The Cognitive Body: From Dynamic Modulation to Anticipation

Starting from the situated and embodied perspective on the study of biological cognition as a source of inspiration, this paper programmatically outlines a path towards an experimental exploration of the cognitive role of the body in artificial agents. Biological cognition is here conceived and synthetically analyzed as a broadly extended and distributed dynamic process emerging from the interplay between body, environment and nervous system. Accordingly, we first analyze a minimalist case study where the ‘body’, through a very simple non-neural internal bio-regulatory system (an ‘energy level’), acts as a self-organized dynamic action selection mechanism. It modulates the activity of the neurocontroller as appropriate to the current context. The availability of the slower non-neural internal dynamic boosts the cognitive potential of the system, constituted of simple reactive components, providing it with the ability to integrate information over time. Then we examine the intrinsic anticipatory potential of viable dynamic attractors. Finally, we propose a new minimally cognitive architecture, currently under development for experimental investigation, where an explicit model for dynamic anticipation might be coherently exploited via bodily mediation. I. THE COGNITIVE BODY What is the role of the body of an organism? The obvious, common sense-answer, would highlight its role in dictating the physical relation between an agent and its environment. Of course the fundamental function that the body of any organism plays, conveniently situating in a spatio-temporal frame of reference its available set of sensors and actuators, cannot be missed or ignored. However, this common-sense interpretation, far from rejected, is broadened in embodied and situated approaches to the study of cognition, both at a theoretical and experimental level [34], [8], [7], [37], [32]. The body shapes the cognitive potential of the agent by specifying the conditions for possible interactions with its environment, whilst constraining their range. A self-organized agent typically depends on and exploits such constraints [21]. The basic idea of a highly systemic approach to the study of cognition was already centrally present in the work of early cyberneticists (i.e. [35], [1]), gestalt psychology [17] and ecological psychology [11]. The sudden rise of cognitive science cast a shadow on such historically prominent intellectual work. This should not come as a surprise: apparently, large bodies of sometimes outstanding scientific knowledge are destined to be reconsidered, or even completely rediscovered, over and over, whenever there is an intellectual need for them. Presently massive research efforts investigate This work was supported by a European Commission grant to the FP6 project “Integrating Cognition, Emotion and Autonomy” (ICEA, IST027819, www.iceaproject.eu) as part of the European Cognitive Systems initiative. The author is based at University of Skövde, School of Humanities and Informatics, SE-541 28 Skövde, Sweden (e-mail: [email protected]). the problem of understanding cognition by a systematic decomposition. In the scientific tradition, reductionism proved powerfully effective in producing sound explanations (with predictive power) of natural phenomena. Nevertheless, it is wrong to infer that all explanations are reductionistic. Such a misconception might be particularly pernicious in an epoch where the scientific community masters and has large availability of the necessary technology to explore the problem of non-linear complexity, and is intellectually engaged in developing the appropriate mathematics to start addressing it. That is, dynamic systems theory offers a natural language to a systemic approach to the study of adaptive behavior and cognition [33]. Much work has already clarified the need for a consistent deployment of the existing mathematical tools and for their further development [16], [32], [3], [4]. Nowadays, a more systemic view of the mind permeates at least a few major theoretical frameworks in the study of cognitive processes. Several authors are currently committed to the underpinning of a theoretical background, where the specific embodiment of an organism has non-trivial cognitive consequences. The body massively pre-/post-processes the information flow to and from the nervous system, and the common phylo-/onto-genesis of body and nervous system provides a deep, distributed integration of bodily and nervous functions (i.e., [6]). Perception and action are not causally sequential activities, but can be seen as an interwoven process, one supporting the other [28], [21]. Nevertheless, we have reasons to think that this might not be the whole story. Instead of treating the body as a mere interface to the world, we should also take into account what happens inside the body of an organism, and its potential cognitive consequences [25], [9], [10]. We find that the hidden, bio-regulatory dynamics developing under the surface of the body are typically largely neglected in the study of cognitive phenomena. As some authors put it, the interaction between bio-regulatory events that take place inside the body of an agent and what is traditionally interpreted as its control system, might be a non-negligible component of its ongoing cognitive processes [25]. In this short paper we comment on our experimental testing of such a proposal, focusing on the role that the intrinsic non-neural bodily dynamics might play in supporting and boosting cognition (section II). Furthermore, we also settle the programmatic foundations for an experimental extension of our work by advocating a new cognitive architecture for the study of anticipatory behaviors (sections IV and V), where non-neural bodily dynamics play a fundamental role. II. THE DYNAMIC ROLE OF THE COGNITIVE BODY: A MINIMALIST CASE STUDY In a recent study [19], [18], we showed how even very simple non-neural bodily states might play a crucial role in the modulation of the activity of an artificial nervous system, e.g. on the behavior generated by an artificial neural network (ANN) implementing the neurocontroller of our simulated robotic agent. We used standard evolutionary algorithms to set the weights and biases for a simple reactive ANN with no hidden layers, driving the motoneurons of a simulated Khepera robot. The system self-organized in order to find a recharger for its energy level (e.g. each instantiation of ANN during the evolution was simply rewarded for the maintenance of a positive level of energy, punished otherwise), thus overcoming its linear energy temporal decay. The energy level constituted one of the sensory inputs to the ANN. During the analysis of the successfully evolved system1, we manipulated the energy level as the control parameter for the whole system [16], [27]. By systematically clamping2 it to a discrete set of possible values, ranging from zero to ’full’, we observed and classified a number of possible behaviors. After exhaustion of the behavioral transients, we basically found three general classes of qualitative behavioral attractors. We observed: exploratory behaviors at the lowest levels of energy, e.g. the agent engaged in loops between potential energy sources and also in external loops broadening its explorations to the rest of the environment (i.e. see trajectory in panel ’A’ of Figure 1); more local behaviors at higher levels of energy (i.e. the agent was closely looping in the neighborhood of a single source as in panel ’C’ in Figure 1); hybrid behaviors, embedding characteristics from both previous classes (as in panel ’B’ in the same figure) for intermediate levels of energy. The relative frequency of the three groups of behaviors was reliably dependent on the current energy level (Figure 2). Back to the evolutionary task, we can examine the implications of the behaviors that we observed in clamped conditions. As the energy level is left free to follow its natural dynamics, it stands for an effective self-organized and dynamic action selection mechanism. Different classes of behavior are locally available to the agent as a function of its current energy level. Apparently, high energy levels imply that a source of energy was recently visited. Given the obvious physical constraints on the agent’s speed, it follows that it must be still in the proximity of the agent, consistent with the selection of local behaviors. On the contrary, low energy levels imply that the recent search 1By the term ’system’, here and in what follows, we refer not just to the explicitly evolved ANN, but to the global system constituted also by the agent, its environment and its non-neural dynamic mechanism of the energy level. Therefore, cognition is here conceived and analyzed as a broadly distributed process; a cognitive aggregate, rather than a localized and proprietary process. 2The term ’clamping’ refers to the injection of a steady energy level as input for the ANN during the whole duration of the replication of the experiment. The agent is free to behave in its environment for a length of time sufficient to exhaust all behavioral transients, yet permitting observations of satisfactory duration. Fig. 1. Sample spatial trajectories for the three classes of behaviors observed in clamped conditions. We can recognize exploratory behaviors (panel ‘A’), local behaviors (panel ‘C’) and hybrid forms (panel ‘B’). The scale of the environment has been systematically varied for the reader’s sake, in order to facilitate a deeper qualitative grasp of the attractor’s morphology. Potential energy rechargers are indicated by red stars. for recharge was not successful. This effectively correlates with broader exploratory behaviors. We can describe, as external observers, these behavioral properties. However, the solution of this minimalistic cognitive task relies on the self-organized dynamics of the whole system. On the contrary, in the traditional approach of cognitive science a similar mechanism would be modeled in terms of explicit representations and memory. So far, we have described our model using the intuitive metaphor of an energy level mechanism, thus evoking biologically plausible dynamics of food intake and metabolism. Nevertheless, our intentionally simple scenario aimed to facilitate the abstraction to general principles: metaphor aside, the fundamental aspect to consider is the coupling of different dynamic systems characterized by time scales that differ by several orders of magnitude (in particular we refer to the dynamics of the sensory-motor and the energy Fig. 2. The intensity of the pixels for each column (corresponding to attractors A-C) represents the relative frequency of the behavioral attractor as a function of the energy level. For energy levels in the interval [0.1, 0.6] we can observe a clear dominance of the attractors in class ‘A’. A similar dominance in the energy interval [0.8, 1.0] is shown by attractors in class ‘C’. The hybrid forms in class ‘B’ characterize intermediate energy levels. level systems). The availability of the slower dynamic of the energy level is exploited during the adaptation of the system. In fact, the neurocontroller receives input vectors which are structured as dynamically related events in a continuous flow (i.e. contexts with a similar, although continuously varying, level of energy). The outcome of the adaptation process allows the system to integrate information over time. Even if the sensory-motor mapping is purely reactive, this is not valid for the motor-sensory mapping and thereby for the behavior of the system as a whole. Extending these observations, we formulate the hypothesis that the access to a collection of attuned dynamic subsystems characterized by intrinsic dynamics at different time scales and the exploitation of such differences, might constitute a general cognitive mechanism, widely operating at the different levels of organization of biological cognition; a mechanism providing the cognitive system with the capacity to structure information on events which are relevant to its survival, with no need for explicit representations, memory or consciousness; a mechanism whose intrinsic time scales that characterize mechanical, chemical and electrical phenomena might be coherently integrated into the cognitive process. III. THE DYNAMICS OF ANTICIPATION Thus far, we have discussed an initial result demonstrating how the dynamics of the body, and in particular its bioregulatory processes, might constitute a powerful mechanism in support of cognition. Operative definitions of anticipatory behavior stress the effect that an estimation of the future state of the system has on its current behavior [26], [5]. We suggest that settling on a dynamic attractor (i.e. see [27]) constitutes an implicit form of anticipation in at least two important senses. First, once engaged with an attractor, the system enters a stable and fully determined regime. Our capacity to predict the trajectory of a strange attractor might well be limited by the confidence in the prediction that we can draw, as errors accumulate critically. Nevertheless, once settled on an attractor that currently satisfies specific functional requirements, the whole dynamic of the system is attuned to a specific flow of events. An example of this attunement and its anticipatory role is Pavlov’s dog, that salivates when food is made potentially available, thus effectively preparing its body for the digestive process. Some authors consider this kind of anticipation so important for an agent that conditioning, the prototypical basic form of learning in organisms, can be interpreted as mainly functional to its potentiation, for originally neutral stimuli become suitable for the elicitation of anticipatory responses [24]. Second, often during a metastable regime3, the closer the system is to a bifurcation the more its trajectory in the phase space tends to develop in the proximity of the attractor that is about to take over after the transition; therefore its future attractor is slowly formed from its own ’ghost’ in the actual current trajectory (see Fig.3, from [16, p.133]). Some authors (e.g. [16]) consider metastability the fundamental state for a complex dynamic system like the brain, for it allows flexible and fast engagement and disengagement with contingent environmental requirements and constraints. On this basis Kelso offers an inspiring dynamic image of biological brains [ibid., p. 26]: The human brain is essentially a pattern-forming self-organized system governed by nonlinear dynamic laws. Rather than compute, our brain ”dwells” (at least transiently) in metastable states: it is poised on the brink of instability where it can switch flexibly and quickly. By living near criticality, the brain is able to anticipate the future, not simply react to the present. All this involves the new physics of self-organization in which, incidentally, no single level is any more or less important than any other. To summarize, the body (in an extended sense that includes its non-neural internal mechanisms) constitutes a fundamental component of the potential dynamical richness of an agent attuned to its environment. From an evolutionary perspective such richness, when autonomously viable, is intrinsically endowed with anticipatory power. IV. THE BODILY PATH OF ANTICIPATION With the present work we intend to present a hypothesis about the possible role inside a cognitive architecture of 3The concept of metastability can be intuitively introduced as a dynamical situation where the system does not express stable states, but a mere tendency towards them [16]. Fig. 3. A system, initially engaged in chaotic regime, approaches a bifurcation from which it will emerge in phase-locked regime. The progressive darkening of the ’relative phase vs. control parameter’ diagrams (plotted in the two panels for two different values of the system’s parameters) announce the formation of the upcoming stable attractor. an explicit model predicting the dynamics of the agentenvironment sensory-motor interaction. Similar to [13], we will use the term emulator to denote such an explicit model. We will introduce our proposal by sketching an example that briefly portraits a prototypical situation. Let us consider a cognitive agent engaged with some activity that might result in noxious or other undesirable consequences. With reference to Fig.4, we have to state a few preliminary assumptions: 1) the global activity determining the current behavioral engagement between the agent and its environment (namely, a behavioral attractor, similar to [19], [30], [29], [14]) is described by a few global variables that compress the specific relevant information for the current sensory-motor activity out of the enormous number of degrees of freedom of the system [16]; 2) the box labeled SENSORY-MOTOR FLOW represents the neural activity associated with the sensory-motor flow of the behavioral attractor; 3) the corresponding non-neural bodily dynamics are summarized by the box labeled NON-NEURAL INTERNAL DYNAMICS; 4) within the global current dynamic is also embedded a neural sensory-motor emulator (box ANTICIPATION), whose evolution is dynamically correlated with the actual sensory-motor flow (similar to [31]), although not necessarily identical to it (as in [36]); 5) the dynamics of the emulator (ontogenetically or phylogenetically adapted) can anticipate, in the sense illustrated above, the sensory-motor consequences of the engagement with a potentially noxious activity, as they follow a faster time scale. Crucially, the capacity to predict the potential negative outcome endows the agent with a massive advantage: it attains the possibility to prepare itself before confronting the consequences of its current behavior, or to inhibit its behavior altogether. If we assume the possibility of a direct interaction between anticipatory and actual sensory-motor dynamics (i.e. a direct path between the boxes ANTICIPATION and SENSORY-MOTOR FLOW in Fig.4), we immediately recognize a critical problem. Which kind of dynamics would eventually emerge after the current action is inhibited or redirected? Obviously, the dynamic structure emerging in the emulator should elicit a viable alternative behavioral attractor. How would that be selected? Generalizing our example in order to also include other situations critical for the agent’s viability, our hypothesis is that, rather than a direct influence on the current behavior, the effect of the prediction of the emulator is actually mediated by the body. The outcome of the emulator affects the actual bodily dynamics (path a-b in Fig.4), and altered bodily quantities transiently act as control parameters for the actual sensory-motor flow (path b-sm). Hence, the problem of the determination of the next behavioral attractor is offloaded on the bio-regulatory dynamics of the body. Destabilized by the input from the sensory-motor emulator, the body reacts as-if actually engaged in such sensory-motor experience, eliciting behaviors that pull back the system towards safe regions. That implies also that the body can achieve homeostatic balance not only in virtue of isolated non-neural internal dynamics, but also by triggering the selection of an appropriate behavior (path sm-b). This mechanism exploits the knowledge ’accumulated’ by the body during a long and complex process of ontogenetic or phylogenetic adaptation, functional to the viability of the agent. Knowledge that, in case of a (theoretically possible) neural path directly coupling anticipatory and sensory-motor dynamics, should be somehow achieved by the nervous system. V. THE BODILY PATH HYPOTHESIS ON TRIAL: TOWARDS