Resituating cognitive mechanisms within heterarchical networks controlling physiology and behavior

Production vs. control mechanisms

Drawing heavily upon the insights of the theoretical biologist Pattee (1972/2012a, 1973/2012b), Winning and Bechtel (2018) argue for enhancing the vocabulary for describing mechanisms by characterizing parts and operations as constraints. ² The notion of constraint was introduced in classical mechanics (Sklar, 2013) to describe the reduction in degrees of freedom available to components when organized into macro-scale objects. Due to constraints, scientists describing macro-scale objects do not need to calculate the behavior of each component along each of its six axes of motion but can treat macro-scale objects as wholes. The parts of mechanisms are examples of macro-scale objects that can be treated as relatively stable objects in accounts of a mechanism’s operation. As emphasized by Hooker (2013), constraints are not only limiting but also enabling. A pipe not only limits how water flows but also enables it to reach distant locations that it would not otherwise reach. Constraints enable mechanisms to perform their activities and, as argued by Winning (2018), account for the causal powers of mechanisms and their parts. For mechanisms to exhibit their causal powers, they also require sources of energy. The constraints in mechanisms direct the flow of free energy (as well as matter). ³ Elevated water, for example, will flow through a set of pipes as a result of the free energy it possesses due to being elevated and then can be used, for example, to grind grain.

Many of the mechanisms that have provided the primary examples in the writings of the new mechanists such as protein synthesis or muscle contraction, make, transform, degrade, or move components of an organism’s body or entities in its environment. In each of these cases one can characterize the work that is done in taking something apart, putting something together, or moving something about. I refer to these as production mechanisms and distinguish them from control mechanisms. Control mechanisms also engage in physical processes such as synthesizing or degrading molecules, but these activities are not the primary end of control processes. Rather, they are a means to altering the operation of other mechanisms that are directly engaged in carrying out physiological or behavioral processes. ⁴ (In some cases a sequence of control mechanisms operates in succession, with only the last operating on a production mechanism.)

The operation of control mechanisms leads to physiological processes or to behaviors, but they have this effect due to the alterations they make in production mechanisms. Control of production mechanisms is essential because the constraints implemented in them are such that they perform work as long as free energy is available. The same is true of human-built machines. Incorporating an on–off switch provides a means for human users of machines to control them so they only draw upon their energy supply and perform their activities when desired. Likewise, control processes are required so that organisms metabolize glucose when adenosine triphosphate (ATP) levels are low and restrict muscle contraction when energy is needed for other activities. ⁵

To explicate this notion of control, it is necessary to differentiate two types of constraint. Most constraints in human-built machines are fixed. Pipes and wires, for example, restrict energy flows to fixed paths and do not change as the machine operates. But a few are flexible—they can be altered through the performance of work on them. Switches and valves are examples. When a human performs work on an electrical switch, she determines whether current will flow through the circuit. The same is true of switches that redirect its activities. By setting the switch at a rail junction, the human operator determines on which track the train will move. Switches can have multiple states, resulting in different ways in which the machine will operate, and can even take continuous values, such as the accelerator in a car. In all of these cases, control is achieved by altering a flexible constraint, thereby altering how energy flows through the machine and what output is ultimately produced. ⁶

A further ingredient is required for the performance of work on a flexible constraint in a production mechanism to count as controlling the production mechanism—the work on the constraint must be determined in part by semantic information. When humans control machines, they do so in response to information either about the current operation of the machine or about other conditions that are relevant to the operation of the machine. In operating the accelerator or brake pedal of a car, a driver draws upon information (e.g., that the car is moving slower or faster than desired or that the car in front of it is stopping). ⁷ By depressing one of the pedals, the driver initiates a sequence of causal processes leading to a change in the constraints in the car, thereby altering its operation. ⁸ Control of biological mechanisms likewise requires that the alteration in the constraints in the controlled mechanism be initiated by the acquisition of information and a causal process then links this to the flexible constraint in the controlled mechanism.

Like all mechanisms, control mechanisms use energy to perform work. There is, however, a qualitative difference in the work performed by a control mechanism. Control mechanisms typically are distinct from those that provide the material and energy used in production mechanisms. MacKay (1964) differentiates inputs in terms of what the energy is used for—in the control case some of the energy is devoted to altering the structure of the mechanism and not for producing the product. Since the flexible constraints of the control mechanism are themselves determined by measurement processes, and thus, information, MacKay’s characterization of control mechanisms as involving “the action of form upon form rather than of force upon force” (pp. 311–312) is apt. Insofar as they carry information, the internal processes of a control mechanism can also alter that information by, for example, combining information from different sources. Accordingly, one can start to characterize these mechanisms as processing information.

Organizing control mechanisms: Hierarchy vs. heterarchy

The relation between control and production mechanisms suggests a multi-level perspective in which a control mechanism is at a higher level than the mechanism it controls. When control relations are iterated, it is natural to view each controller as being at a higher level, giving rise to a hierarchy. This is how control is commonly understood, both in the sciences (Simon, 1962), including psychology (Broadbent, 1977; Newell, 1990) and neuroscience (Badre, 2008; Grafton & Hamilton, 2007; van Essen & Gallant, 1994), and in social organizations such as the military, businesses, or university administrations (as witnessed by the prevalence of organizational charts). Although not strictly required for a hierarchy, two further features are commonly included in that notion: (a) individual production mechanisms typically fall under the auspices of only one controller at the next higher level and (b) there are fewer controllers at each level in the hierarchy, ultimately topping out at a single highest-level controller. Thus, hierarchical control is generally conceptualized in terms of a pyramid as in Figure 1a. The solid arrows represent control operations whereas the dotted arrows reflect information flowing up to the controller. There may be exceptions in any instance of a control hierarchy, as when a given controlled system is controlled by two higher-level controllers, but these are usually viewed as exceptions and generate risks (e.g., that two higher-level controllers issue incompatible commands). A similar hierarchical conception is often assumed, sometimes only implicitly, in thinking about cognitive tasks. Information flows upwards from the senses through a number of spinal cord and subcortical regions until it is funneled through the thalamus to the neocortex which then carries out high-level cognitive tasks. The hierarchical picture is often maintained within the neocortex, with the posterior cortex treated as processing sensory information and prefrontal cortex acting as the central executive that directs the activities of the organism (Badre, 2008; Badre & Nee, 2018; Botvinick, 2008; Fuster, 2015). Decisions made there are then implemented first in the motor cortex and then through a series of subordinate controllers.

OPEN IN VIEWER

Figure 1.

A. A typical hierarchical pattern of control. Nodes with a P are production mechanisms, those with a C are control mechanisms. Dotted, diamond-headed arrows represent upward flow of information while downward arrows represent directions to lower-level control centers. B. A heterarchical pattern in which some units control those at higher levels, there are more controllers at higher levels than at lower levels, and not all lower-level controllers fall under higher-level control. End-edged arrows represent inhibition, which is important in systems in which components are endogenously active.

Hierarchical control such as that shown in Figure 1a reflects how researchers often first begin to conceptualize control. Even Pattee, who contributed much of the framework needed to understand control mechanisms operating on productive mechanisms, typically characterized control as hierarchical. But in 1991, he adopted the term heterarchy, a term initially introduced by McCulloch (1945) in the context of intransitivity of values (e.g., A is preferred to B, B to C, and C to A). Pattee characterized the heterarchical control model as “a distributed causal network that does not define an order relation or special significance to particular local causal links” (1991, p. 220). My use of heterarchy draws from both McCulloch and Pattee, but differs in an important respect. Since heterarchy incorporates the root archy, which characterizes relations in which some items govern others, the order or level relationship should not be totally abandoned. Control mechanisms in particular govern production mechanisms and, in that sense, are at a higher level. I use the term heterarchy when there is a large-scale violation of the features I associated with hierarchy—a transitive ordering of levels, a limit of one controller for a given controlled mechanism, and fewer controllers at higher levels. All three of these fail in the situation shown in Figure 1b. In some cases, one entity controls another shown at a higher or the same level. Many units are controlled by multiple others. And rather than fewer controllers at higher levels, there are more—there is no chief executive. Although I have tried to maintain a level perspective in representing this set of components, these mechanisms, as Pattee suggests, are better viewed as forming a network.

Hierarchical control doesn’t simply arise. Implementing it in human machines requires prior planning to enable the human user to control the machine. Organisms, however, were not designed. Control mechanisms in organisms arose through the course of evolution. There is good reason for this: biological organisms are autonomous systems in the sense they are far-from-equilibrium systems that construct, maintain, repair, and replicate themselves (Moreno & Mossio, 2014). To accomplish this, they must harness flows of free energy (and materials) and direct them to production mechanisms within that serve primarily to build, maintain, and repair the organism itself or its descendants. Notably, the production mechanisms must also build subsequent production and control mechanisms. Given environments that are variable and threaten to disrupt the relatively fragile set of mechanisms constituting an organism, control is required to ensure that production mechanisms build or repair the mechanisms that constitute the organism as needed. When a new control mechanism enhances this ability, it is likely to be maintained in the lineage. Most such new control mechanisms will be quite specific—they will measure one type of information and use that to carry out one control activity. Over time some control mechanisms may be modified to control yet other mechanisms. Given this manner of creation of control mechanisms, many such control mechanisms could evolve to operate on the same production mechanism or on already existing control mechanisms. There is no force promoting hierarchical organization of control mechanisms.

Multiple control mechanisms, connected in a variety of opportunistic ways, may strike many as untenable. In social systems, such as corporations or armies, it is often assumed that ultimate control should reside in a top-level executive. Otherwise, different control mechanisms may operate in incompatible ways on the production mechanisms. If an organism lacks a central executive, what ensures that the different production mechanisms will work together to enable the organism to maintain itself and reproduce? Without a top-level executive, when there are multiple control mechanisms operating on the same controlled mechanism, integration happens at the local level. If there are competing control inputs, the controlled mechanism responds to them as it can. Yet, there are cases in which conflicts arise. This can be a context in which a new, specific control mechanism might arise to settle the conflict.

An example of such conflict, still within individual cells, appears at the origin of circadian rhythms in cyanobacteria. As a result of evolving a mechanism that performed photosynthesis, which provided needed free energy, cyanobacteria created an environment which contained molecular oxygen. Molecular oxygen, however, is a poison to many enzymes and in particular to nitrogenase, which cyanobacteria rely on to fix nitrogen. Some species of cyanobacteria such as Synechococcus elongatus evolved an oscillator that maintains a period of approximately 24 hours. ⁹ This generates a signal that controls gene expression, resulting in some genes, including those needed for photosynthesis, being expressed during daylight, and other genes, including the one coding for nitrogenase, being expressed at night. Segregating their operation to different periods of the day enabled these cyanobacteria to employ at different times of day two production mechanisms whose activities conflicted (Cohen & Golden, 2015).

In S. elongatus much of the genome is regulated in a circadian fashion so that most genes are expressed only during the day or only at night. But it is important to recognize that the circadian clock is not the only mechanism controlling the bacterium’s physiology or behavior. There are control mechanisms regulating metabolism, phototaxis, blooming, etc. These other control mechanisms detect conditions of importance for controlling production mechanisms under their control (Mullineaux, 2014; Schuergers, Mullineaux, & Wilde, 2017). The circadian clock is neither the highest-level controller nor does it appear to be under the control of other control mechanisms. It is one control mechanism in an ensemble of control mechanisms that operate largely independently but with some interconnections. This illustrates heterarchical organization.

Examples of heterarchically distributed cognitive control in organisms with nervous systems

There are many more examples of heterarchically organized control mechanisms performing cognitive functions in organisms without nervous systems, ¹⁰ but since many cognitive scientists and philosophers of cognitive science assume that neurons are required to engage in truly cognitive activity, I turn to examples of neural-based cognitive control mechanisms in animals that operate independently of the high-level cognitive mechanisms that have been the focus of cognitive science. ¹¹ There does not appear to be a strict hierarchy between these mechanisms; rather, they appear to be organized heterarchically with multiple controllers operating on the same production mechanisms, sometimes modulating other still-operative control mechanisms. The first example involves invertebrates that lack a neocortex; the second involves mammals in which the neocortex is rendered non-functional.

Neural control mechanisms in organisms without a neocortex

Invertebrates, which lack a neocortex, perform a host of information-processing activities that would generally be viewed as cognitive, including integrating sensory information, making decisions, and learning. These cognitive functions play important roles in controlling physiological processes and behavior. Even in those invertebrates with a brain, these functions are often performed by multiple control mechanisms. I focus on one case, decision making in the medicinal leech, Hirudo medicinalis, as it is illuminating as to how multiple control mechanisms figure in deciding what behavior the leech performs. Decisions on whether to swim or crawl are made locally in each of the 21 segments of the leech’s body. This raises the possibility that different ganglia will make different decisions. One might suspect that a higher-level control system, perhaps located in the head unit, would direct the activity of each ganglion. Instead, however, overall coordination is achieved via projections between the ganglia that allow each ganglion to be responsive as well to the decisions made in other ganglia and jointly arrive at a decision. Control is heterarchical, not hierarchical.

Turning inside each ganglion reveals further division of decision making between multiple neurons. The decision-making activity of an organism is most evident when a stimulus is not decisive in selecting one behavior or the other. Briggman, Abarbanel, and Kristan (2005) created such a situation by administering electrical pulses to neurons in an exposed ganglion that mimics a touch to the body wall of the intact leech. In response, a leech is equally likely to swim or crawl. Using electrodes that both stimulate and record, the researchers stimulated neurons and recorded the resulting spike train (eight such spike trains from one neuron are shown in Figure 2a; color figure available only in online version of this article). Spike trains which resulted in swimming (shown in blue) exhibit clear differences from those resulting in crawling (shown in red), with differences manifest as early as 4 seconds after the stimulus.

OPEN IN VIEWER

Figure 2.

A. Spike trains generated from a single neuron on consecutive trials after a short stimulus, the timing of which is indicated by a black box above the left end of the first trial. Motor patterns for swimming (blue) and crawling (red) are easily distinguished. B. Trajectories through a state space defined in terms of the first three principal components in which swim and crawl trials diverge (dark blue and red trajectories indicate mean trajectories). Green shows a trajectory that began like those for swimming then diverged to be more similar to those for crawling. From Briggman, Abarbanel, and Kristan (2005). Reprinted with permission from AAAS.

Since Briggman et al. (2005) were interested in how the leech made the decision between behaviors, they employed a voltage-sensitive dye to image activity from 143 of the approximately 160 neurons on the ventral surface of the ganglion during the first 10 seconds post stimulus. Initially pursuing the common assumption that one decision-making neuron would be responsible for initiating the selected activity in each ganglion, they focused on neurons that showed differential responses during this period and tried to alter the decision by hypo- or hyperpolarizing each of them before the stimulus. In no case were they able to change the decision, leading them to conclude that none of these neurons was the executive decision maker. Briggman et al. then explored an alternative hypothesis that the decision was made by an interacting population of neurons. To determine how this might happen, they applied principal components analysis followed by linear discriminant analysis to the data from all the neurons which they were able to image. When they plotted the first three principal components, they showed that, very early in each trail, cumulative activity from the collection of neurons exhibited a distinctive pattern for swimming (blue) or crawling (red). (The trajectory in green is from an atypical trial that began like a swimming trajectory but shifted to a crawling one.) Briggman et al. identified 17 neurons as contributing significantly to this early activity and, as before, hypo- or hyperpolarized each of them. They found one neuron (208) that, when hypo- or hyperpolarized, would bias the decision. Although this suggests that it was the decision-making neuron, neither hypo- nor hyperpolarization alone would initiate behavior. Only stimulation in combination with the nerve shock elicited activity in the population.

The fact that neuron 208 could not produce behaviors on its own, but only bias decisions as part of a population, sheds light on the control mechanisms responsible for decisions between behaviors. Decision-making is not performed by a single control mechanism but from multiple neurons acting collectively. Neuron 208 is especially influential, which the authors speculate may be due to its connections to the central pattern generator for swimming (a downstream control mechanism). In the context in which a decision was elicited by the nerve pulse, the activity of this neuron biased whether the central pattern generator for swimming would control the required muscle.

This example of a decision-making mechanism in leeches reveals several respects in which neural control is heterarchical. Networks of heterarchically organized neurons in each ganglion execute control over muscles in that segment of the leech. These independent control centers, though, coordinate their behavior so that decisions in individual circuits are modulated by activity in other ganglia. Within each ganglion, multiple neurons contribute to the decision—individual neurons can bias the decision, but they do not make the decision. This collaborative effort then activates or deactivates the central pattern generator, a further control mechanism that directly engages muscles.

Neural control mechanisms in decerebrate and decorticate mammals

In mammals, neocortical control mechanisms play a major role in processing information provided by the senses and influencing behavior and it is easy to impute to them all cognitive processing related to behavior. That they are not the only cognitive control mechanisms operative is made evident by the results of research on various mammals that involved removing the neocortex or transecting the spinal cord at particular locations and examining the control over behavior that the organism could still exhibit. ¹² The term decorticate technically refers to lesioning or sectioning that separates only the neocortex from the rest of the nervous system, although often midbrain and limbic structures are also incapacitated. The term decerebrate is applied when sectioning is done lower in the brain such as at the level of the superior colliculus dorsally and the hypothalamus ventrally.

Classic studies of behavioral capacities that are preserved in decorticate preparations were conducted by Goltz and Sherrington in the late 19th century. Goltz (1892) showed that a loud sound would awaken a decorticate dog and that it would move its ears in response to sounds when awake. The ability to process sensory stimuli and select motor responses was, at least to a degree, retained without the neocortex. From studies employing monkeys and cats, Sherrington (1898) described decerebrate rigidity which ensued after removing higher brain regions or separating them from the brain stem:

The elbow joints do not allow then of the usually easily made passive flexion, the knee joints similarly are stiffly extended. The tail is stiff and straight instead of flexible and drooping. The neck is rigidly extended, the head retracted, and the chin thrown upward. (p. 319)

This condition, which Sherrington showed could persist for days, contrasts with the far more flaccid state of muscles if the spinal cord is transectioned somewhat lower, revealing control exercised by areas in the brain stem. In subsequent research, Sherrington (1906; see also 1917) characterized the behavior of decorticate cats and monkeys in which more subcortical structures were maintained, describing one cat that continued to walk on a moving belt, adjusting its walking speed to that of the belt. ¹³ This reveals the capacity of subcortical structures to perform relatively complex processing of information. Findings of this sort inspired Sherrington’s characterization of the integrative nature of the nervous system; on his view, complex behaviors result from the integration of multiple heterarchically organized control mechanisms distributed throughout the brain.

This line of research was continued by a host of 20th-century researchers who revealed even more complex cognitive control of behavior in animals in which the neocortex and other higher cognitive control structures were not available. Bard and Macht (1958; see also Bignall & Schramm, 1974) demonstrated an extensive behavioral repertoire in mesencephalic and pontine cats over a 5-month period, including spontaneous righting, quadripedal locomotion, running, and climbing. When administered a pain stimulus, the cats exhibited fear and escape responses and behaviors characteristic of rage, including hissing, growling, protrusion of the claws, tail lashing, as well as occasional clawing or biting. Other studies have pointed to the capacity to learn. Poltyrev and Zeliony (1929), for example, reported training two decorticate dogs to lift one forepaw in response to a whistle, the other forepaw in response to a sharp percussive sound, and to make no response to a tone. Buchwald and Brown (1973) review studies showing these organisms also exhibit classical conditioning, as do organisms with only their spinal cord preserved.

These studies make clear that the neocortex is not required for a host of behaviors that require complex information processing, including decision-making and learning, raising the question of what are the mechanisms that process information in subcortical structures. One strategy for addressing this question is to examine the behavior of organisms in which different subcortical areas are preserved or damaged. Such research reveals that, if the striatum, the major input region to the basal ganglia, was spared but the thalamus damaged, cats continued to feed spontaneously, to localize low-intensity auditory stimuli, to associate sounds or positions of a feeding dish with food, to clean and groom themselves, and to exhibit estrous, rage, and fear responses when stimulated accordingly. However, when the striatum was damaged and the thalamus spared, component behaviors could still be elicited but were no longer integrated (Bard & Rioch, 1937; Emmers, Chun, & Wang, 1965), suggesting that the striatum plays a role in integrating behaviors.

Extending this strategy by drawing upon studies transecting a cat brain at different locations as shown by the lines labeled a, b, and c in Figure 3, Whelan (1996) identified elements of a network of subthalamic regions that interact in producing behavior. When transected along line a, a cat is capable of sustained bouts of spontaneous locomotion but when transected at b, an exogenous electrical impulse is required to generate activity. The subthalamic locomotor region (SLR) appears critical for this initiation—if it is specifically lesioned, cats lose the ability to initiate action for a period, although they subsequently recover the ability (Shik & Orlovsky, 1976). Stimulating the SLR in the intact cat causes the cat to start looking around and then begin walking in a manner indistinguishable from spontaneous locomotion (Mori, Sakamoto, Ohta, Takakusaki, & Matsuyama, 1989). Among the areas activated by SLR activity is the mesencephalic locomotor region (MLR) in the brain stem. In cats transected at b or c, stimulation of the MLR can elicit different gait patterns (walking, trotting, etc.). In intact animals, the MLR receives input from many regions, including the substantia nigra (SN in Figure 3), the output region of the basal ganglia, and the nucleus accumbens. Evidence suggests that the SN strongly inhibits the MLR, accounting for the lack of spontaneous activity in cats transected at b. Multiple pathways project from the MLR into the brainstem, including to the medial medullary reticular formation (MRF), the last major point of integration before fibers descend to interneurons in the spinal cord. The MRF can both initiate and regulate the step cycle.

OPEN IN VIEWER

Figure 3.

Locations of transections in the cat brain that result in different behaviors being retained or lost. Reprinted from Whelan (1996), with permission from Elsevier.

This line of research reveals a complex collection of control mechanisms that process information so as to control behavior. Much more detailed analysis is required to specify precisely what type of control each region exercises and how interactions between regions figure in the control of complex behavior. For my purposes, the research discussed is sufficient to reveal that, in the absence of control from the neocortex, animals are capable of selecting and initiating actions, integrating information including sensory feedback, and learning. Subcortical regions perform a variety of cognitive activities that contribute to behavior. There is no reason to doubt that these subcortical control networks remain in place when the neocortex is active and figure in fundamental ways in the control of behavior in normal healthy organisms, including us.

Conclusions and implications for studying mechanisms in cognitive science

Cognitive science historically concentrated on high-level cognitive activities and sought to explain these activities in terms of specific mechanisms. An alternative perspective views cognitive activities as fundamental to living organisms and therefore manifest in organisms that cannot perform high-level cognitive activities. I have advanced a framework that focuses on how production mechanisms responsible for physiological and behavioral activities can be controlled by control mechanisms that operate on constraints within production mechanisms. Control mechanisms can also work on other control mechanisms by altering constraints in them. Insofar as these control mechanisms respond to information, integrate information, generate decisions, and are capable of learning, they fit the characterization of cognitive mechanisms.

An important insight one reaches by starting with low-level control mechanisms is recognition that there is no reason that control must be organized hierarchically. The introduction of control mechanisms to regulate specific aspects of production mechanisms will more likely lead to a heterarchical network of control mechanisms as new control mechanisms are established as needed to control specific activities. One might fear that lacking a central controller that regulates all others would lead to chaos as different control mechanisms direct activity in incompatible ways. The success of organisms that rely on heterarchical networks of control mechanisms suggests the opposite—that such distributed control results in robust performance of activities required to maintain the organism.

The perspective of cognition as involving distributed, heterarchically organized control, is generally compatible with calls for construing cognition as embodied and embedded. In particular, in advocating that cognitive activities are distributed throughout the nervous system (and in individual cells), it supports the need to view cognition as embedded. It differs with most proposals for embedding cognition, though, in that it does not focus on the special contribution of the rest of the body, but on control processes, many of which are carried out in the nervous systems.

The account advanced here is most clearly indebted to the contributions of Pattee, Maturana and Varela, and Moreno. In characterizing biological systems in terms of constraints and control processes, Pattee (1972/2012a, 1973/2012b) offered a perspective that greatly enriches the mechanistic perspective that has emphasized only the entities and activities involved in mechanisms. Maturana and Varela (1980) drew out the complementary perspective that the primary activities of biological organisms are directed at creating and maintaining themselves as distinct entities. Moreno (Bich, Mossio, Ruiz-Mirazo, & Moreno, 2015; Moreno & Mossio, 2014) added the perspective that since organized bodies are far from equilibrium, a critical concern for organisms is controlling flows of free energy through them. Understanding this requires a primary focus on organization at various levels in organisms. Adapting this view to the focus of the new mechanists, which focuses on individual mechanisms, these contributions point to the need to understand mechanisms as situated in far-from-equilibrium systems (organisms) that are organized in a manner that enables them to maintain themselves. Production mechanisms perform the work needed to maintain the system far from equilibrium, but for these to be exercised effectively, control mechanisms are required. A step towards appreciating how these mechanisms are organized is to recognize the heterarchical organization of control mechanisms and the need to focus on these.

Authors in this tradition have made only a few direct attempts to integrate their framework with details about the nervous system. Bickhard’s (2015a, 2015b) interactionist proposal is an exception. Accordingly, I will briefly comment on how the account offered here relates to his. A major focus of Bickhard’s proposal is to defend a conception of representation as arising within the nervous system as it engages in interactions with the rest of the body and the external world. Although I concur that understanding how cognitive processes function as representations is important, developing that has not been central to this paper. Accordingly, I will not develop my specific disagreements with his account of representation, but will only note that two components of the analysis of control systems developed here do provide critical foundations for understanding representations—the fact that control systems serve to control production mechanisms the organism needs to maintain itself in a far-from-equilibrium relation to its environment and that this requires them to make measurements of conditions appropriate to the operation of the mechanism. On my view, measurement gives rise to representations that are used by control mechanisms to regulate physiological processes and behavior.

The account of control mechanisms offered here provides perspective on three other features of Bickhard’s analysis (2015a, 2015b). First, he emphasizes that neural processes manifest oscillations. While this is not universally true, oscillations are widespread and are often utilized for regulating production mechanisms. For example, pattern generators are crucial for coordinating epithelial or muscle responses and hence are fundamental to control systems. Oscillations, however, occur at multiple frequencies, arise as a result of different feedback systems, and play different roles. Bickhard notes the neural activity often modulates these oscillations. But often this is due to specific measurements performed by mechanisms that may not themselves exhibit oscillations—in E. coli there is an oscillator regulating the flagellum, but control is achieved by the phosphorylation state of CheY produced by measurements performed by chemoreceptors. Second, Bickhard treats the nervous system as a totally interconnected system—this is what for him results in oscillatory dynamics. In one sense this is true—overall, the nervous system appears to constitute a small world in which there are short pathways of connectivity between any components. But at the time scale that matters for cognitive activities, it is the specific patterns of connectivity among control components that matter. With respect to oscillations, they arise from specific contexts in which positive and negative feedback loops figure in control systems. Viewing these control mechanisms as parts of a heterarchical network encourages analyzing the specific feedback relations that arise within control mechanisms. Finally, Bickhard invokes the notion of microgenesis to characterize cognitive states and the context in which representations appear. While the dynamics of the control mechanisms commonly create transient metastable states, the specifics depend on the particular case and these matter for the control that is exercised.

The framework viewing heterarchical control mechanisms as cognitive has implications for the current approach to cognitive science. As noted at the outset, cognitive science has historically focused on high-level cognitive activities. In recent decades researchers have focused on subcortical structures such as the basal ganglia, cerebellum, and the thalamus, in large part due to the feedback loops between these structures and regions in the neocortex. These approaches still focus on high-level cognitive tasks. The heterarchical control perspective understands cognitive activities as initially arising much closer to the production mechanisms being controlled (as illustrated in both the research on decision making in the leech and behavioral control in decorticate and decerebrate mammals). The behavior exhibited in decorticate cats when both the basal ganglia and thalamus are preserved shows that these structures are not only interacting with neocortical regions but with control mechanisms responsible for regulating muscle activities lower in the nervous system. On the heterarchical control perspective, higher-cognitive processes do not operate autonomously but through interaction with diverse lower-level control mechanisms. A consequence is that the demands on what high-level control processes must do is reduced as they can rely on activities performed by other components of the heterarchical network. To examine this, cognitive science might expand its inquiries far beyond humans. In addition to decorticate mammals, a focus on control mechanisms in other vertebrates and in invertebrates provides a means to identify ways in which cognitive control is realized in multiple control mechanisms. Given the highly conservative nature of biological mechanisms, including control mechanisms, studying cognitive activities in these organisms can give researchers new insights into the sorts of cognitive control mechanisms that are realized in humans. This does not downplay the importance of higher-level cognitive processes in humans but can provide additional insight into the resources provided by other cognitive control mechanisms with which they work.