1 Autonomy and Selfhood: Incorporating A Bodily Self into Intelligent Systems John Jenkinson 1: Introduction It is seldom disputed that the self is (at least) an embodied entity – that selves are somehow tied to or connected intimately with bodies, but in a way that preserves the ontological distinctness of the two entities. This view appears to be supported by our intuitions about the nature of selfhood: we, as selves, have bodies, but we are not our bodies. Some philosophers such as Maurice Merleau-Ponty, however, have denied this distinction, advocating instead that not only is the self embodied, we are entitled to make the stronger claim that the self is bodily. Wayne Christensen and Cliff Hooker have recently developed an account of the emergence of intelligent systems grounded in the functional organization of autonomous systems. The concept of autonomy can be utilized to generate normative functions in adaptive systems, since the closure constraints it requires produce process requirements for the interdependent processes that constitute the unified system. The system can only remain viable as long as these requirements are met, and thus, function can be determined based on the contributions these processes make to the autonomy of the system. Christensen and Hooker’s account develops out of the concept of directed interaction – a capacity of adaptive systems necessary for closure to be met. More sophisticated forms of directed interaction in more complex adaptive systems develop into self-directed anticipative learning. While Christensen and Hooker remain neutral about the ontology of the self involved in ‘self-directed’ interactions, I will argue that this in fact presupposes, or implicitly incorporates, a bodily self. This is significant, for it bears on contemporary and lively debated issues concerning the nature of the relationship between the self and the body in the philosophy of mind, cognitive science, and

2 phenomenology. My argument will proceed in four steps: First, I will elaborate the concept of an autonomous system in relation to Christensen and Mark Bickhard’s account of biological function, which will be necessary for the second step: understanding how intelligence emerges from autonomous systems. Third, I will articulate an argument for the existence of a bodily self grounded in sensori-motor coherence, as developed by Dorothée Legrand.

Finally, I will

demonstrate that self-directed interaction, central to Christensen and Hooker’s account of intelligent systems, requires the mechanism of action monitoring that grounds sensori-motor coherence, thus necessarily incorporating the presence of a bodily self into their account. 2: Autonomy and the Systems Approach to Biological Function Christensen and Bickhard have developed a theory of normative (biological) function based in the autonomy of dynamic systems, where autonomy is “a basic normative dynamical constraint on the functional organization of adaptive systems” (Christensen and Bickhard 2002, 4). Understanding autonomous systems in relation to normative function will provide the first step in arguing that an approach to the emergence of intelligence based in the functional organization of autonomous systems requires an interpretation of the self as bodily. 2.1: Autonomy Within the context of dynamic systems, autonomy should be understood as selfgovernance rather than complete independence, since autonomous systems are always structurally coupled with the environment within which they are embedded, and as such cannot be completely independent from it. Autonomous systems are special instances of more general dynamic systems, which are collections of related processes that are interrelated in such a way

3 that they constitute a single whole that changes over time.1 In the case of autonomous systems, the constituent processes recursively depend on each other for their generation, and collectively constitute a unified whole that is self-determining with respect to the range of interactions possible between the system and its environment (Thompson 2007, 44).

The constant

interchange between the system and its environment implies that autonomous (biological) systems are far-from-equilibrium. Nonetheless, they are both operationally and organizationally closed; operationally closed with respect to the “reentrant and recurrent dynamics” of the system, and organizationally closed with respect to the system’s unity being defined by the selfreferential nature of the relations obtaining between the processes that constitute the system (Thompson 2007, 45).2 There are two key aspects of autonomy relevant to normative function: (i) the system is cohesive in the sense that it interacts with its environment as a causally integrated whole; and (ii) the conditions required for this cohesion are generated by the system itself through its capacity to perform work (Christensen and Bickhard 2002, 17). Autonomous systems are self-governing, then, in the sense that they interactively generate the conditions required for their existence. 2.2: Normative Function The preceding discussion of autonomy can provide a framework for understanding normative biological function. Christensen and Bickhard argue that autonomous systems are the relevant class for understanding biological function, and, subsequently, that contributions to autonomy are the basic instance of serving a function (which is to be contrasted with having a function) (Christensen and Bickhard 2002, 18). To serve a function means that a constituent of a 1

It is important to keep in mind that a system will be regarded as a “single whole” only relative to some observer. An organism can be regarded as a single whole at one level of description, but at a higher level might be regarded as a constitutive part of a larger system such as an ecosystem. 2 It is important to keep in mind that this form of closure is distinct from any sort of thermodynamic closure, for autonomous systems are in fact thermodynamically open.

4 particular (autonomous) system (i) has a dynamical significance, insofar as a particular contribution to autonomy can impact the continued existence of the system as a whole, which makes a causal difference because autonomous systems are integrated with other systems and environments; and (ii) has normative significance, since a contribution can succeed or fail in contributing to autonomy (Christensen and Bickhard 2002, 18). By contrast, having a function is an etiological question about the presence of a function in a system, which can (under some theories) be explained with reference to the selection history of the trait in question. In this respect, etiological and systems approaches to function are not necessarily conflicting. Instead, to at least a minor degree, they can be seen as complementary, since they explain different aspects of normative function.3 As was intimated in the previous section, process interdependence is crucial in developing a framework for normative function out of the study of autonomous systems. Autonomy is realized through the organizational interdependence of the processes that underlie the unified system. In this way, processes depend on the outputs of other processes within the system, creating the closure necessary for autonomy. Indeed, Christensen and Bickhard claim that “[i]ndividual parts and processes serve normative function within autonomous systems because of the way they satisfy the requirements of other processes within the system” (Christensen and Bickhard 2002, 19). The process interdependence of autonomous systems thus generates process requirements among its constituents, since each process depends on the outputs of some other process within the system. The maintenance of the autonomous system requires that these outputs be generated in order for the system to remain viable, which in turn 3

The degree to which they differ would be based in the significance each position places on their particular theory. For example, the systems approach would see the explanation of serving a function as more fundamental to explanations of normative function. This does not, however, preclude an etiological approach to explanations of having a function, only ones who consider such an approach as more fundamental to normative function than explanations of serving a function.

5 can be seen as generating normative functions for the constitutive process within the system. Some processes will be more vital to the autonomy and continued existence of the system, while others, though contributing to autonomy, will not be necessary for it. For example, the loss of vision in both eyes will compromise the visual system of an organism, and will undoubtedly hinder the overall autonomy of it, but because the visual system is not vital to the overall autonomy of the organism-system, it can still survive absent vision. By contrast, the failure of an organism’s heart will result in the collapse of its entire cardiovascular system, which is a vital system of the organism, and will subsequently cause its death. Both the visual system and the cardiovascular system make contributions to the overall autonomy of the organism-system, but the significance of each subsystem to the autonomy of the organism-system will vary based on the vitality of the subsystem in relation to the larger organism-system. Since the process interdependence of autonomous systems creates process requirements that generate normative functions, dysfunction can be understood in this way as well (Christensen and Hooker 2002, 20-21). A process is dysfunctional simply if these process requirements are not being met. For example, it is a process requirement of the cardiovascular system to pump blood throughout the body of an organism, since various vital subsystems of the organism depend on it. If the heart fails to pump blood, then it is no longer meeting the relevant process requirements, and subsequently may be said to be dysfunctional. Conversely, if the process requirements of a given process are being met (e.g. the heart is indeed pumping blood), then the process is serving its function properly. To summarize, the function of a given process can be determined by its contribution to the overall autonomy of the system within which it is

6 embedded, and this will be determined based on whether it satisfies the requirements of other processes that constitute the system, and hence, that it too depends upon.4 3: Autonomy and the Emergence of Intelligence The concept of autonomy can be extended beyond debates concerning the nature of normative function, to aid in understanding the emergence of intelligent systems by providing the foundations for a naturalistic interpretation of cognition. The preservation of closure drives adaptive autonomous systems (systems capable of responding to – adapting to – environmental or internal changes, in order to preserve the overall autonomy of the system) to direct their interactions (with the environment, and between the processes that constitute them) in such a way that the conditions required for the maintenance of closure, and thus autonomy, are met.5 This capacity for directed interaction is fundamental to all adaptive systems, and provides an important starting point for understanding the emergence of intelligence. 3.1: Directed Interaction and Self-Directed Anticipative Learning Directed interaction is the capacity an adaptive system has for augmenting its interactions based on the changing demands placed on the system for the maintenance of closure and autonomy.

Accordingly, more complex forms of directed interaction enable a “steering

capacity” that develops in the system (speaking here of phylogenetic development) in order for it to ensure that the closure conditions are met. The steering capacity is composed of a cluster of abilities, such as “the ability to dynamically anticipate the interaction process, and the capacity to evaluate interaction using normative signals,” from which primitive forms of cognition can arise

4

This is particularly significant to the question of selfhood, for the maintenance of the autonomy of a dynamic system (i.e. the continued closure of the system) marks the co-emergence of an inside and outside world, creating a “boundary” that begins to mark a primitive distinction between self and environment respectively. 5 For example, by providing the matter required for the preservation of certain vital subsystems of the larger system – food.

7 (Christensen and Hooker 2000, 141). However, the step toward cognition requires not only that the interaction be directed, but that it be self-directed in such a way that goal-directed interaction emerges are a result of the ability to integrate affective and contextual information and to anticipate interaction. The ability to anticipate interaction, which is constituted by abilities such as memory and emulation, distal perception, and imagination, endows a huge benefit on the system, thus contributing to its overall fitness, by expanding the time window available for directed interaction through reducing context-dependency and improving context-sensitivity (Christensen and Hooker 2000, 144). Directed interaction is also affected by a system’s ability to evaluate the success and failure of its performance during interactions. This is accomplished by the ability to interpret affective signals normatively: pain implies the presence of a harmful stimulus detrimental to the autonomy of the system, whereas pleasure indicates the success of an interaction. The ability to tease these normative signals apart from their particular instantiations, and to retain the information gained from such interactions can allow the system to learn behaviors more conducive to meeting its closure constraints (and subsequently contributing to its autonomy). This can allow organism-systems to modify their diet, for example, in ways that are more conducive to their continued survival by choosing prey that is better able to satiate them without causing too much energy loss during the hunt. This subsequently constitutes a form of goalconstruction: when hungry, the predator will actively seek out a specific type of prey as its goal, because of the preference it acquired for that prey through its ability to interpret normative signals. Using anticipative abilities, the system can learn to adapt and refine its behavior to better achieve its goals by becoming more sensitive to smaller variations in context. This sensitivity provides additional affective and contextual information gained through the

8 interaction that can then be integrated to better refine the system’s behavior and adapt it to relevant contexts, thereby increasing its ability to reach its goals (Christensen and Hooker 2000, 147). This process of learning, which Christensen and Hooker term self-directed anticipative learning, establishes a positive feedback loop that continually improves the effectiveness of the system in realizing its goals. This method of learning underlies the development of intelligent systems. But how are we to understand ‘self’ in ‘self-directedness’? It is not apparent whether ‘self’ here should be understood minimally, referring to the “reentrant” nature of the system that helps establish boundaries between the system and its environment, or if ‘self’ should be interpreted in a more ontologically robust sense akin to ‘agent’. Christensen and Hooker remain relatively silent on this topic, but I will argue that their theory of intelligent systems in fact presupposes a more ontologically robust interpretation of ‘self’. In particular, I argue that it incorporates a bodily self. 4: The Bodily Self Legrand has recently argued that the there is a legitimate sense in which we as selves are our bodies; that the self is bodily, rather than just being tied to a body – embodied. I will briefly outline the argument she provides for this claim, but the focus of this section will be to set up the relevant aspects of Legrand’s theory that will help demonstrate the convergence between Legrand’s characterization of the sensori-motor roots of the bodily self and Christensen and Hooker’s account of the emergence of intelligence. As such, I will only detail her argument to the extent that it will be useful to that end, and will more or less take for granted the argument’s success. 4.1 Bodily Self-Consciousness

9 It is a relatively uncontroversial claim within the philosophy of mind and cognitive science that self-consciousness is constitutive of selfhood. That is, the possibility of selfhood only comes about as a result of the accessibility of the self to itself. Without this accessibility, or self-consciousness, there can be no self (Legrand 2006, 90). What needs to be demonstrated by the bodily-self theorist, then, is that there exists bodily self-consciousness that is constitutive of the self, and that the subject of bodily self-consciousness is the bodily self. So: If (1) self-consciousness is constitutive of the self, And if (2) bodily consciousness is a form of self-consciousness, Then (3) bodily self-consciousness is constitutive of (a part of) the self, Which implies that (4) (part of) the body is (part of) the self. Hence the definition of a bodily self. (Legrand 2006, 91) The crucial premise is (2), and as such, it is to the question of bodily self-consciousness that we will now turn. The type of self-consciousness being sought is self-consciousness that is immune to error through misidentification, and is consciousness of oneself as a subject rather than an object. This type of consciousness cannot fail to be identified as belonging to oneself. For example, when one experiences a pain, it does not make sense ask who it is that is having the pain. The pain I experience cannot be misidentified as someone else’s pain. This is precisely because in such cases of self-consciousness the subject of the experience is not identified and as such cannot be misidentified. The identification made is regarding the content of the mental state – the pain that I experienced (Legrand 2006, 92). In this respect, self-consciousness that is immune to error through misidentification can be understood as pre-reflective self-consciousness: reflection only occurs after identification, and, as such, immune self-consciousness, as unidentified, is pre-

10 reflective.6 As such, the aim is to find a form of bodily self-consciousness that presents the body as the subject of the experience in such a way that it is immune in the above sense. One can examine the phenomenology of experience to discover the body as subject. When one touches one hand with another, the experience one has shifts between one’s body as the object of experience – the body as experienced – and as the subject of it as well – the body as experiencing. In this respect, the body can be seen as transparent; one is “looking” through the body to the world (Legrand 2007, 504). Significantly, this form of bodily consciousness appears to meet the criteria for immunity: it can be experienced without being identified as an object, and the non-identified form of bodily consciousness can be combined with some aspect that is identifying (e.g. the pain) (Legrand 2006, 99-100). As such, this form of bodily consciousness must be immune to error.7 Legrand proposes that this form of bodily consciousness constitutes a genuine case of bodily self-consciousness, which, given her argument above, entails the existence of a bodily self. 4.2 The Physiology of the Bodily Self Having argued for the existence of immune bodily self-consciousness that entails a bodily self, Legrand attempts to give a positive naturalistic account of the bodily self by seeking to determine its underlying physiological mechanism. To this end, she points out that the self understood as bodily just is “the body as it is acting and perceiving, that is, the body as the point of convergence of action and perception…at a bodily level, to be pre-reflectively self-conscious means to experience action and perception as coherent” (Legrand 2006, 108). What appears to

6

For there remains the possibility that one could identify oneself as the individual experiencing the pain, despite how trivial such an identification would be, but only after the pain is experienced. 7 Legrand goes on to point out that bodily consciousness is not absolutely immune to error, but is immune under normal circumstances. This point, while significant to her project, will not concern us here, for nothing crucial for our current endeavor turns on it.

11 be necessary for bodily self-consciousness is a match between an intention to act, the executed action, and the sensorial consequences of that action (in this case proprioception and exterioception) (Thompson 2007, 252; Legrand 2007, 514). Agency thus appears to play a crucial role in the development of bodily self-consciousness, for it is only based on the ability to form intentions to act which can be matched with information about the outcomes of such intentions, that such consciousness, and subsequently this sense of self, can arise. A plausible mechanism by which this sensori-motor integration takes place is known as “action monitoring”, which allows one to determine perceptual modifications caused by change in external stimuli, and perceptual modifications that are caused by change in external stimuli as a result of one’s actions. This is achieved by allowing the system to indentify divergences between observed environmental changes and expected environmental changes. For example, the lifting of a hand will involve a motor command issuing the movement of my hand. This command is compared to sensorial reafferences (a stimulation resulting from self-locomotion), and when these two different sets of information are recognized as coherent, the action of handmovement is attributed to the agent by the mechanism. Action monitoring allows consciousness of one’s perceptions and actions to be experienced as coherent, enabling a system to experience its actions as its own actions.

This form of integration necessarily involves the body as

characterized by its perceptual and motor activities, and, as a result, underlies consciousness of a bodily self rather than just an embodied self (Legrand 2006, 111). The picture that emerges is of the self understood “as an integrated system characterized by a matching of sensori-motor information,” hence, a bodily self (Legrand 2006, 111). 5: Autonomy and the Bodily Self

12 The extent to which Christensen and Hooker’s and Legrand’s accounts are complementary is striking. To be clear, the two positions are not simply arguing for the same view using different language; the projects undertaken by each position are disparate, yet importantly interconnected. Indeed, it appears to be the case that Legrand’s account of the bodily self – especially her positive naturalistic account of bodily self-consciousness – can be interpreted as supplementing Christensen and Hooker’s account of the emergence of intelligence by providing an account of selfhood that appears to be prior to the emergence of intelligence, and is necessary for it. Indeed, their account appears to imply the existence of a bodily self despite their silence on the ontology of selfhood. 5.1: Self-Directed Interaction and Action Monitoring Recall that in more complex autonomous systems, directed interaction is improved by the development of a steering mechanism that enables the system to better meet its closure conditions. Further, this mechanism will be self-directed in such a way that it allows the system to anticipate interactions and to integrate affective and contextual information. The emergence of intelligence is supposed to have its roots in this capacity for self-directed interaction. Most importantly, though, advanced forms of directed interaction necessitate an ability on the part of the system to evaluate the success or failure of its performance in interaction. This ability, I argue, presupposes the presence of an action monitoring mechanism, and subsequently, presupposes the presence of a bodily self. Interactions can only be normatively evaluated, and integrated into the system for the sake of self-directed anticipative learning if the actions executed by the system can be attributed to the system, i.e. the system needs to be able to understand itself as the author of its actions in order to be able to learn from the performance of those actions. The mechanism of action

13 monitoring briefly outlined in §4.2 can aid in this endeavor, since it allows the system to distinguish between changes in the world that are and are not caused by the system, thus allowing the system to be able to claim that the actions executed were indeed its actions. Indeed greater sensitivity to affective and contextual information will plausibly come as a result of more fine-tuned action monitoring, since it would make explicit the aspects of the action that were, and were not under the control of the system when executing the action and integrating information about the action’s result. The connection between action monitoring and selfdirected anticipative learning can be made clearer by looking at paradigmatic cases that involve or lack both capacities. In less complicated adaptive systems, environmental change can be detected, but it would be inaccurate to claim that the system can make a differentiation between a real change in the environment and a change that it caused as the result of the execution of some action, or at least in any way that is meaningful. Christensen and Hooker claim that mosquitoes are not selfdirected, specifically because they cannot anticipate interactions, and thus cannot modify their responses context sensitively (Christensen and Hooker 2002, 11). This makes sense given that the ability to modify responses in interactions context sensitively would require a higher order normative signal evaluation, which would require the ability to monitor actions in order to attribute them to the agent. And this coheres with a common intuition that mosquitoes don’t possess a sense of bodily self in the robust sense being discussed (though, it is possible that they might have some sort of primitive “core” self). Conversely, cheetahs are self-directed agents because they can modify their behavior context sensitively, which will require an ability to monitor action. The ability to modify behavior, Christensen and Hooker claim, is aided by the highly integrative nature of cheetah behavior, which incorporates many types of normative

14 signals in order to act context-sensitively (Christensen and Hooker 2002, 11). This further supports the incorporation of a bodily self, for recall that the bodily self just is the integration of sensorial and motor signals through the mechanism of action monitoring in order for them to be experienced as coherent, thus enabling action to be attributed to the agent. 5.2: Naturalized Intelligent Selves It should be noted that this view is entirely compatible with differing degrees of bodily selfhood, just as there are presumably differing degrees of self-directed anticipative learning. Based on the degree to which action can effectively be monitored, there may be a correspondingly strong or weak sense of selfhood. Also, I have remained silent on whether the two positions I have been incorporating are viable. I think that they are, but at this point I will leave the question open for debate. Both positions are relatively recent contributions, so I am sure much debate will be had. The point presently being made is that in order for the emergence of intelligence to be articulated in a way that is naturalistic, it needs to be grounded on theories that, too, are naturalistic.

It appears to be the case that Christensen and Hooker’s theory

presupposes a sense of self, so this presupposition had better also be articulated in a way that is naturalistic. Legrand’s theory of a bodily self does just this, for it explains the relationship between the body and the self in a way that is arguably more naturalistically viable than embodied self theories (i.e. by “identifying” the self with the body, rather than merely explaining the connection abstractly). 6: Conclusion The process of self-directed anticipative learning cannot be understood absent a naturalistically viable account of a self-agent capable of distinguishing the results of its action from natural change.

Because the bodily self is grounded in the mechanism of action

15 monitoring, which is capable of realizing this distinction, entailing that the bodily self is both prior to and in some sense necessary for this capability, Legrand’s account of bodily selfhood effectively supplements Christensen and Hooker’s account of the emergence of intelligence. Obviously, much more work needs to be done in order for the incorporation of the two accounts to be complete. In particular, the bodily-self theory needs to incorporate an account of the phylogenetic emergence of bodily selfhood in order to more explicitly specify the relationship between the emergence of intelligence and the emergence of bodily selfhood, and the relationship between the emergence of bodily selfhood and autonomous systems more generally. This will help determine whether intelligence and bodily selfhood are co-emergent phenomena, or if bodily selfhood is necessarily phylogenetically prior. What should be clear, however, is that there appears to be a strong connection between the emergence of intelligence and the emergence of bodily selfhood. It seems highly probably from the discussion in §5.1 that bodily selfhood is necessary for intelligence (in the form of selfdirected anticipative learning). If this is the case, then we have reason to believe that there is a strong adaptive advantage for organisms that possess a bodily self, since it permits the development of self-directed anticipative learning. This can help us understand the significance of the phylogenetic development of the self, and can aid in determining the biological function of selfhood. My goal, however, has simply been to begin this dialogue by arguing that the emergence of intelligence must incorporate a sense of the self as bodily.

16 Bibliography Christensen, W.D. and M.H. Bickhard (2002): 'The Process Dynamics of Normative Function', Monist, 85 (1): 3-28. Christensen, W.D. and C.A. Hooker (2000): 'Autonomy and the Emergence of Intelligence: Organised Interactive Construction', A. Etxeberria, A. Moreno and J. Umerez (eds.) The Contribution of Artificial Life and the Sciences of Complexity to the Understanding of Autonomous Systems, Communication & Cognition, 17, Special Edition: 133-158. Christensen, W.D. and C.A. Hooker (2002): 'Self-directed agents'. In MacIntosh, J. (Ed.) "Selfdirected agents". In J. MacIntosh (ed.), Naturalism Evolution & Intentionality, Canadian Journal of Philosophy, Special Supplementary Volume 27: 19-52. Legrand, Dorothée (2006). “The Bodily Self: The Sensori-Motor Roots of Pre-Reflective SelfConsciousness.” In Phenomenology and the Cognitive Sciences 5: 89-118. Legrand, Dorothée (2007). “Pre-Reflective Self-Consciousness: On Being Bodily in the World.” Janus Head, 9(2): 493-519. Thompson, Evan (2007). Mind in Life: Biology, Phenomenology, and the Science of Mind. Cambridge: Harvard University Press.