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

Complex cognitive systems can have very large and multifaceted goals, and this arises through evolutionary and ontogenetic scale-up and elaboration of primitive goals that arose from constraints of thermodynamics and homeostasis. The initial, most primitive feature of a living system is preferences – the fact that some states of the world are better for its welfare than others. This enables learning from positive and negative reinforcement, which leads to an explosion of computational and behavioral capabilities. Preferences evolve into goals to the extent that a system grows in complexity and causal power and becomes able to act in the world in ways that are likely to move it toward preferred regions of its state space (initially, focused on simple metabolic survival, but ranging all the way to complex human psychological needs and perhaps beyond). It is likely that any life we observe today (which has passed the filter of selection) has preferences and is good at optimizing for them (making this criterion a useful part of the very definition of life), but it is possible that our initial efforts at artificial life may construct some truly primitive, transitional cases that can maintain a degree of coherence in a sheltered laboratory environment without an efficient goal-seeking capacity.

A very simple organism can only have preferences about what is occurring at the current time, in its immediate environment. A more complex organism whose causal structure enables associative learning can pursue or avoid stimuli that are several steps removed in space, time, and causal connection from whatever it is choosing among. This kind of learning enables associations between stimuli that impinge on very different sensors in different parts of the organism and leads to behavioral preferences about stimuli that are in themselves neutral (do not cause damage or provide immediate reward) but are linked to future positive and negative outcomes by past experience (indirect causal connections stretching across body distance and life history).

Others have already made the link between associative learning and the emergence of self (Ginsburg et al., 2019). Very complex selves can have preferences about abstract states that are very far away indeed (such as human beings who are genuinely troubled by the ultimate fate of our star and are actively working toward long-distance space travel and the fate of humanity as a whole). However, the spectrum of Selves is not a simple linear one because of the very wide range of possible natural and artificial agents we do (and will increasingly) encounter. The biological world offers numerous corner cases of swarm individuals, but a good conceptual framework will also include organic artificial life, engineered (computer-based) artificial intelligences, and potential exobiological discoveries. Can highly diverse Selves, with very different material structures be compared with each other in any meaningful way? I suggest that a universal rubric, applicable regardless of the physical implementation, can be defined by focusing on the information processing and goal-directed activity of any given system.

The cognitive boundary of an Individual [a “Center of Concern” (Murase and Asakura, 2003)] is the most distant (in time and space) set of events that this system can measure and attempts to regulate in its goal-directed activity. It is a surface indicating what things this system can possibly care about (conversely, it defines preferences as the spatio-temporal domain of states that serve as inputs and outputs to the system). In advanced agents, it is also the boundary of the self-model. A range of such systems is illustrated in Figure 2: using one axis for time and one axis to represent three dimensions of space enables a semi-quantitative representation within which individuals of very variable cognitive capacity can be plotted in the same virtual space. Recent work in artificial life has already begun to characterize the cognitive domains of very diverse kinds of complex systems (Beer, 2014).

The edges of a given Agent’s goal space define a sort of “computational light cone” – the boundaries beyond which its cognitive system cannot operate. For example, a tick has a relatively small cognitive boundary, having very little memory or predictive power in the temporal direction, and sensing/acting very locally. A dog has much more temporal memory, some forward prediction ability, and a degree of spatial concern. However, it is likely impossible for a dog’s cognitive apparatus to operate with notions about what is going to happen next month or in the adjacent town. Human minds can operate over goals of vastly greater spatial and temporal scales, and one can readily imagine artificial (organic or software-based) Selves with properties that define every possible shape in this space (and perhaps change their boundaries over evolutionary and individual timescales). As will be seen in the next section, expanding the horizon is what enables information (in the Shannon sense) to acquire meaning, because data become causally linked to distant and past experiences, and acquires implications for future expectations. The formalism also suggests semi-quantitative definitions of maximum cognitive rate (the speed at which information propagates across an agent’s body); this and other similarities to the space-time diagrams of Relativity remain to be explored.

Clearly, much more work is needed to fully flesh out this rubric in a way that makes it immediately applicable in ethology, AI, and artificial life. One implication of this view is that there is not necessarily one unique primary level of organization; rather, Individuals existing at different levels in a given system can be putatively revealed by experiment and analysis that identifies allostatic set points - the goals of systems defined at different levels of organization and the spatio-temporal boundaries of measurements and actions taken by such mechanisms at each level. It is of course non-trivial empirical work to ascertain goals, especially for novel or alien ones. Moreover, systems and subsystems will have different, overlapping, and not always aligned goals. Thus, another key implication is that systems house multiple levels of coexisting coherent Selves.

Presumably, the volume of possible cognitive light cones keeps expanding with evolutionary time (although some lineages can contract it, permanently or during specific ontogenetic life stages). Over evolutionary scales, there have been “inflationary leaps” – innovations in body structure which drastically increase the cognitive boundary of viable selves. The formalism places no upper bound on the size of the cognitive light cone. Thus, while we understand “diminished capacity” for human beings in a legal and psychological setting, it is important to also consider “increased capacity”: what would a being with a much larger cognitive boundary than a typical Homo sapiens be like? It is interesting to consider whether this is the sort of outcome envisaged in classical traditions that posited evolution of human consciousness to a state of “Buddha-hood” – a cognitive boundary so large that such an individual would be capable of, for example, being directly concerned about the individual welfare of a myriad of beings. It is likely that efforts in artificial life, biomedical enhancement, or engineered AI will eventually give rise to much larger Selves, since it is unlikely that today’s human cognitive sphere is the maximum possible one.

Next, we consider a possible sequence of evolutionary scenario for the expansion of cognitive boundaries.

What evolutionary pressures lead cognitive boundaries to expand, resulting in the variety of agents observed in the biosphere? A sequence of phases can be hypothesized (setting aside the abiotic origin of chemical systems with feedback). I propose that the “atom” of this cognitive hierarchy is homeostasis (Bernard, 1865; Cannon, 1932; Wiener, 1961; Man and Damasio, 2019). The ancient origin of living forms is predicated on the ability of some simple systems to achieve coherence and distinction from the external environment, against perturbations, maintaining spatial and metabolic integrity by ensuring certain parameters stay in specific ranges (Maturana and Varela, 1980; Luisi, 2014). This homeostatic persistence is the origin of cognitive goals – the setpoints of subcellular biochemical circuits that react to perturbations in a way that maintains parameter range are the first, tiny examples of integrated goal-seeking systems. Homeostatic setpoints (actually, ensembles and not single points) are encoded by biophysical properties that guide activity toward implementing those states. The ability to operate toward a region in state space may be the primitive origin of complex cognitive systems that can entertain counterfactuals (remember or anticipate events that are not occurring right now).

Minimization of anti-homeostatic stress is a powerful driver, already suggested to be an important factor in evolutionary innovation (Erkenbrack et al., 2018; Miller et al., 2018; Wagner et al., 2019). A second step to the simplest homeostatic loop is the inclusion of a richer set of “hidden layers” (in the neural network sense) - additional biochemical nodes between the sensors and effectors of a given system that enable a degree of memory. By introducing a delay between inputs and outputs (Flack, 2012, 2017) and by including additional feedback loops that can maintain state after transient stimuli, a much more powerful homeostatic circuit is created, which uses memory of past events to anticipate, not merely react to, environmental challenges. The molecular basis of such anticipatory circuits, and biochemical machinery that responds to different frequencies in signals, has been described in single cells, such as ERK signaling dynamics and metabolic networks (Toettcher et al., 2013; Bugaj et al., 2018; Katz, 2018). Chemotaxis, such as seen in bacteria, is an early example of this because it exploits time delays in environmental sensing as a kind of memory to enable prediction/anticipation that enables it to optimize travel up nutrient gradients (Vladimirov and Sourjik, 2009; Scherber et al., 2012). The earliest forms of such memory are direct and not representational, e.g., slime molds learn to cross salt bridges to get rewards by storing the salt itself in their bodies as the engram of the memory used to guide future decisions (Vogel and Dussutour, 2016; Boussard et al., 2019).

More advanced creatures possess richer networks that allow complex forms of learning in which memories and goal states are implemented by settings (configurations) of internal mechanism several biochemical steps removed from the actual state in question (perhaps the origin of symbolic representation in more complex minds). Reactive homeostasis evolves into predictive allostasis (Schulkin and Sterling, 2019), under the pressure to predict signals from environment and other elements of the biosphere (Seoane and Sole, 2018). Importantly, memory can serve as the beginning of modularity because learning essentially groups diverse stimuli into compressed representations: complex states of affairs become remembered as compact biophysical engrams – this is the essence of the kind of modularity when a simple biophysical event kicks off the formation of a complex morphogenetic cascade such as building a “hand” in embryogenesis. In associative learning, some complex state gets functionally “hooked” to a trigger stimulus; this stimulus may be much simpler (in terms of information content) than the state it represents or the activity that it will kick-start. In somatic control networks, this is implemented as developmental modules – complex downstream morphogenetic activities that are initiated by a single driver. It has often been noted that modularity confers benefits for evolvability (Schlosser and Wagner, 2004; Wagner et al., 2007), but it is not always appreciated that like many other key aspects described above, it has an origin continuous with (and is driven by) control and proto-cognitive capabilities of the simplest living agents.

The recent discovery that very simple bioelectric states can trigger complex organ formation, such as induction of complete eyes in various parts of the animal (Pai et al., 2012), are consistent with a primary role of bioelectric signaling in both learning and morphogenetic control mechanisms (Levin, 2012).

The sensory machinery, implementing inputs to the cognitive agent, is extremely ancient. Complex animals’ sense of touch, taste, temperature, hearing, etc. are mediated by the exact same ion channel mechanisms that were already discovered by bacteria and used to make sense of their world (Tasneem et al., 2005; Blackwell, 2006; Liebeskind et al., 2012; Nilius and Honore, 2012). A rich sensory input layer, consisting of receptors and channels on a cell’s membrane enables the dynamics of Active Inference to operate (Friston et al., 2012; Pezzulo et al., 2015), and is itself highly regulated by cellular transcription and translational machinery in accordance with the main tenet of Perceptual Control Theory: behavior is the control of perception (Powers, 1973). Systems minimize surprise and optimize variational free energy by tuning their internal states in a way that optimally predicts future stimuli (Friston and Ao, 2012; Friston, 2013; Badcock et al., 2019). This is as true for nervous systems as it is for individual cells during metazoan development (Friston et al., 2015; Ramstead et al., 2019). This kind of predictive coding is data compression – a mechanism that works to represent a long history of environmental inputs into a compact representation via tuning of internal states. This basic learning and inference capability is the primal origin of “understanding” and intelligence in advanced brains, which can be defined as the process of inferring patterns from raw data and making “maps” of regularities in observations that occupies every Self from the simplest of organisms to the scientist working on extracting deep theory from data (Dingle et al., 2018).

The advent of multicellularity has been proposed to result from the drive to minimize surprise: a cell need only to surround itself with its progeny, in order to ensure a much more predictable milieu (the least surprising object in the world is a copy of yourself) (Fields and Levin, 2019). If these “front line infantry” are kept in place and suppressed from proliferating (and differentiated), one immediately gets the kind of arrangement seen across biology, from stem cell/soma systems to queen/worker dynamics in insect colonies. The surrounding body becomes an informational shield (Markov blanket) for the stem cell (Friston, 2013); adding epithelial and mesenchymal cell layers enables a kind of neural-like network in which each layer provides an increased level of abstraction from environmental stimuli, enabling deep tissues to perceive not simply raw signals but highly processed, dimensionally-reduced, compressed data that form a kind of primitive “knowledge” about the outside world’s patterns.

This transition from single cell state to multicellular tissue is a crucial step in the expansion of the cognitive boundary and the creation of more complex agents with larger goals (Figures 3, 4). By receiving information from neighbors, a cell in the center of a tissue can get (filtered) data about events occurring at a considerable distance. Because of the finite spread of signals through tissue and the predictive capacities of networks, this expands the cognitive horizon of a cell in both space and time. Crucially, by organizing into a (partial) electrochemical and informational pool or syncytium, all of the cells are able to measure and detect events occurring within the same boundary, creating a larger individual that emerges from the collective. The new Individual is an integrated whole because its subunits are no longer exclusively have their own local, distinct microenvironments (and thus internal models of the world) but are constrained to share a bigger, common reality due to the spread of information and influence among them. Such collective individuals can have a higher problem-solving capacity than their members, because they can support a layered architecture with experience-dependent communication channels (synapses, broadly defined), emergence of “virtual governors” with beneficial properties with respect to control capacity (Dewan, 1976), and more complex state space with more attractors and thus can compute meta-system properties not accessible to the single agents (Hofstadter, 1979; Crutchfield et al., 1998; Cenek, 2011).

It is proposed that morphological complexity and multicellularity are driven by greedy infotaxis (Vergassola et al., 2007) – collecting as much information as possible, from as far away (in both space and time) as possible, to minimize surprise and optimize prediction. The more a cell is connected to other cells in networks, the more processing capacity and the bigger the horizon of what the compound individual can potentially sense, remember, and store (Crutchfield et al., 1998). Merging into communicating networks also enables more noise tolerance and robustness – the field of artificial neural networks exploits the fact that some kinds of networks facilitate extracting patterns from data and can ignore irrelevant details while picking up on salient patterns. Inevitably, the scope of the Self expands with this integrated, growing cognitive domain. What molecular mechanisms might underlie this in cellular systems?

A fundamental facilitating step in this process is provided by developmental bioelectricity: the exchange of ionic signals within and among cells to propagate instructive influence at multiple scales of organization (Levin, 2012). Electric circuits are very convenient for processing information (a fact that has not escaped the attention of computer engineers and evolutionary mechanisms of neural systems), because they facilitate integration of information over spatial domains, and form feedback loops (such as voltage-gated ion channels) that readily implement memories and homeostats (Pietak and Levin, 2017; Cervera et al., 2018). This proposal, of ancient bioelectrical systems of coordination at the dawn of multicellularity, has now been confirmed by elegant studies of brain-like integration of spatial information by ion flows in the proto-bodies known as bacterial biofilms, which show how group-scale goals can emerge in simple physiological systems (e.g., nutrient sharing in the bacterial collective) (Prindle et al., 2015; Humphries et al., 2017; Lee et al., 2017). In more advanced cell types, this same scheme is implemented not by extracellular ionic waves but by intracellular electrical synapses known as gap junctions – highly regulatable valves that allow cells to selectively share bioelectric state with neighbors and implement activity-dependent plasticity (memory) (Palacios-Prado and Bukauskas, 2009; Zeng et al., 2009; Mathews and Levin, 2017).

Individual cells will all measure local conditions at their cell surface. When coupled into networks via gap junctions however, the whole collective can measure and act upon the same sensory data – the Umwelt expands to create a larger individual in which the comprising cells share a unifying picture of the world (Figure 4). It is thus no coincidence that most anesthetics used to remove cognition and sensory experience, whether in plants or animals, are gap junctional (bioelectric) uncouplers (Mantz et al., 1993; Wentlandt et al., 2006; Liu et al., 2012; Gremiaux et al., 2014; Baluska et al., 2016). Thus, physiological connectivity is the binding mechanism responsible for the appearance of larger unified Selves. The coordination of cells toward a single goal (body patterning) is now known to be in part controlled by the activity of bioelectric networks mediated by gap junctions (Levin et al., 2017; Mathews and Levin, 2018; Mclaughlin and Levin, 2018; Pietak and Levin, 2018). Consistent with the isomorphism between patterning control and behavioral control, the same scale-up occurs in morphogenesis: subcellular chirality that determines morphology of single-cell organisms is amplified during early development by a system of gap junctions and ion channel-driven bioelectrical cues into animal-wide axial patterning (the appearance of a new anatomical level – “whole body symmetry”) (Levin, 2003, 2006).

The developmental control bioelectric network shares mechanistic evolutionary history with the nervous system, and the conservation extends not only to mechanisms of targeting connections [such as neural-like mechanisms deployed in plant pollen tubes (Palanivelu and Preuss, 2000)] but to the ubiquitous use of neurotransmitter molecules downstream of bioelectric driving forces [e.g., GABA and the serotonin-like Auxin used in many plants (Ramesh et al., 2015)]. An intermediate form between somatic bioelectric networks and neural networks sheds light on how local gap junctional connections in tissues were refined into more targeted connections on the scale of several cell diameters: tunneling nanotubes. This axon-like structure present in many cell types is an intermediate cellular appendage that has a gap junction at its end, allowing limited directed connections between cells in a tissue, and mechanistically presaging neural axons and their electrical synapses (Wang et al., 2010; Sherer, 2013; Ariazi et al., 2017). The general bioelectric system was speed-optimized during the development of nervous systems for behavior (Van Duijn et al., 2006; Keijzer et al., 2013; Jekely et al., 2015), but uses fundamentally the same set of mechanisms to optimize the input-output relation within and between the internal milieu mechanisms and the outside world (Molnar-Szakacs and Uddin, 2013; Fields et al., 2019). Increasing the correlation length of such control could have driven the development of targeted, long cellular structures (axons): to move the whole animal, not just individual cells, the electrical information needs to be scaled considerably, and nervous systems are an ideal extension of ancient, pre-neural bioelectric signaling paths to enable further scaling of the cognitive boundary of cells into organs, organisms, swarms, and societies.

A prediction of this proposed scaling is that swarm organisms should fall prey to the same kinds of cognitive illusions and specific failures of rationality as do vertebrate brains, which has indeed been observed in ants and slime molds (Sakiyama and Gunji, 2013, 2016; Tani et al., 2014; Beekman and Latty, 2015). Another consequence is that the cybernetics of associative learning in networks is agnostic as to the spatio-temporal scale and physical implementation, being widespread from molecular networks and inorganic physics to whole evolving populations (Cragg and Temperley, 1955; Mcgregor et al., 2012; Power et al., 2015; Watson and Szathmary, 2016; Kouvaris et al., 2017; Lopez Garcia De Lomana et al., 2017).

“The self is not something ready-made, but something in continuous formation through choice of action.”

A number of ideas from cognitive neuroscience, information theory, computer science, cybernetics, and engineering converge on key questions in biology with respect to the multi-scale interface between body structure and mind. The scheme described above seeks to (1) define Individuals and Selves in a way that facilitates taxonomy, comparison, and communication with evolved, created, biological, artificial, and exo-biological agents, and (2) propose a plausible naturalistic framework for the evolutionary scale-up of cognition from earliest origins of life, hypothesizing about the forces that drove it and the major transitions along the continuum. The goal of this research program is to show how complex agency and goal-directedness evolves naturally from ancient mechanisms. The evolutionary pressure to survive in a challenging world leads (in order) from simple homeostasis to infotaxis, memory, anticipation, spatio-temporal scale-up of measurement and prediction, and large-scale global goals (system-level agency). The implications of these hypotheses extend beyond philosophy and evolutionary biology. Practical strategies for regenerative medicine (control of cell collectives in vivo toward macrostates such as “healthy structure and function”), exobiology, and robotics/AI are impacted by our view of what defines a coherent Self.

A unified, integrated cognitive Self or Individual can be defined with respect to the integrated ability to pursue specific goals via a homeostatic process that resists perturbations. Goals also define positive and negative reinforcement for that agent, thus enabling communication with/training of highly diverse intelligences. The ability to pursue goals is a major ratchet for evolution because it smooths the selection landscape: the potentially destructive effects of individual mutations are often made up for by the regulative ability of other mechanisms to accomplish specific outcomes despite changes of circumstance (e.g., cells providing blood vessels and tendons to make a coherent finger when a new hand bone is induced).

An Agent’s cognitive world can be quantified and characterized, enabling comparison with others (regardless of their material implementation), by estimating the spatio-temporal boundaries of its area of concern: the volume in space and time over which the agent is able to take measurements, exert influence, and functionally link disparate events (learning, association).

The borders of the temporal and spatial events of which a given system is capable of measuring and acting map out a “cognitive light cone” – a boundary in the informational space of a mind. These borders can grow or shrink, on evolutionary or ontogenic time scales, as the organization of an agent changes. The key is a balance of selective information sharing, via “synapses” – structures of arbitrary physical construction which share the feature that they can regulate the passage of signals based on the state of other such elements. Too little sharing results in a failure to bind subunits into a new Self. Too much sharing results in a homogenous soup with insufficient differentiation of modules and abstraction of information across distinct layers.

Cancer is a (reversible) shrinking of the computational boundary of a biosystem: by isolating itself from the surrounding tissue’s physiological signals, a cell’s cognitive boundary shrinks to the small size it had before multicellularity. Cancer cells are not more selfish than somatic metazoan cells – they are equally selfish but their Self is now scaled down to a single cell (which will reproduce and migrate as much as it can), whereas the normal physiological binding in healthy tissues binds each cell to a common goal represented by the large network – the construction and repair of a specific large anatomy. These ideas connect naturally to gene-level views of selfishness (Dawkins, 1989); if it is useful to think of genes as selfish agents, it is doubly plausible to view cells and tissues as such, since the latter have much more capacity for activity and computation. Future work will develop, in the contexts of ontogeny and evolution, how the optimization of specific states of affairs (pursuit of goals – selfish or otherwise) occurs simultaneously at many scales of biological organization.

Agents scaled up by evolving from basic homeostatic loops, driven by active inference (surprise minimization) via addition of delays (memory), anticipation (inference), and networks (spatially-distributed processing that enables learning and progressive abstraction/generalization from data). Gathering into larger collectives with optimal informational structure (Klein and Hoel, 2019) improves the computational (predictive) capabilities and gives rise to functional relationships (memories, encoded goal states, test-operate-test-exit loops) that exist over and above any individual member (Miller et al., 1960). These levels coexist, enabling numerous coherent Selves of different scales to be implemented by any collection of living matter.

Infotaxis (the drive for better actionable intelligence about the regularities/patterns in the world, and in the agent’s own mechanisms) encourages cells to connect in groups via signaling. On the cellular level, this is implemented by diffusion (bacterial biofilm proto-bodies) or direct connections via gap junctions or neurons.

Collecting into a syncytium enables all of the cells to share the same data and access the same memories (illustrated, e.g., by the ability of trained slime molds to pass on information to naïve hosts by fusion [Vogel and Dussutour, 2016)], This shared information structure extends to the edges of the large collective, which binds small, individual competent sub-agents into a larger unified Self. These principles likely apply beyond cells in organs, to swarms of whole organisms such as bees and termites (Seeley, 2009; Turner, 2011), as do the dynamics of breakdowns in coordination which share important similarities for example between cancer and social insect colonies (Amdam and Seehuus, 2006).

The hypothesis of scale-free cognition does not rely on cooperation per se – it builds up apparent cooperation from selfish agents minimizing their stress (surprise) and competing for information. Greed, at the single-cell level, for information (infotaxis) drives cooperativity, as each unit expands its measurement boundary (communication with neighbors) and thus inevitably becomes part of a bigger self with bigger set points serving as homeostatic attractors. It only looks like cooperation from a perspective of a higher level, because the higher level of organization shows an integrated Self which appears, necessarily, cooperative.