Foundations of a Philosophy of Collective Intelligence

Philosophy, artificial intelligence and cognitive science have long been dominated by the presupposition that intelligence is fundamentally individual. Recent work in cognitive science clearly undermines that notion. Increasingly, intelligence is seen not as having its locus in the individual, but in the network of relationships that the individual has with the external world and other individuals. At the same time, there has been an increasing neo-Heideggerian focus on the role of embodiment and anti-representationalism, as shown by work ranging from robotics to dynamical systems. While philosophers are carefully trying to justify this development, the most significant computational phenomenon by far the World Wide Web is a veritable explosion of representations. In its latest stage, the Web has become increasingly more the realm of representations used for social real-time co-ordination, as a tool for “collective intelligence.” In order to make sense of these developments, we first summarize the differences between the Cartesian assumptions of classical artificial intelligence and the neo-Heideggerian embodied cognitive science. Then we show both how Brian Cantwell Smith’s story of representations can be built on top of a neo-Heideggerian story. A combination of a refined version of Smith’s rehabilitation of representationalism with the Extended Mind Hypothesis can explain the emergence of collective intelligence and its mediation through representations, and so the wide-scale success of the Web. Finally, we reconsider the notions of autopoiesis, the individual body and embodiment itself in light of collective intelligence. 1 The Individual Challenged The paradigmatic problem of both analytic philosophy and cognitive science is to explain the intelligence of the human individual: What properties of the individual human deserve credit for intelligence, and why? The answers seem to be self-evident; the unique combination of language and consciousness of the individual is the foundation of intelligence, both of which are not obviously found in ants, trees, or computers. Language and consciousness both seem to be incarnations of a reasoning process that leads to flexible, adaptive behavior, the general purpose reasoning mechanism of Descartes. Ranging from Frege and Russell onwards, philosophy of language sought to explain the relationship of the logical grammar and the world in order to explain why language is so effective, while more recently philosophers have been flocking to the rather mysterious “hard” problem of consciousness. On a more empirical vein, artificial intelligence attempts to understand intelligence through building mechanisms that display intelligence. Yet after the failure of classical artificial intelligence to produce intelligence in computers that could scale out of 1 University of Edinburgh email: [email protected] 2 It should be noted that artificial intelligence is usually the application of reigning theories in philosophy, and classical artificial intelligence was based on the “Language of Thought” representationalism in philosophy of language [5]. very small domains, a strain of research based primarily in robotics have shown that the very details of the implementation can produce intelligent behavior without representations, much less consciousness or reasoning [3]. This empirically-driven focus on embodiment has signalled the greatest change in artificial intelligence since its inception, and is explained by Wheeler as a shift from a classical Cartesian paradigm to a neo-Heideggerian programme [32]. Despite this revolution, one assumption that analytic philosophy, classical AI, and the new embodied AI all share is that the fundamental unit of analysis should be the individual. Recent empirical work in psychology and cognitive science has increasingly challenged the assumption that intelligence is irreducibly individual. It has shown that for complex tasks such as ship navigation that the success of the action relies on the co-ordination of multiple individuals [16]. Evidence from decision-making shows that the “wisdom of crowds” in other words, decision-making guided by the aggregate of information in a social network reliably makes better decisions than any individual [4]. Furthermore, work in developmental psychology has shown that the ability to point in children is more than an expression of a linguistic demonstrative, but rather an effort to produce a shared intentionality by directing the attention of others to the same object [31]. Some evidence from neuroscience the explosion of frontal cortex, long thought to be the seat of reasoning, evolved to keep track of interactions within a social network [8], and that the presence of mirror neurons provides a set of neurological mechanisms that allow individuals to share the same neurological state [9]). More recent work in tracking the behavior of individuals finds that their behavior ranging from movement to turn-taking in conversation can be reliably tracked by appealing to the behavior of others in their social network with a high degree of accuracy (over 40 to 80% of variation over a wide variety of tasks) without any appeal to planning, reasoning, or verbal language [24]. Pentland claims “that important parts of our personal cognitive processes are caused by the network via unconscious and automatic processes such as signaling and imitation, and that consequently, important parts of our intelligence depend upon network properties.” Instead of locating the intelligence in the individual, intelligence can be located in the collective aggregate of individuals. Collective intelligence does not necessarily mean the sharing of a cognitive state by, for example, mirror neurons. Intelligence can be exhibited by a network of individuals where each individual is specialized in a particular task so that no two individuals share the same cognitive state (skills, activity, and so on) per se, but that the successful action depends on the activities of the entire network. The classic cognitive ethnographic example by Hutchins is the piloting of a ship, where correct piloting of the ship depends on each individual, ranging from the navigator to the steersmen, completing their task [16]. Furthermore, it is not the simple aggregate or organization of individuals in a network that deserves credit for intelligence, but the conjunction of this social network with their environment. The environment should not be considered static, but dynamically shaped by the actions of intelligent behavior. However, some of the knowledge needed for success is not just embodied in individuals, but embodied in the environment, in their artifacts such as compasses and maps, and the very shape of the boat itself. This leads us to consider the example put forward by Herbert Simon of the apparent complexity of an ant’s path as it steadily marches towards food on the beach: “Viewed as a geometric figure, the ant’s path is irregular, complex, and hard to describe. But its complexity is really a complexity in the surface of the beach, not the complexity in the ant” [27]. Although this may be true in some cases, it would be to primitive to describe the ant totally to be at the mercy of its environment. Intelligence in general collective or not leave traces behind in the environment. The classic example is the pheromone trace of the ant, in which a traces get reinforced as more ants use a particular trail, has been shown to be an efficient way of navigating the environment. This shows how individuals with limited memory can use the shaping of the environment as an external memory. Culture, ranging from design of cities to Wikipedia, can be considered collective cognition extended into the environment. This usage of the environment has a number of advantages over direct individual-to-individual communication. As noted by Heylighen, there is no need for simultaneous presence, so interaction can be asynchronous, and individuals can even be anonymous and unaware of each other. This allows highly organized successful actions to be performed by individual that, due to limited memory and knowledge, would be unable to achieve success otherwise [14]. To modify Pentland’s thesis: The collective activity of individuals and their modifications to the environment are responsible for intelligence. While at first this thesis seems intuitive, it goes against much of the practice of both classical cognitive science and philosophy that have a tendency towards individualist reductionism. While the question of whether or not this thesis is actually true is a distinctly empirical question, the philosophical ramifications of this thesis should be developed to see if they are in conflict or continuous with the neo-Heideggerian framework currently being championed in philosophy and AI. Two points of conflict immediately become apparent. Although Heidegger himself is unclear, the neo-Heideggerian framework as articulated by Wheeler understands intelligence as a function of the situated being in the world, not a collective of beings in a shared world [32]. Furthermore, the neo-Heideggerian framework does not explain the reshaping of the environment by intelligence, in particular the creation of representations, not just representational explanation. Representations are seen as crucial by many for the emergence of collective intelligence, which Hutchins traces his “distributed cognition” to “the propagation of representational states across representational media” [16]. The neo-Heideggerian framework is most associated with robotics that exhibits “intelligence without representation,” and in contrast collective intelligence is most associated with the advent of the Web, a veritable explosion of representations if ever there was one. To tackle these problems, we will focus on them in reverse order. First, after explaining the rising neo-Heideggerian framework in cognitive science by contrasting it with the classical Cartesian framework, we will show how representations can be built into the framework. Then, by pushing on the Extended Mind thesis, we will show how the neo-Heideggerian framework allows collective intelligence, including those that use representations. We can then use this framework to understand the explosion of collective intelligence on the representation-heavy Web, and finally try to reconstruct a notion of what should replace the individual in philosophy. 2 Neo-Heideggerian Embodiment The philosophical assertions made by proponents of neoHeideggerian programme must be summarized in order to see if they are continuous, or in contradiction with, a theory of representationbased collective intelligence. This is difficult, as like classical artificial intelligence, the move towards embodiment in AI has mainly been one of empirical work where the philosophical assumptions have for the most part been implicit in the work itself. Just as Dreyfus unearthed the philosophical presuppositions of Cartesian classical artificial intelligence, Wheeler has effectively summarized the assertions of embodied AI and based them firmly on a reading of Heidegger, which we call the neo-Heideggerian programme [32]. The neo-Heideggerian programme is best understood in contrast with the neo-Cartesian programme of classical AI. Wheeler digests this programme into three main assumptions: • The subject-object dichotomy is a primary characteristic of the cognizers ordinary epistemic situation • Mind, cognition, and intelligence are to be explained in terms of representational states and the ways in which such states are manipulated and transformed. • The bulk of intelligent human action is the outcome of general purpose reasoning processes that work by retrieving just those mental representations that are relevant to the present behavioral context and manipulating and transforming those representations in appropriate ways as to determine what to do It should be noted that at first glance these neo-Cartesian assumptions are based on the individual being the locus of intelligence. That is surely how at least Descartes thought of it: The singular subject is operative in “cogito ergo sum.” The first of the Cartesian points seems to have an implicit individual subject, while the second remains neutral, and the third also seems to have an implicit human individual as the subject. Wheeler then makes the fairly accurate assessment that “word on the cognitive-scientific street is that classical systems have, by and large, failed to capture in anything like a compelling way, specific styles of thinking at which most humans naturally excel” [32]. However, all hope is not lost for AI if it can only lose its neo-Cartesian assumptions. Based on a survey of current work in AI, ranging across robotics, artificial life, and dynamical systems, Wheeler unifies these diverse works on four new assertions, which he states as follows [32]: • The primacy of online intelligence: The primary expression of biological intelligence, even in humans, consists not in doing math or logic, but in the capacity to exhibit...online intelligence...a suite of fluid and flexible real-time adaptive responses to incoming sensory stimuli. • Online intelligence is generated through complex causal interactions in an extended brain-body-environment system: Online intelligent action is grounded not in the activity of neural states and processes alone, but rather in the complex causal interactions involving not only neural factors, but also additional factors located in the non-neural body and the environment. • An increased level of biological sensitivity: Humans and animals are biological systems and that matters for cognitive science. • A dynamical systems perspective: Cognitive processing is fundamentally a matter of state space evolution in certain kinds of dynamical systems. Is there any bias towards an individual subject in these assertions? It seems present in a subtle manner in the first assertion since the very idea of “incoming sensory stimuli” presumes an individual that is processing these stimuli. The second and third assertion also seem to take for granted that our primary subject is not just an individual, but a biological individual. This is put into perspective by the second assertion that “not only neural factors, but also additional factors located in the non-neural body and the environment” play a critical role, a point we will return to with a vengeance. Wheeler and his philosophical fellow-travellers such as Clark [5] spend much of their time on the question of whether or not there is any room whatsoever for internal representations inside these individuals. Rejecting Clark’s notion of “decoupling” as sufficient but not necessary for cases he believes demands a representational explanation, Wheeler argues for some, albeit limited role for representations that pins representations on the two notions of homuncularity and arbitrariness. Since it is too involved to argue over homuncularity and arbitrariness here, we shall instead focus on how Brian Cantwell Smith’s revival of decoupability can be built on a neo-Heideggerian framework. We shall just comment that Wheeler’s general framework is not incompatible with our notion of collective intelligence and his account of representations is not too far from our account. 3 Representations Revisited The very idea of representation is often left under-defined and is as a consequence given near-magical powers by certain theories of language and classical AI. While it is hard to pin down a reigning definition, the classic definition stems from the notion of a “symbol” given by Simon and Newell’s Physical Symbol Systems Hypothesis [22]: “An entity X designates an entity Y relative to a process P , if, when P takes X as input, its behavior depends on Y .” First, the very idea of “being a representation” is grounded in the behavior of a process, and behavior depends on having access to the representation. Thus, the target of representation (i.e. what is represented, the “thing designated”) will depend on the process the representation is used in, i.e. a representation is never context-free. Second, there is clearly decoupling “for this is the symbolic aspect, that having X (the symbol) is tantamount to having Y (the thing designated) for the purposes of process P ” [22]. This definition seems to have an obvious point of conflict with the neo-Heideggerian agenda, for it reflects the infamous “subject-object dichotomy” due to its presupposition of at least three distinct a priori entities, the subject (P ), the representation (X), and the object (the “target” of the representation, Y ). To the extent that these distinctions are held a priori, then the definition is the very exemplar of the neo-Cartesian programme of classical AI. An escape-hatch from this Cartesian dead-end would exist if there was a way within the neo-Heideggerian program to tell the story of how representations come to be without an a priori subject-object dichotomy. Brian Cantwell Smith tackles this by developing a theory of representations that does not presume an individual [28]. Smith starts with the example from Lettvin and Maturana, a frog tracking a gadfly across the sky [17]. The frog sees the fly, and begins tracking it with its eyes as it flies. The frog and the gadfly are both physically connected via light-rays. Borrowing an analogy from physics, everything is composed of non-distinct fields of energy, so it would be a presupposition to talk about a frog, a fly and light as individual objects. All that exists is some sort of pre-individual flow from which individual objects may emerge. At the moment of tracking, connected as they are by light, the frog, its light cone, and the fly are a system, not distinct individuals. An alien visitor might even think they were a single individual. When the fly goes behind a tree, and the fly emerges from the other side of the tree, the frog’s eyes are not focused on the point the fly was at before it went behind the tree, but the point the fly would be at if it continued on the same path. Components of the flux are now physically separated, with a mutually distinct o-region and s-region. The s-region is distinguished from the o-region by virtue of not only its physical disconnection but by the s-region’s attempt to “track” the o-region, ”a long-distance coupling against all the laws of physics” [28]. After disconnection (and possibly more cycles of disconnection and re-connection) the s-region can stabilize as an individual subject and the o-region as an individual object, and with considerable work on the subject’s side to “track” its object a representation is created by the subject using some form of dynamically incoherent memory. Both subject and object are then full-blown individuals, with the subject possessing a representation of the object[28]. The individuals are not a-priori distinct, but coconstitute each other. According to this explanation subject and objects co-evolve, with the physical processes used to track the object being the representation. In order to clarify and make abstract Smith’s analogy and explicitly connect it to Simon and Newell’s definition, we can divide Smith’s process into what I have called the representational cycle [10]. In order to explicate why precisely the s-region differs from the o-region, we rely on Rocha and Hordijk’s work on evolving representations, in particular their idea of dynamically incoherent memory [25]. Dynamically incoherent memory is defined as a type of memory not changed by any dynamic process it initiates or encounters. In this manner, it serves as memory that does not degrade or radically alter, but can maintain itself over time. To phrase this outside of the language of dynamical systems, we would say that “dynamically incoherent” might be a misleading word. Instead, what Rocha means is that the subject must have a some sort of memory that is capable of maintaining coherence in terms of its physical structure against, ”the vagaries and vicissitudes, the noise and drift, of earthy existence” as Haugeland would say [11]. The cycle can then be put into four stages