The Swimmer

An Integrated Computational Model of A Perceptual-motor System

William R. Uttal, Gary Bradshaw, Sriram Dayanand, Robb Lovell, Thomas Shepherd

• Publisher: Psychology Press
• Year: 2014
• ISBN:
  9781317782414
• Edition: 1
• Topic: Psychologie
• Subtopic: Histoire et théorie en psychologie

This research monograph describes a large programming project in which an underwater organism, capable of perceiving, learning, deciding, and navigating, is computationally simulated. The developed computational model serves as a contemporary theory of perceptual-motor performance, embodying much of what is known about human vision and some of what is known about other cognitive processes. This artificial intelligence project has substantial contributions to make to the development of autonomous underwater vehicles. It also makes a specific theoretical statement about the organization and nature of organic perceptual motor systems that may be useful to psychologists, neuroscientists, and theoreticians in a number of other fields.

1	Introduction

A recent explosion of activity in the development of computational models of human perceptual, decision making, and response processes has been made possible by modern developments in cognitive theory and computer technology. A wide variety of programmed algorithms has been developed that simulate and describe what are usually only the parts and pieces of perceptually guided behavior. In the visual domain, for example, algorithms have been developed that perform such visual functions as edge enhancement (e.g., Argyle, 1971), object segmentation based on brightness, color, or texture (e.g., Landy & Bergen, 1991), and the generation of object depth from disparity, shading, or contour lines (e.g., Horn & Brooks, 1989). Other models have been developed that describe and analyze performance during motor tasks such as walking or manipulation or that imitate such performance in the form of a mechanical device.

A characteristic of this present joint theoretical science and practical engineering endeavor is that most current energy has been aimed at the fine-tuning of the algorithms representing separate segments, aspects, or components of the total perceptual-motor process. Current goals have mainly been to invent or improve the individual algorithms that carry out isolated components of the total process rather than to integrate them into a coherent aggregate that performs some behavioral task in its entirety.

Similarly, much of the effort to study perception and performance in humans or animals has been directed at examining the components of human cognition. Researchers in this field tend to study fragments of mentation rather than the whole process. However, in the last decade it became clear that most cognitive processes, no less than complete perceptual-motor tasks in general, are actually accomplished by the integration of numerous attributes or mechanisms rather than driven by a single dimension of the stimulus or psychobiological mechanism. We showed this to be the case for a sampled surface reconstruction perceptual task (Uttal, Davis, Welke, & Kakarala, 1988). Anderson (1981) makes this concept a central point of his important work, and Massaro and Friedman (1990) have added greatly to our understanding of attribute and dimensional integration by their careful comparison of a number of alternative theories of information integration.

The promise of this present work, in this same spirit, is that there is much to be learned, both by the student of artificial intelligence and by the student of natural intelligence, by considering an entire perceptual-motor process at once. The premise upon which the present work is founded, therefore, is that perceptually guided behavior is not simply line detection, nor stereopsis, nor control of movement in space. Rather, it is a complex of functionally integrated processes, of which these are only a few examples, that mutually affect and interact with each other. Visual inputs evoke perceptual or interpretive responses that are necessary to define the parameters of the environment and the objects it contains. The interactions between the stimuli, the natural or synthetic organism, and the environment set goals, the achievement of which requires specific meaningful and relevant responses. The process is a highly adaptive and dynamic feedback system in which the very act of behaving in an appropriate manner to close on those goals can alter the sensory inputs as new relationships occur during that behavior.

The integrated computational model that we present here emphasizes traditional psychological variables such as a figure from ground extraction, reconstruction of forms, closure, discrimination, recognition, and some primitive decision making. It is not unique, it must be acknowledged in another closely related context—the engineering of artificially intelligent systems. Several other efforts in that context were directed at the design of autonomous, automobile robots. Among the most notable of these projects comparable to our own, have been those summarized in a recent article by Busby and Vadus (1990). However different our long-term goals may be, all of these efforts obviously share common methods and techniques.

The best way to communicate our plan is pictorially. Figure 1.1 presents the global outline of our project as initially planned. At present, all of the blocks have been programmed and installed in an operating version of the model with the exception of the motion and contours and optical flow modules. This flowchart indicates the rudiments of both the horizontal and vertical integration, but the full implications of this very abbreviated flowchart can only be appreciated when it is realized that any of the connecting lines may also itself be instantiated in the form of a set of serial and parallel processes.

The flowchart is also a theory of human perceptual processes in that it analogizes some of the perceptual mechanisms that are necessary for human vision. It outlines, however briefly, the relationships we believe to exist among the various components that underlie perceptual-motor behavior.

In the rest of this book, we report the development of this computational model of an entire perceiving and responding organism. Although we place primary emphasis on the integration of a suite of algorithms and processes, we also report considerable progress in improving the structure of a number of individual programming modules. Each of the functional steps in our simulation is considered to be tentative and replaceable as the technology for executing that particular subprocess evolves. The simulated perceiving and behaving entity we describe in this report is thus an integrated, but easily modified, concatenation of a wide variety of actions, interactions, and transformations. The computational model representing that entity simulates or analogizes the actions that “must” be carried out (in Marr’s, 1982, terms) by an organic nervous system that would function in an equivalent manner when confronted with the same environment. However, it is important to note that it may do so by means that are quite different than those used by an organism. Indeed, we have made no effort in this work to specifically imitate psychological or physiological mechanisms. Rather, it is the sequence and integration of processes and transforms that we are simulating. Thus, our mathematical tools are typically chosen from the ordinary forms of classical analysis (e.g., integral equations and Fourier transforms) or from a set of simple computational interactions rather than from some currently popular connectionist concept of parallel neural nets.¹

The actions that have to be simulated range from the initial transduction of photic energy to the appropriate interpretation of a stimulus scene to the construction of a map of the world of the simulated entity to a specific effector response. Thus, our simulation—to the degree that it is successful—represents a descriptive transformational theory of this kind of behavior in an entire organism as opposed to a restricted model of a single internal stage of information processing. We repeat: It is not intended to be a valid reductive explanation of the actual inner workings of the brain or mind of any real organism, but a description of the transformations and processes that have to be accounted for if we are to understand the complexities involved in perceptual-motor behavior.

FIG. 1.1. A flowchart showing the horizontal and vertical integration of the various components of our model of perceptual-motor processes. With the exception of the motion and contours and optical flow modules, which have not yet been programmed, all of the other components of the system are described in this book.

Another general premise of our project is that both psychological theory and computer technology have developed to a point where it is now possible to consider collecting many existing algorithms, procedures, ideas, and theories into an integrated model of a complex perceptual-motor skill. While we feel that we are pushing the frontiers in this kind of theorizing, we do not feel that we are beyond the current limits of science and technology. We are convinced that what we have chosen to do has been, from the outset, a realizable goal simply because it is synchronous with modern developments in both computer technology and cognitive science.

To pursue such a project without violating those state-of-the-art boundaries, we have had to limit the universe in which we work. It would be inappropriate and grandiose to suggest that a universal model of perceptual-motor performance could be developed with current knowledge and technology. Therefore, we have chosen, as a prototypical microworld system to be studied, an underwater SWIMMER capable of acquiring images of food objects, recognizing and discriminating among salient (edible) and irrelevant (inedible) objects, establishing a three-dimensional world model of its environment and the objects in it (including itself), and then demonstrating its understanding by swimming through a turbulent ocean to those objects. Involved in such a simulation must be consideration of visual, localization, interpretative, decision making, and motor functions as well as some challenging new problems of how one integrates all of these functions.

An important technological goal of our work is to incorporate these functions into a single integrated programming structure that operates completely automatically, but is easily modified as new techniques for improving the various transformation stages become available. To this end, we also had to develop an executive operating system environment that permits flexible and convenient construction of alternative theories or models as our understanding evolves.

We discovered that new empirical, practical, and theoretical problems arise when simulating such an integrated ensemble of psychological and/or computational processes. The difficulty of experimenting with such a complex perceptual-motor function is vastly greater than the clean-cut experimental task of examining, for example, isolated psychophysical relationships. Indeed, much of psychophysics’ progress is based on what was called the Method of Detail by John Stuart Mill or the Methode of René Descartes. That is, on techniques that sought, in the ideal, to hold all variables and components of a complex process constant except the one that was intentionally manipulated.

Successful studies of quasi-realistic and less constrained (i.e., in terms of the numbers of variables allowed to covary) complex systems, however, often require more naturalistic and explorative approaches in which many variables of the system’s perceptual input may be difficult to control and its behavioral output difficult to measure. Simulations of the kind explored here provide an alternative and convenient means of studying some of the more complex interactions than might be possible in the conventional laboratory or field situations.

Practically, and here we refer to the practical details of the simulation (i.e., the programming) process itself, researchers concerned with the study of elaborate programing efforts of the kind we describe are confronted with a new set of problems that are not faced when working with carefully isolated components of the total system. For example, how can the output of one process be automatically transformed into the input to the next in such an integrated simulation? This is the problem referred to as automation (a slightly different use of the word than in other contexts), faced by engineers developing working robotic or tele-operator systems. Automation is an especially challenging and unstudied issue because most component-oriented projects can, by definition, ignore the many difficulties involved in communication between the separate program segments.

The same problem of defining the sequencing of information flow also exists for cognitive modelers and theoretical neuroscientists. These more biologically oriented scientists confront virtually identical challenges; for example, how do we explain the neural coding and integration mechanisms by means of which the signals from the receptors are conveyed to and interpreted by subsequent levels of neural processing? It is quite clear that the two groups of researchers—engineers, on the one hand, and psychologists and neurophysiologists, on the other—are often attacking exactly the same problems, although from the differing perspectives of their respective sciences.

Theoretically, a computer model of a complex perceptual-motor action is also a much more intricate mixture of many different challenges than a putative description or explanation of only one step of the overall behavior. Not only does such a model require that determinations be made of the nature of the individual processes that compose the composite behavior, but also how they interact. As we have learned more and more about the codes and activity patterns at the periphery of the nervous system, our attention has necessarily turned to the ways in which the various centers and nuclei operate in the central nervous system. Biological and cognitive hypotheses leap forth from this change in the conceptualization of the relationship between cognitive process and neural structure.

This change in emphasis immediately raises a major philosophical or conceptual issue. We assume implicitly that the complex perceptual-response behavior we simulate is, in biological fact, a composite or concatenation of segmentable and separately assayable components; that is, that our program modules represent mental processes that may be literally characterized as being microgenetic as opposed to a single inseparable process with different aspects from which it may be examined. Surprisingly, some recent, although preliminary, evidence (Posner, Petersen, Fox, & Raichle, 1988) using positron emission tomography (PET) adds a kind of physiological reality to the microgenetic concept by showing separate physiological response for separate cognitive components. Other earlier, less direct arguments for this kind of psychological process can be found in the work of Bachmann (1980, 1992), and of course most experimental psychologists behave as if they accept this point of view when they go about their laboratory research studying putatively isolatable cognitive components. Thus, the idea of microgenetic components (i.e., constituent psychological processes) is a fundamental component of contemporary views of mind. They are also a necessary conceptual, if not physiologically and anatomically instantiated, strategy for psychological research; at the present time our methodology and mathematics simply do not have the power to deal with a totally holistic model of cognition.

A related epistemological point concerns the significance (as opposed to the mechanical details) of computational or mathematical models of cognitive processes. Uttal (1990) has discussed the issue of the meaning of computational models. He argued that there is no way to verify or validate the actual internal neural mechanisms suggested by a model because of constraints that were elucidated by workers in the fields of automata theory (Moore, 1956), chaos theory (Gleick, 1987), and combinatorics (Stockmeyer & Chandra, 1979; Tsotsos, 1990), among many other arguments drawn from psychology, mathematics, physics, and computer science. If this view is sustained, then any simulation or computational model could be, at best, only an analog (i.e., a process description that follows the same course as some internal mechanism) describing, to a more or less successful degree, the series of events unfolding within a process. But, it could never be substantiated or verified as a homolog (i.e., a true and valid reductionistic explanation of the actual internal neural or cognitive mechanisms). It is, thus, not only conceivable but likely that any hypothetical subprocess embodied in a programmed algorithm (as well as the overall system of a model such as the present one) will always be indeterminate with regard to the validity with which it can represent the actual internal structure and logic of the perceiving brain. This is not a startling new revelation, but an idea that has long standing in the history of science. Unfortunately, it is often forgotten in the enthusiasm for a new computational or mathematical theory, particularly in cognitive science or neuroscience.

Another conceptual issue must be frankly faced here. This concerns the ambitiousness of our project. So far our experience on this project confirms that it is clearly infeasible, given the current technological state of the computational modeling art and psychological theory, to produce a broadly competent perceptual-motor model² able to operate in many different microworlds. There is no question that a simulation of a truly intelligent entity that even begins to approach completeness is still beyond our capability. Nevertheless, what we accomplished does support our contention that it is plausible to pursue the development of a model capable of a circumscribed, but integrated, kind of perceptual-motor behavior that goes well beyond the scope of the individual processes themselves. To do so, we built upon the base of currently available specialized algorithms and pyramided the results obtained with several decades of research on hypothetical cognitive components into both a more broadly based theoretical model and a more broadly representative and realistic simulation than has heretofore been presented.

In this book, we report some of the progress made toward such an integrated theory and a simulation of such a well-defined, but strongly constrained, perceptual-motor process. We describe the main steps th...