The New Science of the Mind Page 7

  In this way, a process-long multiplication-that seems to require the postulation of internally instantiated mathematical symbols can be reduced to an internal process of pattern recognition and completion coupled with a process of manipulation of external mathematical structures. The entire process whereby this mathematical task is achieved is, therefore, a hybrid that straddles elements both internal and external to the neural network. Neither of these elements involves the postulation of internally instantiated mathematical symbols, and neither requires postulation of internally instantiated rules governing the transformation of such symbols. The role that is thought to be played by internal symbols and rules has largely been usurped by external symbols and rules, plus the ability to manipulate these symbols as required. The internal residue consists of nonsymbolic pattern recognition and completion operations.

  This strategy, in common with the accounts of perception and memory we have so far examined, works by off-loading some of the cognitive burden of a task onto the world. Pattern-mapping operations seem to place relatively low demands on the human cognitive system in the sense that they seem to be operations that can be implemented fairly easily in such a system. Thus, once again, we find the replacement of a certain type of internal representation-mathematical and other sorts of formal symbolwith an inner process that the human brain finds far easier to implement. And any explanatory slack engendered by this replacement is taken up by the organism's ability to act on the world around it. This attenuation of the role of internal representation coupled with an augmentation of the role of action is a hallmark of the prospective new science.

  One way of thinking of the development of situated robotics is as a more general development of this basic idea. This development is organized around a distinction between what Clark (1989) calls horizontal and vertical microworlds. A microworld is a restricted domain of study. Given that we cannot understand intelligence all at once, we need an approach that breaks it down into easily digestible slices. However, this can be done in two ways. A vertical microworld takes a small piece of human-level cognitive competence as an object of study. Thus, famous research programs in Cartesian cognitive science have focused on things like the production of the past-tense forms of English verbs, playing chess, and planning a picnic. There are several worries engendered by this approach. Perhaps most significantly, in solving problems of this sort-in modeling vertical microworlds-we might find ourselves (unwittingly) employing neat, design-oriented solutions-ones quite unlike biological solutions driven by the need to make use of extant structures. As Clark (1997, 13) puts it: we may be chess masters courtesy of pattern-recognition skills selected to recognize mates, food, and predators. But if we were looking for a neat, design-oriented solution to modeling chess-playing competence, pattern-mapping operations are far from the first place we would look.

  A horizontal microworld, on the other hand, consists in the complete behavioral competence of a relatively simple creature (whether real or imaginary). The idea is that by focusing on these horizontal microworlds we can simplify the problems posed for human-level intelligence but without forgetting biological imperatives (e.g., making use of extant structures and solutions, coping with damage, real-time response, the integration of sensory and motor functions, and so on). The most influential work in robotics and artificial intelligence of the past two decades has focused on the understanding of horizontal microworlds.

  This is where something very like neural networks enters the picture. Many horizontal microworlds are susceptible to modeling by a type of architecture that is similar, in crucial ways, to a neural network: a subsumption architecture. Care must be taken here, because one of the most influential advocates of the use of subsumption architectures is careful to distinguish them from neural networks. Rodney Brooks (1991, 147) writes:

  Neural networks is the parent discipline of which connectionism is a recent incarnation. Workers in neural networks claim that there is some biological significance to their network nodes, as models of neurons. Most of the, models seem wildly implausible given the paucity of modeled connections relative to the thousands found in real neurons. We claim no biological significance in our choice of finite state machines as network nodes.3

  Given their simplicity, the claim of biological significance for neural networks was, of course, always a tenuous one. But this acknowledgment should not blind us to the important similarities between subsumption architectures and neural networks. A subsumption architecture is, like a neural network, composed of "layers" of circuitry, and to the extent these layers communicate it is not through the transmission of complex messages but, rather, by way of simple signals that turn one layer off or on when another layer has entered a certain state. Whether we regard subsumption architectures as a species of neural network is largely a matter of stipulation. But even if, like Brooks, we are not inclined to do so, we should be aware of the deep similarities between the two.

  Brooks has developed an array of simple "creatures," composed of a number of distinct activity-producing subsystems or "layers." Each of these layers-and this is characteristic of subsumption architectures-is a complete route from input to action. Layers communicate with each other by way of simple signals that either augment (i.e., excite) activity in another connected layer, interrupt that activity, or override it. (The expression "subsumption" architecture derives from the fact that a layer can, in these ways, subsume another layer's activity but cannot communicate in more complex ways.)

  Designed along these lines, a creature might be composed of layers of the following sorts (Brooks 1991, 156):

  Layer 1 Object avoidance via a ring of ultrasonic sonar sensors. These cause the creature (a mobot-mobile robot) to halt if an object is dead ahead and allow reorientation in an unblocked direction.

  Layer 2 If the object-avoidance layer is currently inactive, an onboard device can generate random course headings so the creature wanders.

  Layer 3 This layer can override the wander layer, and instead set up a distant goal to take the creature into a new locale.

  Layers can be added incrementally, and each new layer yields a new complete functional competence. These creatures do not require a central planner, processing unit, or reservoir of data. Instead, they are driven by a collection of competing behaviors; and these competing behaviors are themselves driven by environmental inputs. In such a creature, there is no precise dividing line between perception and cognition. Nor is there any point at which perceptual codes are translated into a central code to be shared by onboard reasoning devices. Rather, the environment itself guides the creature, courtesy of some basic behavioral responses, to success.

  Subsumption architectures decompose systems in a way quite different from that employed by traditional Cartesian cognitive science. The latter decomposition is based on the identification of vertical microworlds. A subsumption architecture, on the other hand, decomposes systems not by local (i.e., vertical) functions or faculties, but rather by global (i.e., horizontal) activities or tasks. Thus Brooks writes:

  [This] alternative decomposition makes no distinction between peripheral systems, such as vision, and central systems. Rather, the fundamental slicing up of an intelligent system is in the orthogonal direction dividing it into activity producing subsystems. Each activity or behavior producing system individually connects sensing to action. We refer to an activity producing system as a layer. An activity is a pattern of interactions with the world. Another name for our activities might well be skill. (Brooks 1991, 146)

  The crucial idea is that the activity performed by layers presupposes, and makes no sense without, the system's environment. The activity of each layer consists in close interaction with the environment and can only be understood in terms thereof. So the structures of the respective aspects of the environment are at least as important as the structures of the internal portions of the corresponding layers in rendering the different activities intelligible.

  Again, we find the pattern of expl
anation characteristic of the cluster of theories that vie to make up the new science: the attenuation of representation combined with the augmentation of action. Traditional approaches would attempt to build into the creature or mobot a set of representations and series of rules that governed what to do with those representations. The approach employed by Brooks and others eschews this. Instead, mobots are supplied with subsumption architectures, each layer of which is a functionally complete neural network. And any explanatory slack with respect to the capabilities of the mobot is explained in terms of the mobot's ability to utilize structures in its environment in the right sort of way. With this strategy, no role is assigned to representations in anything like the traditional sense. There is no need for such an assignment. As Brooks famously put it: the world is its own best representation.

  6 Conclusion

  This survey of the cluster of theories that make up the new science is neither comprehensive nor complete. I have, for example, left out any discussion of dynamicist approaches to cognition (van Gelder 1994), and the important accounts of situated child development based on these approaches (Thelen and Smith 1994). However, the survey has been comprehensive enough to serve its purposes. These purposes are twofold. First, enough has been said to show that there is a significant body of scientific research being conducted into the nature of mental processes that diverges in crucial respects from traditional, Cartesian approaches based on rules and representations. Second, enough has been said to identify, at least in broad strokes, the principal points of divergence between the traditional rules and representations approach and this cluster of alternative theories. The cluster of theories-the new and the not so new-is characterized, relative to the traditional approach, by attenuation of the role of representation combined with augmentation of the role of action. The two are connected in that the attenuation of the role of representation is made possible only by the augmentation of the role of action: any explanatory gap engendered by the attenuation of representation is filled by the augmentation of action.

  It is the augmentation of action that makes this cluster of theories antiCartesian-at least ostensibly. Underlying this augmentation is a vision of cognition, and perhaps other mental processes, not as something occurring exclusively inside the brains of organisms but as something that organisms achieve, in part, because of what they do in and to the world that is outside their brains-whether their bodies or the wider environment.

  None of this is, as yet, particularly clear. The function of the next chapter is to inject at least some of the required clarity.

  1 Cartesian Cognitive Science (A Recapitulation)

  Traditional cognitive science has been a continuation of the Cartesian vision of mental states and processes as located exclusively inside the head of any given subject, person, or organism. Cartesian cognitive science developed this vision by way of a theoretical apparatus of mental representations and operations performed on those representations. That is:

  1. Cognitive processes consist in the manipulation and transformation of structures that carry information about the world.

  2. These information-bearing structures are known as mental representations.

  3. Mental representations are structures located in the brains of cognizing organisms.

  Mental representations are typically regarded either as brain states or higher-order functional properties realized by brain states. Since, on either interpretation, mental representations are things that are to be found in the brain, and only in the brain, their manipulation and transformation are also processes that occur in the brain. Since cognitive processes simply are these manipulative and transformative operations, cognitive processes are, therefore, ones that occur in the brain: cognitive processes are brain processes or are processes exclusively realized by brain processes. The manipulation and transformation of mental representations is, of course, not random, but takes place according to certain principles or rules. Therefore, this view is sometimes called the rules and representations approach toward understanding cognition. Both the representations and the rules by which they are manipulated and transformed have an implementation that is exclusively neural: they occur inside the brains of cognizing organisms.

  Different ways of developing cognitive science can, in varying ways and to varying extents, abandon this rules and representations conception of cognition. Neural network, or connectionist, approaches, for example, are commonly thought of as abandoning the rules component of this picture, at least as this had hitherto been understood. Some-both advocates and detractors-even take them as abandoning the representations component also; but I think a more plausible way of understanding neural networks is as modifying, rather than abandoning, the guiding concept of representations.

  For our purposes, however, these differences between classical and connectionist approaches to cognition are unimportant. For they share a common assumption; and it is this assumption that is going to provide the target for the central arguments of this book. Even if we abandon rules and representations for spreading patterns of activation over neural assemblies, we are still dealing with structures and processes that are internal to the brain. Whatever else is true of cognitive processes, they are processes occurring inside the brains of cognizing organisms. That is something on which both classical approaches to cognition and the connectionist alternative can agree. Non-Cartesian cognitive science is defined by its rejection of this common assumption.

  However, as we saw in the opening chapter, the new science is made up of different strands. And even if they all do reject the assumption that cognitive processes always occur inside brains-and at the end of chapter 1, we started to unearth some reasons for doubting this-they do so in quite different ways and for quite different reasons. It is to an examination of the ideas of embodiment, embeddedness, enactedness, and extendedness that we now turn.

  2 The Mind Embodied

  To begin with, recall my admonition in the opening chapter concerning talk of the mind. There is a persistent tendency in most of us to think of the mind as something that lies behind and holds together our various mental states and processes; to think of it as something to which all those states and processes belong. If we think of the mind in this way, then cognitive science-in both Cartesian and non-Cartesian forms-is not a science of the mind but of mental processes. Of course, if we think of the mind as nothing more than a network of mental states and processes, things are different. On this, what is often known as the Humean view of the mind-for Hume might have held a view similar to it-cognitive science would indeed be a science of the mind.' However, in the absence of any consensus on whether or not the mind is in this sense Humean, I shall continue to regard the new science of the mind as, fundamentally, a science of mental states and processes. Therefore, what I am going to call the thesis of the embodied mind is more accurately rendered the thesis of embodied mental processes. This is a bit of a mouthful, so I shall continue talking of the thesis of the embodied mind; but it should be understood that it is a thesis concerning mental processes and not the mind as this is perhaps commonly understood.

  According to this thesis, at least some-not all by any means, but some-mental processes are constituted not just by brain processes but by a combination of these and wider bodily structures and processes. This thesis has been defended by Shapiro (2004) and Damasio (1994), among others. Here, I shall focus on the contribution of Shapiro since I think this is more germane to the concerns of this section-that is, providing a conceptual foundation for understanding the claims of the thesis of the embodied mind.

  Shapiro (2004) provides a sustained attack on what he calls the separability thesis (ST). According to ST, minds make no essential demands on bodies. A humanlike mind could very well exist in a nonhumanlike body. Against this, Shapiro defends what he calls the embodied mind thesis (EMT). According to EMT, "minds profoundly reflect the bodies in which they are contained," and, therefore, "it is often possible to predict properties of the body based on knowl
edge of properties of the mind" (Shapiro 2004, 174). In essence, Shapiro's arguments for EMT are based on the idea that

  psychological processes are incomplete without the body's contributions. Vision for human beings is a process that includes features of the human body.... Perceptual processes include and depend on bodily structures. This means that a description of various perceptual capacities cannot maintain body-neutrality and it also means that an organism with a non-human body will have non-human visual and auditory psychologies. (Ibid., 190)

  For example, in processing visual-depth information, the brain deploys disparity information from two eyes. Were there more than two eyes or fewer, or if the distance between the eyes differed, the processes in the brain that compute depth from disparity would require significant revision: "Human vision requires a human body" (ibid., 191). The same is true of other perceptual abilities. The human auditory system is calibrated to the distance between human ears. Because of this distance, sound will reach each ear at slightly different times, and the difference carries important information about the direction from which the sound is emanating. So, the brain uses the distance between the ears as a way of determining the direction of the sound source. The brain is thus calibrated to this distance: if you change the distance, you would also have to change the calibration of the brain. Moreover, the fact that sound passes through a head of a particular size and density provides further important information about the direction of the sound source, and so on.

  Here is an analogy employed by Shapiro in developing the EMT thesis (ibid., 185ff.). Imagine an instruction manual for piloting a submarine. This manual is, of course, specific to submarines. Although there might be some overlap, it would be largely useless for teaching you how to pilot an airplane. There is no general-purpose manual-or program-that affords skill-transferability from piloting a submarine to an airplane to a bus, and so on. The submarine manual, Shapiro argues, is analogous to the computational algorithms underwriting mental activity. That is, in order to work properly, the rules on which the brain runs-the rules by way of which it manipulates and transforms mental representations-depend on the nature of the underlying architecture that implements or realizes them. And this architecture, Shapiro argues, includes not just the brain but wider bodily structures. The rules by way of which the brain manipulates and transforms mental representations depend, partly but crucially, on the body in which that brain is embodied. Indeed, Shapiro argues that, in one crucial respect, the analogy is not deep enough: "the presence of the sub does not change the information that the instructions contain. Destroy the submarine and the instructions in the manual remain unchanged" (ibid., 186). However, the instruction manual for human cognition would make no sense in the absence of the bodily structures in which, in part, they are implemented. Shapiro characterizes his argument as an attack on the idea of body neutrality. This is the idea that "characteristics of bodies make no difference to the kind of mind one possesses," and this is in turn associated with the idea that the "mind is a program that can be characterized in abstraction from the kind of body/brain that realizes it" (ibid., 175).