Competency in Navigating Arbitrary Spaces as an Invariant for Analyzing Cognition in Diverse Embodiments
The tight link we have developed between motion in spaces and degrees of cognition across scales suggests that it may be possible to develop models of evolutionary search itself as a kind of meta-agent searching the fitness space via active inference and other strategies [62,63,215,216]. In this light, evolution is still not claimed to be a complex meta-cognitive agent that is knowingly seeking specific ends, but on the other hand, it may not be completely blind either. It may be possible to develop models of minimal information processing that better explain the ability of the evolutionary process to solve problems, to choose which problems to solve, and to give rise to architectures that not only provide immediate fitness payoffs but also perform well in entirely new environments.
A key opportunity for new theory concerns what tools could be developed for a system to detect that it is part of a larger system that is deforming its action space with nonzero agency. It may not be possible due to the Gödelian limits for a system to fathom the actual goals of the larger system, of which it forms a part, but how does an intelligent system gain evidence that it is part of an agent with some “grand design” versus living in a cold, mechanical universe that does not care what the parts do? The Lovecraftian horror of catching a glimpse of the fact that one is a cog in a grandiose intelligent system may be tempered by mathematical tools that enable us to have more agency over which aspects of the externally applied gradients we wish to fight against and which gradients we gladly roll down.
We foresee great promise in the application of the mathematical framework of category theory [217,218], which provides the conceptual and formal tools needed to model the relationships between arbitrary spaces. Any of the spaces discussed here, together with the search operations acting within that space, can be considered a category. The theory provides, in this case, rigorous tools for determining whether multiple paths through the space yield the same outcome and, even more interestingly, whether paths through different spaces, such as a path in a morphological space or a path in a 3D behavioral space, yield the same outcome. We defer such analysis to future work. Some preliminary steps in this direction, characterizing arbitrary QRFs as category-theoretic constructs, can be found in [170,178].
There are numerous analogies to be explored with respect to porting conceptual tools from relativity to study scale-free cognition. The use of cognitive geometry and infodesics [219] ties naturally to general relativity. Other examples include the following:○Gravitational memory (permanent distortions of spacetime by gravitational waves [220]) to link the structure of action spaces to past experience;○Inertia in terms of resilience to stress (anatomical homeostasis as a kind of inertia against movement in the morphospace and other spaces);○Acceleration and force in a network space, where every connection in a network could be modeled via a “spring constant” or, even better, an LRC circuit. With feedback, interesting oscillations can appear, which can be harnessed as computations;○The ability of one system to warp the action space for another, such as warping the morphospace for the embryonic head by specific organ movements, generates an analog of “mass”;○Bioelectric circuits could be modeled as warping the morphospace in the same way wormholes warp physical space. The two points at opposite ends of a wire are, for informational purposes, the same point, even if they are on opposite sides of the embryo. Neal Stephenson stated, “The cyberspace-warping power of wires, therefore, changes the geometry of the world of commerce and politics and ideas that we live in’’ [221]. The gap junctions’ control of morphogenetic bioelectric communication deforms the physiological space to overcome distance in the anatomical space. Neurons do this too, as do mechanical stress in connective tissue and hormones;○Links also could be made to concepts of special relativity. For example, doppler effects in morphogenesis have already been described [222].
Moreover, the limited speed at which information can propagate through tissue naturally defines a minimal “now” moment, a temporal thickness for the integrated agent below which only submodules exist, in effect illustrating the relatedness of space and time by the propagation speed of information signals within living systems.
Specific models of morphogenetic control (embryogenesis, regeneration, cancer, etc.) that rely on navigation policies with diverse levels of cognitive sophistication need to be created and empirically tested. Can craniofacial remodeling be understood as a “run and tumble” strategy? Can evolution of morphogenetic control circuits be understood as the evolution of abstract vehicle navigation skills, thus porting knowledge from evolutionary robotics and collective intelligence to developmental biology [157,223,224,225]?
Similarly, such models need to be developed to understand allostasis in transcriptional, metabolic, and physiological spaces, modeling and then developing minimal Braitenberg vehicles [44,226,227,228,229] as real devices to implement biomedical interventions such as smart insulin and neurotransmitter delivery devices.
Regenerative medicine needs to be moved beyond an exclusive focus on the micro-level hardware (genomic editing and protein pathway engineering) to include interventions at higher levels. Using tools from behavioral science such as training in various learning assays can manipulate the lower-dimensional and smoother space of tissue- and organ-level incentives (described in more detail in [35,36]). Much as evolution exploits multi-scale competency to maximize the adaptive gains per change made, bioengineers and workers in regenerative medicine can take advantage of behavior shaping of cellular agendas and plasticity, working in a reward space. Interestingly, this was well-appreciated by Pavlov, whose early work included training animals’ organs in addition to the animals themselves. He understood the physiological space, and his experiments on training the pancreas and other body systems can now be performed with much higher-resolution tools. More broadly, impacting and incentivizing decision-making modules at higher levels is much more likely to produce coordinated, coherent outcomes than interventions at lower levels [230], resulting in fewer side effects in pharmacology and avoiding unhappy monsters in synthetic bioengineering. The future of biomedicine will look much more like communication (with unconventional intelligences in the body) than mechanical control at the molecular pathway level. This includes signaling to exploit the control policies of cells in the morphospace for regenerative control of growth and form [35,36] and exploiting gene-regulatory networks’ abilities to learn from experience to modify how they move in the transcriptional space while healthy and in the case of disease [79,83,231,232,233,234,235].
Computer engineering and robotics also afford many opportunities for testing and applying this framework. Incorporating biological concepts into a computing system design has been explored in the abstract [236,237], at the level of system design [238,239], and with neuromorphic hardware [240,241]. The present work suggests further directions, including developing frameworks for working with agential materials (like the cells that make up Xenobots), which requires distinct strategies from those used with passive materials or even active matter [242,243,244,245], creating evolutionary simulations and human use tools to explicitly address multiple scales of organization and problem solving.
More broadly, artificial intelligence can benefit from enhancing current neuromorphic approaches with systems based on much more general, ancient intelligence, creating systems with motivation and agency from the ground up by taking embodiment seriously from an evolutionary perspective. The classic Dennett and Minsky debate about how much real-world embodiment matters for artificial intelligence can now be reframed in more general terms: embodiment is critical indeed, but it does not have to be in the classic 3D space. Embodiment in other action spaces can drive the same intelligence ratchet described above. New general AIs are likely to be developed gradually from minimal systems driven by the dynamics described above, which eventually scale homeostatic action into advanced metacognition. One specific strategy that can be suggested is the creation of an unsupervised agency estimator, which seeks to make models of its environment anywhere on the spectrum of persuadability [9]. This system will not only be useful for human scientists (freeing their hypothesis-making from the mindblindness [246] that limits imagination with respect to unconventional intelligences); it can also be used in an “adversarial” mode with evolving intelligences, a cycle that increasingly potentiates both the intelligence and the ability to detect it.
Human beings evolved with conscious access mostly to data from the outside world, including sensitivity to only a very small slice of the myriad of actions occurring in their various physiological, cellular, metabolic, and morphogenetic control systems. As a result, while our cognition is finely tuned to identify levels of agency in the behavioral actions of objects moving in a 3D space, we are intrinsically bad at recognizing intelligence in unfamiliar guises. If, for example, we grew up with a keen internal sense of our blood chemistry and all the things our pancreas, liver, kidneys, etc. were doing to maximize our health, or if we could directly sense changes in gene expression, we would have no trouble recognizing these as agents exhibiting competency and degrees of intelligence in other spaces. Thus, it is essential to develop a substrate- and scale-invariant theory of agency and intelligence. Specifically, we have sketched such a theory while maximizing the empirical, testable, and practical implications over philosophical wrangling. We have shown in particular that living systems operate in multiple spaces at different scales. These include the transcriptional, physiological, and morphological spaces as well as the more familiar 3D behavioral and social spaces. The computational mechanism of VFE minimization drives behavior in these spaces toward attractor states that enable allostasis. This framework suggests a number of new theoretical and experimental approaches in both biological and hybrid biological-artificial systems.
Our model is committed to an observer-dependent non-binary approach that takes evolution and developmental biology seriously to emphasize gradual origins, deep unification of basic principles, and ubiquitous real-time change. Policies and mechanisms guiding such seemingly diverse behavior as magnets’ “mindless” movements to reach each other, biochemical networks moving down energy gradients, bacteria swimming up nutrient gradients, moths’ repeated attempts to reach a light, and human goal-directed behavior must be on the same continuum because modern biology offers no discrete magical event that separates them; a single framework is needed. In this light, all intelligences are collective intelligences, and biological systems are nested dolls of agents with agendas that cooperate, compete, communicate, and interact within and across levels of organization. Agents of highly diverse implementation model their environments and themselves in accordance with an active inference framework, which drives the way they navigate information spaces. While dynamical systems theory describes how a system can go through a defined space, multi-scale agency models explain the shape of the space relative to specific observers and agents [247].
A key aspect of life is top-down control, where higher levels deform action spaces so that lower levels can be less intelligent and more mechanical while enabling the larger system to occupy the more adaptive regions of various spaces. The system evolves via a ratchet mechanism that begins with simple homeostasis and takes advantage of its modularity to measure, remember, and act over progressively larger and more complex states. Evolution pivoted this basic trick across spaces from the metabolic and physiological spaces through the anatomical morphospace, where ancient bioelectric network mechanisms became used to propel animals through a 3D space using the same strategies they originally relied on to move their anatomical configuration through morphospace during regulative development and regeneration.
The “brain in a vat” dynamic, where agents have no access to the objective ground truth about their brain–body–environment relations but must build actionable models of themselves on the fly, helps understand (and design) highly diverse systems in which all of these aspects can be changed in a modular way: by evolution and by engineers. This helps clarify why biological systems are so highly evolvable and dissolves the existing concepts of objective privileged viewpoints of boundaries between the self and the world. These ideas have numerous implications for understanding the origin of various control systems in neuroscience and the efficiency of evolution in creating extremely robust problem-solving machines. This in turn suggests the same kinds of strategies (targeting high-level reward spaces) that are useful for workers in biomedicine, AI, and synthetic bioengineering who seek to manipulate and build complex adaptive systems.
We view this framework as only the beginning of an empirically grounded understanding of agency that provides a conceptually integrated picture of the world. It is crucial to develop such frameworks that abandon untenable binary distinctions in favor of predictive, actionable models that are compatible with the modern understanding of gradual evolution and developmental biology. Forthcoming advances in synthetic bioengineering and AI will result in a diversity of agents in our midst that will dwarf Darwin’s challenge to classical categories. Future developments in transformative regenerative medicine, automation, and ethics require a firm conceptual foundation for understanding agency in a physical world.
We thank Dora Biro, Pranab Das, Karl Friston, Edward Harvey, Benjamin Levin, and many other members of the community for helpful discussions and Wesley Clawson for comments on a draft of the manuscript. We thank Julia Poirier for editorial assistance with the manuscript.
Conceptualization: M.L. and C.F. Writing and editing: M.L. and C.F. All authors have read and agreed to the published version of the manuscript.
M.L. gratefully acknowledges the support of the Templeton World Charity Foundation (grant TWCF0606) and the John Templeton Foundation (grant 62212). The opinions expressed in this publication are those of the authors and do not necessarily reflect the views of the John Templeton Foundation.