We are all collective intelligences – not just ants and bee colonies, but all of us, because we are made of active parts (cells), which have to become aligned to work together to give rise to a system that has memories, preferences, and goals that don’t belong to any of them. The collective organism operates in new problem spaces to which its parts do not have access (such as anatomical morphospace). This emergence of higher-level Selves, through the actions of components, is one of the most fascinating questions of science and philosophy: how do integrated Selves, which are more than the sum of their parts, arise? Where do their goals, preferences, and competencies come from? Our lab is working on these questions, as well as on the origin of novel systems’ goals (ones that couldn’t have been directly set by evolutionary selection in the case of synthetic proto-organisms). We also work toward tools for detecting, understanding, communicating with, and constructing such emergent Selves. You can see some of that discussed in my recent talks where I describe how we use the ability of cellular collectives to achieve specific anatomical shapes as a model system to understand collective intelligence. Groups of living cells navigate the space of possible anatomical outcomes (morphospace) by using their genetically-provided toolkits to achieve the correct species-specific target morphology even when circumstances, or their own parts, change.
Collective intelligence and cognitive glue
Here I want to talk about a key question in this field – the nature of cognitive glue: what policies, properties, and features of cells and various laws (physics, computation, etc.) enable this binding toward common purpose? A cognitive glue mechanism is one that aligns subunits (e.g., cells) toward the same model of themselves, of the outside world, and of their goals, enabling them to work together to reach that goal reliably and despite novel circumstances, perturbations, etc.
Crucially, cognitive glue mechanisms enlarge the size of goals that a system can represent and re-shapes the action landscape toward common interests in a way that induces cooperation and effective alignment. In the case of our most apparent cognitive Self, which drives conventional behavior in 3D space, we emerge via the functioning of a collection of neurons in our central nervous system. Here’s an example illustrating this.
The rat has been trained to press a lever and get a delicious pellet as reward. But note that no individual cell has both experiences: the cells at the bottom of the paws touch the lever, the cells in the gut receive the sugar reward – who owns the associative memory between the two events? The “rat”, held together by (among other things) neural networks that bind a huge number of cells into a coherent whole that can own memories (and goals, and preferences) which none of them individually have.
Morphogenesis as collective intelligence
What about morphogenesis – the process by which complex bodies self-assemble? Embryogenesis, regeneration of organs (in species fortunate enough to be capable of it), and on-going suppression of cancer and aging, all require cells to work together to reliably reach and maintain a specific target morphology – a tight anatomical specification. The most remarkable and salient aspect of this process is not its fidelity however – it’s the ability of morphogenesis to reach the same target morphology despite a wide range of novel scenarios, interventions, and surprise changes of external environment and internal components). Here is a talk discussing how we use morphogenesis as a model system to learn to detect, manipulate, and create collective intelligences that operate in non-obvious problem spaces. When cast as navigation of the anatomical morphospace, it becomes clear that morphogenesis is actually the behavior of a collective intelligence; this insight has enabled our lab (and now others) to make many new discoveries with implications for many fields from AI to cognitive science to biomedicine. Enormous advances in regenerative medicine and in our understanding of evolution will be reaped when the mechanisms, capabilities, and communication interfaces to this somatic intelligence are characterized.
To really understand the need for alignment in morphogenesis, consider an embryonic blastoderm – a sheet of tens of thousands of cells we call “an embryo”. What are we counting there – what is there one of? What we’re really counting when we look at a sheet of cells and see an “embryo” is commitment to a world-model: the fact that all of the cells are aligned to the same goal – the same journey in anatomical space that they will do their best to undertake as a collective. But that glue is easily interrupted – mechanical or physiological barriers placed within the blastoderm (such as scratch wounds shown in the lower series of panels below) separate it into multiple individuals, each of which will take their own journey and become separate beings (conjoined, or not, twins, triplets, etc.). So how many individual beings can come forth from the excitable medium of an embryonic blastoderm? 0, 1, 4 or possibly more potential selves. The number of selves and the borders between distinct selves and the outside world are defined by the workings of specific cognitive glue mechanisms (this also has fascinating parallels with dissociative identity disorders of the cognitive intelligence, in terms of fragmentation and unification of individual goal-directed intelligences within a substrate).
The same issue comes up when trying to set up boundaries between regenerating structures (indeed, all of development can be seen as regeneration of all of the body’s structures from 1 cell). So how do cells control their internal parts (molecular networks) and all act together to, for example, regenerate the right number of fingers in an amputated salamander limb, when no individual cell knows what a finger is? That kind of creative problem-solving in multicellular bodies requires the tiny cognitive light cones of their cells to be merged together into a much larger one that subsumes massive construction projects (like, “build a limb of this shape and size”) instead of the much smaller physiological and metabolic goal states pursued by individual cells. On the developmental and evolutionary time scales, cells become connected into networks (e.g., bioelectrical) that can store larger goals by providing extended memory, anticipation, and representation capacities. Here is a diagram that shows the limited size (in time and space) of single cells vs. cellular collectives and their goals:
Memory anonymization as a kind of cognitive glue
How are cells orchestrated toward specific anatomical outcomes as goal states that, once achieved, cause their activity to cease? In other work (Figure 9 here) we focused on memory anonymization as one kind of cognitive glue:
The idea that gap junction (electrochemical synapses) connections between cells make it harder for them to keep individual identities (memories) because physiological events that serve as memory engrams propagate from one cell into its neighbor, which then can’t tell whose memory it was. It’s not just a false memory for that 2nd cell, it’s a true memory for the collective. Breaking down the informational boundaries between cells results in a partial collective mind where the individual identity partially disappears in favor of group dynamics and their informational mind-meld.
Gap junctions are part of the bioelectric networks governing growth and form. But they are not just another piece of biophysics that one needs to take into account in developmental biology. I have long argued that electrophysiological phenomena, which hold together our familiar behavioral Self evolved from older mechanisms that performed precisely the same function for our morphogenetic self. In other words, bioelectricity is a powerful mechanism for implementing cognitive glue that underlies integration of information across space and time, from bacteria to embryos to brains. But it is likely not the only one, and we are now studying another one – stress sharing – that may or may not have any bioelectrical component.
Stress sharing as cognitive glue
More recently, we began testing the hypothesis that another kind of cognitive glue in morphogenesis is stress sharing, and that stress in morphogenesis is mediated by the same ancient genetic components that were first used to detect abnormal states in microbes and unicellular organisms.
The fundamental process of reliably reaching a specified target morphology, starting with a single cell (egg), different body (metamorphosis), or damaged body (regeneration), and stopping when it is achieved, is a process of error minimization. It’s a loop, like any homeostat, in which the system continuously measures its progress relative to a setpoint within the action space and tries to reduce the delta between current state and final state (our lab has been studying for maby years how this final state is a pattern memory encoded in bioelectrical properties of cells, but other mechanisms could be involved as well). The important thing is that unlike in a thermostat, where the system is made of passive components and can be engineered with an explicit goal, life consists of an agential material – cells with their own agendas, which must be explicitly coordinated toward a specific goal in order to dominate their innate agendas (the story of cancer). Specifically, many problems (such as regeneration, repair, etc.) must involve coordination of many cellular events spread throughout the body, such as making sure that the various parts of a complex structure match in terms of size, orientation, location, symmetry, etc.
Individual cells are pushed through their homeostatic paces by error estimates – measurements (sensing of internal states) that tell the system how far off from desired setpoint it is. When the error falls below an acceptable limit, they can stop adjusting. That works fine for single-cell level goals (like keep the local pH to a certain level) but morphogenesis requires being motivated toward very large-scale states: does this Axolotl limb have the right number of fingers? etc. How to get all of the relevant components to care about the same problem?
We conjectured that stress – a system-level signal that reflects distance from setpoint but is not localized to a single locale, can be used to engage distant bodyparts into coordinated action. Let’s imagine how it might work. The following example concerns geometric positioning of cells, but the basic principle can hold in all kinds of spaces – physiological, gene expression (transcriptional), etc.
In this simplified diagram of one typical scenario during morphogenesis, stress sharing is hypothesized to help solve the problem. Cell I (red) is motivated to move upward until it reaches its correct position in a morphogen gradient. But cell I would not get to its intended spot because all the other cells’ target states are met and they have no incentive to move enough to let cell I pass, no matter how much stress cell I is feeling due to the error state of its current position – they are content, their own stress is low, they won’t move. Cell I’s problem is not their problem and a permanent defect will remain.
But if the molecule(s) whose level represent the distance between its positional setpoint and its current position can leak (or be exported) our of cell I, they will affect nearby cells. The leaking, propagating stress signal indicates a problem somewhere in the collective. But since stress molecules are conserved among all cells within a given body, the neighboring cells cannot tell that the problem is not their problem – they are motivated to reduce stress, and the presence of the conserved stress-carrying signal affects them directly. They will be stressed by their microenvironment, which raises their plasticity – reduces their threshold for moving and making changes and motivates them to act in order to reduce their own stress. In this example, they become more willing to move around (to find a better position for themselves), which enables them to let cell I pass through. When it gets to the correct position, its stress levels reduce, it stops generating and propagating its stress signal, and the whole cell sheet’s stress levels reach low levels because the entire collective is in a low-energy state with respect to all of their setpoints.
In this way, a systemic stress response mechanism, implemented by conserved stress molecules that leak from specific cells, could facilitate large-scale problem-solving by de-localizing incentive to work toward group goals (cooperation without built-in altruism, since all the cells are simply trying to reduce their own local stress levels). The leak mechanism enables the collective to bend the energy landscape for the individual cells to exploit their homeostatic optimization behavior towards a tissue-level patterning goal. The best thing about a stress sharing mechanism is that it doesn’t rely on altruism: by leaking stress, my problems become your problems. You are motivated to help me because it reduces your own stress to do so, you do not have to have a special mechanism for helping others. It’s a kind of emergent collective coordination that can evolve simply by making stress indicator molecules leaky.
Seems to make sense conceptually, but would this really work?
Testing the stress sharing hypothesis in simulations
Lakshwin Shreesha (also here) and I tested this idea in a computational model of developmental biology (in silico embryogeny), described in the final paper here and in an open preprint here.
Fig. 1. Schematic of experimental setup
(A): A simplified diagram of one typical scenario during morphogenesis in which stress sharing is hypothesized to help solve the problem. Cell I (red) is motivated to move upward until it reaches a specific position in a morphogen gradient. (A1) Before and after, in the absence of stress sharing: cell I will not get to its intended spot because all the other cells’ target states are met and they have no incentive to move enough to let cell I pass, no matter how much stress cell I is feeling due to the error state of its current position. Thus, a permanent defect will remain. (A2) Before and after, in the presence of stress sharing: if the stress molecule(s) representing the delta between its gradient setpoint and its current position can leak (or be exported) our of cell I, they will affect nearby cells. Since stress molecules are conserved among all cells within a given body, the neighboring cells will be stressed by their microenvironment, which raises their plasticity and enables them to let cell I pass through. When it gets to the correct position, its stress levels reduce and the whole cell sheet’s stress levels reach low levels because the entire collective is in a low- energy state with respect to all of their setpoints. In this way, a systemic stress response mechanism, implemented by conserved stress molecules that leak from specific cells, could facilitate large-scale problem-solving by de-localizing incentive to work toward group goals (cooperation without built-in altruism, since all the cells are simply trying to reduce their own local stress levels). The leak mechanism enables the collective to bend the energy landscape for the individual cells to exploit their homeostatic optimization behavior towards a tissue-level patterning goal. (B): Process of cell movement with and without stress sharing in embryos: The target task was for cells within a randomly initialized 2D matrix/embryo to re-arrange themselves towards a pre-set target pattern (a smiling face binary image in our setup). Stressed cells (B.I) (i.e, cells which are not in their correct positions to form the pre-set morphogenetic pattern) were allowed to move by picking other stressed cells as target and moving towards them through local swaps. By default, a stressed cell was not allowed to disturb other fixed cells (cells which were already in their correct positions). However, when given the ability, stressed cells could share their stress with their fixed neighbors, urging them to provide a path for movement towards the target (B.II). Consequently two kinds of competencies were setup: One with stress sharing (B.II), and one without (B.IV). In the case where sharing was not allowed, a stressed cell surrounded by other fixed cells would have no way of moving through, causing it to get blocked locally (B.IV). In the case where sharing was allowed, blocked stressed cells would spread their stress to an area of size 3×3 around them (shown in orange), thereby communicating an intent of movement; in response, a channel would be created by the fixed cells for the stressed cell to move towards the target location. (Stages I, II, and III). With each resolved cell movement, the stress map was updated (B.III). Such a process occurred iteratively until either the competency limit was reached or the target pattern was formed.
Here is, the evolutionary cycle:
Fig. 2. Genetic algorithm schematic
(A): Initialization: Prior to evolution, a population of genomes which served as the substrate for evolution were initialized. A genome in our framework was a randomly initialized two-dimensional matrix containing two cell-types (0, indicated in black, and 1, indicated as white) (A.I). The genome also carried a marker indicating the type of development an embryo could undergo during the first stage of evolution i.e: development. Three kinds of reorganizaton processs were defined: 1. Stress sharing based reorganization, 2. Reorganization without stress sharing, and 3. Hardwired (no reorganization). A target for evolution was also set. To this end, we chose a downsampled binary image of a smiling face as the target pattern which evolution had to optimize over (A.II). (B): Development: The first stage of evolution involved a process of cellular reorganization of the initialized matrices. Reorganization would be based on the gene marker encoded in the genomes of each embryo. This stage served to mimic the developmental stage observed during an organism’s lifetime. Post development, the reorganized cell-structure was termed the embryo’s phenotype. Thus, the developmental process served as an independent layer separating the genotype of an embryo from its phenotype. (C): Selection: Prior to development, the fitness of each genome was recorded by comparing it with the target (l2 − distance). And Post development, the fitness of the corresponding phenotype was also recorded using its l2 − distance with the target. During selection, phenotypic fitnesses of the top 10 % of the population were chosen and, their corresponding genomes were selected (Darwinian selection) for evolution in the next generation and the rest were discarded. (D): Mutation: The selected for genomes constituted 10 % of the original population. To repopulate the population back to its original strength, selected genomes were mutated by randomly swapping cell- pairs within the genome. The resulting mutated genome served as a new embryo. Post re-population, a new cycle of evolution began marked by the onset of development.
The basic experiment compared the ability 3 different kinds of simulated embryos to complete morphogenesis (we’re using the “smiley face” target because this is a basic pattern for the frog embryo’s electric face). Hardwired embryos (black line in the graph below) had a direct mapping from genotype to phenotype – like nematodes (for example), it was a very fixed kind of development which always went the same way. Other embryos had a more realistic developmental process between genotype and phenotype, which was able to make decisions in anatomical space and execute homeostatic ability to keep working until the overall anatomy was correct. When not sharing stress (controls, orange line), they started off well but were eventually overtaken by the direct evolution of the hardwired ones’ genomes. The best performance was seen in the homeostatic embryos whose cells could share stress – they achieved perfect morphogenesis earlier than others (green), and had the best outcomes.
Thus, we found that stress sharing improves the morphogenetic efficiency of multicellular collectives. Moreover, we found that stress sharing influenced the future fate of distant cells in the multi-cellular collective, enhancing cells’ movement and their radius of influence, consistent with the hypothesis that stress sharing works to increase cohesiveness of collectives. Finally, we found that anatomical goal states could not be inferred from observation of stress patterns (there was no obvious 1:1 mapping of the stress pattern among the cells and structure they were trying to build), revealing the limitations of knowledge of goals by an external observer outside the system itself.
What’s next:
Interestingly, one thing we observed was that the effective “circle of concern” – the radius that individual cells tried to manage – was increased by stress sharing – they began to actively try to manage (i.e., care about) bigger domains of their world. To make an extension from morphological to cognitive spaces: the cognitive light cone (the size of the biggest goals an agent can work towards) is intimately tied to the intelligence. Tell me the size of things that stress you out and I can roughly guess what kind of mind I’m dealing with. If you only care about the sugar concentration over minutes in a few micron-sized area, you might be a bacterium. If you are stressed out by events in a few hundred yard radius but can’t possibly worry over what happens several miles away 3 weeks from now, you might be a dog. If you’re stressed by the sorry state of the planetary climate cycles long after you will be dead, you might be a human. However, if you can really (i.e., actively and in the linear range) care about trillions of other sentient beings, you are not human but something greater. In this paper we discuss the connection between intelligence, cognitive light cones, compassion, empathy, and related concepts, which is an area that needs to be developed further because of its implications for alignment of AI and other future beings.
In the meantime, next steps are to test these ideas in real morphogenesis in the lab. We’ve got some good candidates in mind for the stress-bearing molecule and we’re going to study how they up- and down-regulate during specific scenarios, and how inhibiting their propagation (or. preventing cells from caring about their levels) affects morphogenesis. Actually we already have some pretty cool preliminary data on this; stay tuned!
Featured image by Midjourney.
Other images and diagrams by Jeremy Guay of Peregrine Creative.
Leave a Reply