The Computational Boundary of a "Self": Developmental Bioelectricity Drives Multicellularity and Scale-Free Cognition

There is a fundamental symmetry between anatomical control mechanisms and cognitive mechanisms. Co-evolution and exaptation drove the mutual enlargement of mechanisms that control patterning and behavioral goals. The same dynamics operate in unicellular systems, multicellular systems, and colonial/swarm organisms and most concepts from memory to cancer are found at every level of organization in biology, from memory in transcriptional networks to regeneration of termite nest structures.

Neurons utilize bioelectric computational strategies that were discovered and exploited by evolution as far back as bacteria. There are no sharp distinctions between neural networks and non-neural somatic bioelectrical networks (although they function on different time scales). The functional isomorphism between patterning and cognitive processes is also reflected in the ancient molecular conservation of mechanisms: ion channels and neurotransmitter molecules are ubiquitous across the tree of life. Bioelectric integration helped evolve control strategies and cognitive content across the continuum from chemical networks to human minds – it illustrates an important mechanism of early evolution. But clearly, numerous aspects of physics (from stress forces to diffusing chemicals/pheromones) are exploited by evolution, or could be exploited by engineers, across the wide range of possible systems.

There is a deep functional scale-invariance between the decision-making of cells in building body structures, the workings of an insect colony, and the integrated behavior of a human “Person”: these are the cybernetic processes of learning and parameter optimization implemented by large numbers of units pursuing infotaxis and homeostatic goals at whatever scale the sensory channels permit.

A conceptual unification is proposed as Scale-Free Cognition: one major control knob is the boundaries between self and world. These boundaries are malleable, and can shift at different time scales, to sizes limited by what the underlying hardware supports. This parameter determines the scope of the self and implements the continuum leading smoothly from cell- > body- > swarm. Signaling between animals in an ecosystem (Pais-Vieira et al., 2013; Kingsbury et al., 2019; Zhang and Yartsev, 2019) is not fundamentally different than signaling within the brain – both are examples of information propagating through a network of locally-competent micro-agents. Others have pointed out the parallels between the dynamics of cancer and ecosystem-level degradation (Degregori and Eldredge, 2019). Thus, multiple levels of approach to living systems are a priori equally valid, with no unique privilege for the lowest (molecular) levels at which everything looks like a mechanism. In any given example of biology or artificial life, the most appropriate level of analysis or description is to be determined empirically: it is the one that facilitates prediction and control with the least effort (by the experimenter or by the system itself), and gives rise to unified understanding that drives the most novel, robust research programs at the bench.

This view is consistent with Dennett’s stack of homunculi (Dennett, 1991). Here, the bottom homunculus is a basic homeostatic loop, which is required for even the simplest life form to persist against entropy (Varela et al., 1974; Maturana and Varela, 1980). Each succeeding level of that same homunculus has a tiny bit more reach in physical space (how far it measures and how far its signals propagate) and in time (how long it remembers past signals and how far it can anticipate) - the homunculus is not entirely blind but has a horizon within which it can perceive. The bottom homunculus measures very simple, immediate things (chemical concentrations on the surface of a single cell), and has a feedback loop that keeps that one variable within a life-compatible range. The others see a little more and manage slightly more complex allostasis. This naturally scales into “goal-directed systems” and enables massive improvements in evolvability because these components make random moves in genotype space into acceptable (and sometimes brilliant) ones in phenotypic space. The functional modularity enabled by learning and goal-directed feedback loops radically modify the evolutionary search space landscape, because they can improve the phenotypic quality of mutations. Mutations that by themselves would have been useless or harmful to the organism (because they usually adjust one feature without directly implementing the many others needed for functional advantage), can be made workable if the various parts are flexible and alter their behavior to reach the same attractors despite shifting surroundings.

Developmental and regenerative biology are full of these examples, like the rearranged tadpole faces whose organs then move around to make a normal frog face (reviewed in (Lobo et al., 2014; Pezzulo and Levin, 2015, 2016)). Likewise, manipulations that cause an ectopic bone to form in the embryonic hand do not simply result in a single, out of place bone - the nearby muscle cells, blood vessels, and tendons make up for this novel circumstance and build an extra, workable finger. This helps explain Darwin’s original puzzle of evolvability because the body is not a static canvas or merely a set of features emerging from micro-level rules. Morphogenetic goal-directedness is a major evolutionary ratchet, smoothing the anatomical fitness landscape by hiding the otherwise potentially disruptive effects of random mutations. This is a point that has been made in the context of gene-regulatory circuits, whose self-organizing capacity takes a lot of the heavy lifting off of the evolutionary process (Kauffman, 1993).

When the subunits exhibit their own level of competent, local intelligence, many mutations do not need to occur before a feature is useful. One mutation can change how something operates, but the other parts may be able to continue to pursue their same anatomical goal even though conditions have changed. This somatic plasticity is ubiquitous in biology because complex agents inevitably consist of micro-agents that were selected on the basis of proto-cognitive competency. This view of modularity from the perspective of basal cognition complements traditional genetic accounts. This is likely also the source of cognitive plasticity. Tadpoles engineered to develop with eyes on their tails instead of in their usual spot can see quite well, despite the fact that the eyes connect to the spinal cord, not the brain (Blackiston and Levin, 2013; Blackiston et al., 2017). Brains, like other developing organs, are not hardwired but are able to ascertain the structure of the body and adjust their functional programs accordingly – a strategy that is already being pursued in robotics, including the use of electrophysiological principles (Bongard et al., 2006; Cheney et al., 2014). This self-discovery phase is an important area for future research (Ramstead et al., 2019).

The way regulatory pathways are organized (primitive intelligence of the parts, exploiting attractors in state space and modularity to attain goals at various levels despite changing conditions) makes complex traits much more easily evolvable. Organs and tissues have local goals (in the dynamical systems and control theory sense of ‘goals’) and their ability to reach these goals despite damage and changing environmental conditions (something for which evolution selected from day one) is what enables random mutations to power evolution in a rapid timeframe. From the perspective of the cells, a mutation is just another attack of a hostile environment - they need to keep going despite this perturbation, and they adjust their activity as they would to any external stressor. On this view, the organization of cybernetic control structures at different levels originating from simple homeostatic survival advantages, are the source of agency at multiple scales (Arnellos and Moreno, 2015).

This view has important implications for treatment of injury and disease (Levin, 2011). Most of the focus of biomedicine today is on the cellular hardware, seeking molecular- or cell-level interventions. The problem is that even with full genomic information and stem cell derivatives, directly creating an organ such as a human hand directly from individual cells is likely to be beyond our capabilities for a long time. Moreover, the complex, emergent nature of morphogenesis establishes an important inverse problem: how to manipulate low-level rules to achieve system level outcomes such as repair or regeneration of organs damaged by traumatic injury or disease? I have previously suggested that a better strategy than cellular re-wiring at the genetic level may be to edit the pattern-homeostatic set point and let wild-type cells build to the new spec. For example, the ability to re-write target morphology in planarian regeneration was recently shown, by altering the bioelectric state in flatworms, without genomic editing, giving rise to permanently double-headed worms (Durant et al., 2017; Levin et al., 2018). Similarly, a brief stimulus (24-h exposure to chemical cocktail) kick-starts a limb regeneration module that drives growth for 11 months with no further intervention (Herrera-Rincon et al., 2018). This ability to initiate very complex, long-lasting morphogenetic cascades via a simple trigger demonstrates the utility of understanding the modular control structure of large-scale patterning systems so that the most efficient (least effort, simplest) intervention can be deployed, at the right level of organization, to induce a desired system-level outcome.

What is the difference between a mechanism and decision-making (choice)? This has been widely discussed in the philosophy literature (Barandiaran and Moreno, 2006; Haig and Dennett, 2017) and is now becoming an important issue for bench biologists working in basal cognition of somatic cells and synthetic biology (Perkins and Swain, 2009; Balazsi et al., 2011; Reid et al., 2013, 2016; Mitchell and Lim, 2016; Paul et al., 2016; Vesty et al., 2016; Bugaj et al., 2017) as well as for workers in artificial life (Juel et al., 2019). Without attempting to deal with this profound question in full, it can be mentioned that the perspective taken herein suggests that several factors contribute to a smooth transition between biochemical mechanism and agency: integration of remote events into the causal chain (spatial distance, and temporal distance – memory/anticipation), and stochasticity (distance in terms of predictive capability). The greater the area of the cognitive light cone that comprises the events that cause a particular outcome, the more likely that we will gain explanatory and control power by treating it as a choice and not a push. If an outcome depends on things that happen far away, and in the past (memory) or future (planning), then it is likely to be empirically useful to consider that process a choice. This heuristic has the attractive natural corollary that the bigger the cognitive horizon (scope of self) of an agent, the more apparent freedom that agent has (when causes are very far away, in space and time, from a given action is when folk psychology usually labels an event as a free choice).

Bioelectric signaling is an ancient aspect of physics, which evolution exploited for its computational capacities. Neural Hodgkin-Huxley equations can be derived from first principles of information theory (Gatenby and Frieden, 2017), but the use of bioelectrics long predates nervous systems. Why is bioelectricity, implemented by ion channel and gap junction protein hardware, so well-suited for implementing the flexible software of life? Voltage-gated ion channels and gap junctions are voltage-gated current conductors – basically transistors, from which as we know all manner of advanced computing devices can be constructed. Such channels readily enable loops which either amplify small signals (positive feedback) or implement robustness (negative feedback), because they create voltage changes that then control their own activity (open or close the channel), in a physiological cycle that occurs at the post-translational level (not requiring any gene-regulatory mechanisms). Also, they very naturally link up into networks that enable spatial integration of signaling. Thus, any electrical circuit that enables feedback loops and memory is sufficient for very primitive software behavior in the following sense. It can stably occupy more than one state, and shifting it from one state to the other does not require changing out the hardware (altering protein profiles by transcriptional changes). What makes electric circuits, whether ion channel-mediated or electronic, good for software is that they easily implement dynamical systems that have multiple attractors and can be made to occupy one physiological state or another by transient inputs (signals), not requiring changing the hardware (i.e., mutating the DNA encoding that hardware).

This is also true to some extent of physical forces and cytoskeletal controls, which can enable extremely rich morphogenetic and behavioral processing in unicellular organisms (Faure-Fremiet, 1953; Ford, 2017) and can scale up to tissue-level controls (Roper, 2013).

Some agents can transcend one system and re-set its goals from a vantage-point of a meta-system (Hofstadter, 1979). While many animals are capable of a wide repertoire of diverse goal-seeking activity, human minds appear to be (for now) in a unique position. A fish can successfully meet the most basic goal of “survival” because it only sees a short time into the future, and surviving for that short time is usually achievable. However, humans (who are prone to reflection on these issues) can infer that “survival” is not actually an achievable goal on timescales that we can think about; uniquely among goal-pursuing systems, we alone know that this most basic of all life goals is doomed to failure. We are the only known agents who must change their goals, because our cognitive horizon is, along the temporal axis, greater than our lifespan and we can envision goals that extend beyond our own life. Therefore, we as cognitive agents are perhaps uniquely motivated to set other goals ahead of survival (Frankl and Vance, 2007). This potentially includes pursuit of the most radical goal of all, to cease functioning as a unified goal-seeking system entirely. “To be, or not to be?” – when did this question first get asked, phylogenetically? A cabbage cannot commit suicide. A human can; could a non-human primate? As with other kinds of behavioral capacities (Bronfman et al., 2016), this capability likely represented an important phylogenetic cognitive transition. Individual cells have this capacity [even unicellular organisms (Gordeeva et al., 2004)]; it is unclear if metazoan organs do. The synthetic biologists may want to ask, how would we create an organism capable of such a meta-goal?

Needless to say, many details of this set of ideas remain to be worked out. However, the above-described perspective makes a number of specific predictions which suggest experimental approaches to test and exploit this perspective (Levin, 2011). Future empirical work will reveal whether Scale Free Cognition is a useful way to organize what we know about the mechanisms and algorithms of life and life-as-it-could be (Langton, 1995; Walker and Davies, 2013). Directions for future research, suggested specifically by this synthesis, include:

Rapid and successful adaptation of biological systems, at the cell and tissue level, to novel perturbations should be considered with respect to the cognitive task of inferring which genes to activate (microscale) to deal with a physiological stressor (macrostate) (Elgart et al., 2015, 2016; Soen et al., 2015; Schreier et al., 2017). Is there a quantifiable sense in which biological systems model themselves? Can tools such as information theory or other approaches be used to identify signatures in time-series data of regulatory events of attempts by gene-regulatory and bioelectric networks to coarse-grain themselves – to discover their own causal structure for optimally efficient self-control?

Can an in silico evolutionary system be built, containing both genetic and physiological components, which simulates the scheme described above (homeostasis and infotaxis) and illustrates the emergence of different scales of cognitive horizons over time? Can we directly observe the evolution of multicellular goals from networking of agents with single-cell homeostatic goals?

Because of the mechanistic and functional commonalities between cognition and patterning, neurotransmitters are ancient control mechanisms functioning also outside the nervous system in fungi, plants, and animals. It is already known that perturbation of these pathways can induce low-level effects [perturb developmental patterning (Buznikov and Shmukler, 1981; Hernandez-Diaz and Levin, 2014; Sullivan and Levin, 2016)] but might there be higher level effects on the physiological boundaries of the cognitive sub-agents within the body from exposure to psychedelics, which have long been claimed to expand “transpersonal boundaries” or induce “ego death” (Bouso et al., 2015; Carhart-Harris et al., 2016)? Specifically, the implementation of multicellularity by the drive to minimize predictive error (Fields and Levin, 2019) suggests that cells need a way to keep other cells nearby. This suggests the testable hypothesis that the addictive power of opiates and similar molecules derives from such ancient roots. Functional experiments targeting addiction pathways can readily be attempted in early models of multicellularity and in somatic/stem cell co-cultures in vitro, to identify or disprove an ancient role for addictive neurochemicals in this process.

It would be instructive to start fleshing out the cognitive light cone diagrams with behavioral data from real organisms and available AI agents, as well as single cancer cells in vitro and in vivo, and tumors. Additional concepts in this space may need to be added, such as conservation principles; e.g., the amount of “mental energy” for caring about things and making decisions is certainly limited in humans (Weippert et al., 2018); can we measure and define the thermodynamic/metabolic costs of decision-making in basal cognition? More broadly, the analogies between geometric spaces and cognitive ones needs to be explored further, as is beginning to be done via geometric information theory (Balduzzi and Tononi, 2009; Oizumi et al., 2016; Sengupta et al., 2016; Calcagni, 2018). Indeed, the notion of spaces with different geometry has been pursued both in the neuroscience of perception and representation (Zhou et al., 2018b; Gilead et al., 2019) and in developmental biology of patterning (Jaeger et al., 2008; Jaeger and Monk, 2014). It may even be possible to go backwards, asking what cognitive structure accompanies a particular space-time geometry – a concept that might have been presaged by Newton’s concept of space being the “Sensorium of God”.

A major question in AI is motivation (McShea, 2013) – will complex engineered systems set novel goals, and if so, how? It is usually extremely obvious how to negatively motivate a living organism, but how does one punish an artificial agent? It may be that until we have artificial constructs that are, at their core, homeostatic goal-seeking systems, which will be as easy to motivate as living things, significant general AI will not be possible. While biology has competent agents at each scale (McShea, 2012), most human artifacts do not – robots are often made of reliable but very dumb parts [although modular designs with sub-goals have already been shown to be a useful strategy (Cudhea and Brooks, 1986)]. It is a prediction of this view that adaptive, useful robotics will require components that are themselves competent and goal-seeking. The fact that robot cancer is not a known problem today is likely the same reason we do not have highly adaptive, robust robotics: selection of activity (feedback based on positive and negative reinforcement toward local goals) is applied only at one level. Subunits rich enough to make a good cognitive Self will occasionally defect and go off the reservation as human cells sometimes do. This suggests a roadmap for engineering agents that are fundamentally based on homeostatic goals and selfish infotaxis at multiple levels of organization.

The importance of physiological connectivity in keeping cells harnessed toward a global body plan and away from cancer has long been known. The lack of multicellular information is a potent carcinogen; e.g., plastic barriers (but not the same plastic powder) placed between cells induces tumors, as do gap junction blockers (Oppenheimer et al., 1953; Bischoff and Bryson, 1964; Yamasaki et al., 1995). Data in amphibians already show that modulation of the bioelectric state and connectivity of cell networks can prevent and reverse tumorigenesis (Chernet et al., 2015, 2016); thus, the next work in this subfield should focus on the discovery of bioelectric communication-inducing technology [whether through gap junctional opener drugs or bioelectric nanomaterials (Jayaram et al., 2017) for cancer reprogramming as an alternative to chemotherapy]. Another possibility is that stress (physiological, or informational – signals that cannot be predicted by a cell) could be a factor that motivates cells to reduce their sensory/action surface and contract, in effect abandoning organismal membership when it begins to generate more stress than its presence reduces by its protective action. The current research on the relationship between stress and cancer (Benndorf and Bielka, 1997) could be modeled from the perspective of surprise minimization.

The proposed crucial role of bioelectrics in multicellularity predicts that it should be possible to induce the formation of metazoan-like bodies in an otherwise unicellular organism by forcing the expression of appropriate electrogenic proteins and gap junctions. Similarly, it may be possible to induce dissolution of an entire metazoan body by appropriate changes of bioelectric dynamics.

Bioelectrics should be widely conserved in regeneration across Kingdoms; transcriptomic analyses of regeneration events from damaged bacterial biofilms to plants to metazoan models of limb and organ regeneration should all show consistent use of electrogenic proteins in the events that enable cooperation toward self-limiting morphogenetic cascades. Likewise, it should be the case that neurotransmitter signaling (a key downstream response module for bioelectric change) should be highly conserved, revealing effects of cognitive modulator drugs on function in bacterial biofilms and primitive cognition models such as Physarum (Maier et al., 2018).

The evolution of neural control systems from pre-neural morphogenetic mechanisms suggests that there should be a lot of interplay between these systems in development. The requirement of a functional CNS for effective organ regeneration (Brockes, 1987) and for correct patterning in development (Herrera-Rincon et al., 2017; Herrera-Rincon and Levin, 2018) has been shown in a range of model species, and the tools of optogenetics can now be used to dissect the information content of nerve-mediated signaling as instructive cues for both normal morphogenesis and regulative development in which bodies adapt to major changes in architecture (Oviedo et al., 2010).

The continuum view of cognition suggests that the cognitive ability to learn from experience and act to maximize specific parameters to increase welfare is universal across scales. Much work remains to characterize and exploit phenomena such as cardiac memory (Chakravarthy and Ghosh, 1997; Zoghi, 2004), and learning in bone (Turner et al., 2002; Spencer and Genever, 2003) and in gene-regulatory networks (Herrera-Delgado et al., 2018). Emerging technologies for real-time, closed-loop controls (Bugaj et al., 2017; Perkins et al., 2019) now enable interrogation of cellular cognition in a variety of somatic contexts ex vivo. We have previously suggested the use of behavior shaping and training paradigms as a strategy for synthetic morphology and regenerative medicine that complements bottom-up rewiring at the molecular level (Pezzulo and Levin, 2015, 2016; Mathews and Levin, 2017).

Similarly, it should be possible to communicate with (train) swarm organisms. Moreover, the scale-free cognition hypothesis suggests that multi-human systems could have their own degree of cognition. We speak this way informally (the Supreme Court has opinions, a town may send letters to specific people, and countries are often portrayed as having intentions and behavioral traits in political planning). Tools of information theory and reinforcement learning techniques could be tested on the behavior of human social groups, to determine if there is a true psychology of an integrated Self that can be predicted and managed as a coherent goal-seeking individual. It is fascinating to note that swarm cognition, in the sense of a new supervenient Self which makes choices and reaps the consequences, is a concept proposed long ago in pre-scientific thought about cognition (Hutchinson and Sharp, 2008).

Ascertaining the functional set points (goals) of a given system has implications for open questions in exobiology and artificial life. First, it facilitates identification of a system’s boundaries – empirical hypotheses about what subset of a complex biosphere is practical to consider a unified Individual. This is sometimes obvious in the world of terrestrial biology, but it is an important potential problem to be encountered in research in Life as it Could Be (Langton, 1995). Second, it enables communication with highly diverse intelligences (evolved or constructed), with whom we may share no language and few mental constructs. Given a novel life form, how does one know when successful communication has taken place? A basic form of communication is behavior shaping: if a system can be made to act in a pre-determined way through positive and negative reinforcement, then one is sure that at least one thing has been communicated: the desired activity. Of course this is a very minimal notion of communication, but it is a practical and useful starting point in scenarios where the biology or engineering is truly novel. This requires understanding what motivates the system (i.e., what conditions it seeks vs. avoids) and takes advantage of its innate goal-seeking behavior.

Most of the prior discussion focused on objective, functional traits, and capabilities and is compatible with several views on consciousness. Nothing in this model explains why specific functional features give rise to first-person experience (the so-called Hard Problem). But, given that at least some nervous systems do give rise to such experience and that the differences between neural networks and non-neural ones (aside from temporal scale) are minimal, it is natural to hypothesize that there is something-it-is-like to being a tissue or organ and making decisions. It is not claimed of course that a pancreas, in its striving to keep homeostasis, has any self-awareness in the human sense, but it may have as much proto-consciousness as a simple neural network and indeed diabetes has already been modeled as a kind of cognitive disorder (Arntfield and van der Kooy, 2011; Goel and Mehta, 2013). The model does take a stand on the perennial “combination problem” and can perhaps be seen as a form of panpsychism (Chalmers, 2013).

One final comment concerns an interesting intersection of the above model with non-Western views on consciousness. It is striking that the process which Zen practice is meant to reverse – attachment to past memories and high valence for future expectations/fears – is precisely the process suggested to be responsible for the creation of complex Selves. It is unclear whether it is beneficial (or even possible) to truly live in the moment and let go of past memories and future expectations, but anyone who succeeded in doing this would achieve precisely what Zen promises: the dissolution of the self (Flanagan, 2011; Josipovic, 2014, 2019). According to the above model, the Zen ideas of stamping out desire (goal-directed activity, preferences for specific states of affairs) are exactly correct in that this would lead to a dissolution of the ego (Self) and the freedom from the law of cause and effect that governs the Individual’s actions. By turning off memory, anticipation, and striving, the essential glue that creates a cognitive Self is dissolved. By erasing the set point toward which the feedback loop expends energy to accomplish, the higher level integrated Self disappears, leaving nothing but the constituent parts (smaller Selves in their own right). This is completely different from killing the individual components, and it can be asked whether someday we will develop a biochemical path to the Nirvanic Void, which releases the global self by breaking the integrating communication channels but leaves all of the subunits healthy and free to pursue their local goals. Re-creating the unification into, and liberation from, larger scale unifying Selves would be a true pinnacle of synthetic biology and artificial intelligence engineering.

The author confirms being the sole contributor of this work and has approved it for publication.

The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.