5.11: Levels of Embodied Cognitive Science

Classical cognitive scientists investigate cognitive phenomena at multiple levels (Dawson, 1998; Marr, 1982; Pylyshyn, 1984). Their materialism commits them to exploring issues concerning implementation and architecture. Their view that the mind is a symbol manipulator leads them to seek the algorithms responsible for solving cognitive information problems. Their commitment to logicism and rationality has them deriving formal, mathematical, or logical proofs concerning the capabilities of cognitive systems.

Embodied cognitive science can also be characterized as adopting these same multiple levels of investigation. Of course, this is not to say that there are not also interesting technical differences between the levels of investigation that guide embodied cognitive science and those that characterize classical cognitive science.

By definition, embodied cognitive science is committed to providing implementational accounts. Embodied cognitive science is an explicit reaction against Cartesian dualism and its modern descendant, methodological solipsism. In its emphasis on environments and embodied agents, embodied cognitive science is easily as materialist as the classical approach. Some of the more radical positions in embodied cognitive science, such as the myth of the self (Metzinger, 2009) or the abandonment of representation (Chemero, 2009), imply that implementational accounts may be even more critical for the embodied approach than is the case for classical researchers.

However, even though embodied cognitive science shares the implementational level of analysis with classical cognitive science, this does not mean that it interprets implementational evidence in the same way. For instance, consider single cell recordings from visual neurons. Classical cognitive science, with its emphasis on the creation of internal models of the world, views such data as providing evidence about what kinds of visual features are detected, to be later combined into more complex representations of objects (Livingstone & Hubel, 1988). In contrast, embodied cognitive scientists see visual neurons as being involved not in modelling, but instead in controlling action. As a result, single cell recordings are more likely to be interpreted in the context of ideas such as the affordances of ecological perception (Gibson, 1966, 1979; Noë, 2004). “Our brain does not simply register a chair, a teacup, an apple; it immediately represents the seen object as what I could do with it—as an affordance, a set of possible behaviors” (Metzinger, 2009, p. 167). In short, while embodied and classical cognitive scientists seek implementational evidence, they are likely to interpret it very differently.

The materialism of embodied cognitive science leads naturally to proposals of functional architectures. An architecture is a set of primitives, a physically grounded toolbox of core processes, from which cognitive phenomena emerge. Explicit statements of primitive processes are easily found in embodied cognitive science. For example, it is common to see subsumption architectures explicitly laid out in accounts of behaviour-based robots (Breazeal, 2002; Brooks, 1999, 2002; Kube & Bonabeau, 2000; Scassellati, 2002).

Of course, the primitive components of a typical subsumption architecture are designed to mediate actions on the world, not to aid in the creation of models of it. As a result, the assumptions underlying embodied cognitive science’s primitive sense-act cycles are quite different from those underlying classical cognitive science’s primitive sense-think-act processing.

As well, embodied cognitive science’s emphasis on the fundamental role of an agent’s environment can lead to architectural specifications that can dramatically differ from those found in classical cognitive science. For instance, a core aspect of an architecture is control—the mechanisms that choose which primitive operation or operations to execute at any given time. Typical classical architectures will internalize control; for example, the central executive in models of working memory (Baddeley, 1986). In contrast, in embodied cognitive science an agent’s environment is critical to control; for example, in architectures that exploit stigmergy (Downing & Jeanne, 1988; Holland & Melhuish, 1999; Karsai, 1999; Susi & Ziemke, 2001; Theraulaz & Bonabeau, 1999). This suggests that the notion of the extended mind is really one of an extended architecture; control of processing can reside outside of an agent.

When embodied cognitive scientists posit an architectural role for the environment, as is required in the notion of stigmergic control, this means that an agent’s physical body must also be a critical component of an embodied architecture. One reason for this is that from the embodied perspective, an environment cannot be defined in the absence of an agent’s body, as in proposing affordances (Gibson, 1979). A second reason for this is that if an embodied architecture defines sense-act primitives, then the available actions that are available are constrained by the nature of an agent’s embodiment. A third reason for this is that some environments are explicitly defined, at least in part, by bodies. For instance, the social environment for a sociable robot such as Kismet (Breazeal, 2002) includes its moveable ears, eyebrows, lips, eyelids, and head, because it manipulates these bodily components to coordinate its social interactions with others.

Even though an agent’s body can be part of an embodied architecture does not mean that this architecture is not functional. The key elements of Kismet’s expressive features are shape and movement; the fact that Kismet is not flesh is irrelevant because its facial features are defined in terms of their function.

In the robotic moment, what you are made of—silicon, metal, flesh—pales in comparison with how you behave. In any given circumstance, some people and some robots are competent and some not. Like people, any particular robot needs to be judged on its own merits. (Turkle, 2011, p. 94)

That an agent’s body can be part of a functional architecture is an idea that is foreign to classical cognitive science. It also leads to an architectural complication that may be unique to embodied cognitive science. Humans have no trouble relating to, and accepting, sociable robots that are obviously toy creatures, such as Kismet or the robot dog Aibo (Turkle, 2011). In general, as the appearance and behavior of such robots becomes more lifelike, their acceptance will increase.

However, as robots become closer in resemblance to humans, they produce a reaction called the uncanny valley (MacDorman & Ishiguro, 2006; Mori, 1970). The uncanny valley is seen in a graph that plots human acceptance of robots as a function of robot appearance. The uncanny valley is the part of the graph in which acceptance, which has been steadily growing as appearance grows more lifelike, suddenly plummets when a robot’s appearance is “almost human”—that is, when it is realistically human, but can still be differentiated from biological humans. The uncanny valley is illustrated in the work of roboticist Hiroshi Ishiguro, who,

built androids that reproduced himself, his wife, and his five-year old daughter. The daughter’s first reaction when she saw her android clone was to flee. She refused to go near it and would no longer visit her father’s laboratory. (Turkle, 2011, p. 128)

Producing an adequate architectural component—a body that avoids the uncanny valley—is a distinctive challenge for embodied cognitive scientists who ply their trade using humanoid robots.

In embodied cognitive science, functional architectures lead to algorithmic explorations. We saw that when classical cognitive science conducts such explorations, it uses reverse engineering to attempt to infer the program that an information processor uses to solve an information processing problem. In classical cognitive science, algorithmic investigations almost always involve observing behaviour, often at a fine level of detail. Such behavioral observations are the source of relative complexity evidence, intermediate state evidence, and error evidence, which are used to place constraints on inferred algorithms.

Algorithmic investigations in classical cognitive science are almost exclusively focused on unseen, internal processes. Classical cognitive scientists use behavioral observations to uncover the algorithms hidden within the “black box” of an agent. Embodied cognitive science does not share this exclusive focus, because it attributes some behavioral complexities to environmental influences. Apart from this important difference, though, algorithmic investigations—specifically in the form of behavioral observations—are central to the embodied approach. Descriptions of behavior are the primary product of forward engineering; examples in behavior-based robotics span the literature from time lapse photographs of Tortoise trajectories (Grey Walter, 1963) to modern reports of how, over time, robots sort or rearrange objects in an enclosure (Holland & Melhuish, 1999; Melhuish et al., 2006; Scholes et al., 2004; Wilson et al., 2004). At the heart of such behavioral accounts is acceptance of Simon’s (1969) parable of the ant. The embodied approach cannot understand an architecture by examining its inert components. It must see what emerges when this architecture is embodied in, situated in, and interacting with an environment.

When embodied cognitive science moves beyond behavior-based robotics, it relies on some sorts of behavioral observations that are not employed as frequently in classical cognitive science. For example, many embodied cognitive scientists exhort the phenomenological study of cognition (Gallagher, 2005; Gibbs, 2006; Thompson, 2007; Varela, Thompson, & Rosch, 1991). Phenomenology explores how people experience their world and examines how the world is meaningful to us via our experience (Brentano, 1995; Husserl, 1965; Merleau-Ponty, 1962).

Just as enactive theories of perception (Noë, 2004) can be viewed as being inspired by Gibson’s (1979) ecological account of perception, phenomenological studies within embodied cognitive science (Varela, Thompson, & Rosch, 1991) are inspired by the philosophy of Maurice Merleau-Ponty (1962). Merleau-Ponty rejected the Cartesian separation between world and mind: “Truth does not ‘inhabit’ only ‘the inner man,’ or more accurately, there is no inner man, man is in the world, and only in the world does he know himself” (p. xii). Merleau-Ponty strove to replace this Cartesian view with one that relied upon embodiment. “We shall need to reawaken our experience of the world as it appears to us in so far as we are in the world through our body, and in so far as we perceive the world with our body” (p. 239).

Phenomenology with modern embodied cognitive science is a call to further pursue Merleau-Ponty’s embodied approach.

What we are suggesting is a change in the nature of reflection from an abstract, disembodied activity to an embodied (mindful), open-ended reflection. By embodied, we mean reflection in which body and mind have been brought together. (Varela, Thompson, & Rosch, 1991, p. 27) However, seeking evidence from such reflection is not necessarily straightforward (Gallagher, 2005). For instance, while Gallagher acknowledges that the body is critical in its shaping of cognition, he also notes that many aspects of our bodily interaction with the world are not available to consciousness and are therefore difficult to study phenomenologically.

Embodied cognitive science’s interest in phenomenology is an example of a reaction against the formal, disembodied view of the mind that classical cognitive science has inherited from Descartes (Devlin, 1996). Does this imply, then, that embodied cognitive scientists do not engage in the formal analyses that characterize the computational level of analysis? No. Following the tradition established by cybernetics (Ashby, 1956; Wiener, 1948), which made extensive use of mathematics to describe feedback relations between physical systems and their environments, embodied cognitive scientists too are engaged in computational investigations. Again, though, these investigations deviate from those conducted within classical cognitive science. Classical cognitive science used formal methods to develop proofs about what information processing problem was being solved by a system (Marr, 1982), with the notion of “information processing problem” placed in the context of rule-governed symbol manipulation. Embodied cognitive science operates in a very different context, because it has a different notion of information processing. In this new context, cognition is not modeling or planning, but is instead coordinating action (Clark, 1997).

When cognition is placed in the context of coordinating action, one key element that must be captured by formal analyses is that actions unfold in time. It has been argued that computational analyses conducted by classical researchers fail to incorporate the temporal element (Port & van Gelder, 1995a): “Representations are static structures of discrete symbols. Cognitive operations are transformations from one static symbol structure to the next. These transformations are discrete, effectively instantaneous, and sequential” (p. 1). As such, classical analyses are deemed by some to be inadequate. When embodied cognitive scientists explore the computational level, they do so with a different formalism, called dynamical systems theory (Clark, 1997; Port & van Gelder, 1995b; Shapiro, 2011).

Dynamical systems theory is a mathematical formalism that describes how systems change over time. In this formalism, at any given time a system is described as being in a state. A state is a set of variables to which values are assigned. The variables define all of the components of the system, and the values assigned to these variables describe the characteristics of these components (e.g., their features) at a particular time. At any moment of time, the values of its components provide the position of the system in a state space. That is, any state of a system is a point in a multidimensional space, and the values of the system’s variables provide the coordinates of that point.

The temporal dynamics of a system describe how its characteristics change over time. These changes are captured as a path or trajectory through state space. Dynamical systems theory provides a mathematical description of such trajectories, usually in the form of differential equations. Its utility was illustrated in Randall Beer’s (2003) analysis of an agent that learns to categorize objects, of circuits for associative learning (Phattanasri, Chiel, & Beer, 2007), and of a walking leg controlled by a neural mechanism (Beer, 2010).

While dynamical systems theory provides a medium in which embodied cognitive scientists can conduct computational analyses, it is also intimidating and difficult. “A common criticism of dynamical approaches to cognition is that they are practically intractable except in the simplest cases” (Shapiro, 2011, pp. 127– 128). This was exactly the situation that led Ashby (1956, 1960) to study feedback between multiple devices synthetically, by constructing the Homeostat. This does not mean, however, that computational analyses are impossible or fruitless. On the contrary, it is possible that such analyses can co-operate with the synthetic exploration of models in an attempt to advance both formal and behavioral investigations (Dawson, 2004; Dawson, Dupuis, & Wilson, 2010).

In the preceding paragraphs we presented an argument that embodied cognitive scientists study cognition at the same multiple levels of investigation that characterize classical cognitive science. Also acknowledged is that embodied cognitive scientists are likely to view each of these levels slightly differently than their classical counterparts. Ultimately, that embodied cognitive science explores cognition at these different levels of analysis also implies that embodied cognitive scientists are also committed to the notion of validating their theories by seeking strong equivalence. It stands to reason that the validity of a theory created within embodied cognitive science would be best established by showing that this theory is supported at all of the different levels of investigation.