populations. Here the objective is to allow users to simulate hypothetical situations to gain an understanding of the range of possible outcomes and how various factors interact (e.g., an individual may run several simulations to see how highway traffic could be affected by a public event). In each of these cases, the accuracy and usefulness of tools will depend on the accuracy with which they forecast human behavior. We assert that plausible cognitive models are necessary to attain a realistic portrayal of human behavior, including the complex interactions between different individuals and groups.

ESSENTIALS FOR A PLAUSIBLE COGNITIVE MODEL

The preceding section offered a brief description of the development process on which we base claims concerning the plausibility of our cognitive modeling framework. We do not assert that the framework developed by our group is the only plausible computer model of human cognition. Our framework embodies a particular level of representation and a specific combination of cognitive processes. Nothing precludes other frameworks that operate at different levels of representation or emphasize alternative combinations of cognitive processes. Furthermore, the precise computational mechanisms for representing cognitive processes may vary across frameworks, with alternative frameworks producing equivalent outcomes. Although alternative models may provide for plausibility, we cannot overstate the importance that models be based on a systematic development process rooted in science concerning human cognition.

The following sections summarize properties of human cognition that we believe are fundamental and have included in our cognitive modeling framework. Figure 1.1 provides a graphical depiction of this framework. Our framework addresses a specific combination of cognitive processes, and there is varying overlap with other reasonably plausible models.

Concepts

The basic unit in our cognitive framework is the concept. Concepts correspond to the most elementary units of cognition, enabling an entity to recognize and respond to stimuli. It may be noted that here, recognize implies knowledge, which is consistent with the contention that cognition is knowledge-mediated action whether observable action or mental processes.

Within our framework, cognition begins with recognition of meaningful regularit...

Table of contents

• Cover Page
• Title Page
• Copyright Page
• Contributors
• Preface
• Part One: Introduction
• Part Two: Theoretical and Empirical Basis For Cognitive Systems
• Part Three: Illustrations of Cognitive Systems
• Part Four: Topics in Cognitive Systems