Answering Schrödinger's question: A free-energy formulation

A node, defined as an interacting unit of a network, is itself a network composed of smaller nodes interacting at a lower hierarchical level, producing nested neural networks that extend from neurons, to macrocolumns, and to macroscopic brain regions. From [72]. (B) A simple cortical hierarchy with ascending prediction errors and descending predictions. Superficial pyramidal cells (red triangles) compare expectations (at each level) with top-down predictions from deep pyramidal cells (black triangles) at higher levels. Neuromodulatory gating or gain control (blue) of superficial pyramidal cells determines their relative influence on the deep pyramidal cells that encode expectations by modulating their precision. From [93]. (C): The variational neuroethology of human cognition and biobehaviour. F(s˜(a),μ(i)|m(i)) represents the free-energy of the sensory data (over time), s˜(a), and the states μ of an agent m(i) ∈ s that belongs to a subgroup s ∈ c of class c. Action (a) governs the sampling of sensory data, and the physical states of the phenotype (μ) encode beliefs or expectations (and expectations of the mean of a probability distribution). Free energy minimisation dynamics vary across timescales, ranging from neurocognition in real-time (i.e., perception and action; learning and attention); neurodevelopment throughout the lifespan; epigenetic mechanisms that minimise free energy across generations (e.g., kin); and the process of adaptation, which involves the optimisation of human generative models over time and conspecifics via the inheritance of adaptive priors [104].

These temporal processes are captured by Tinbergen's four levels of inquiry, which appeal to a dynamic causal hierarchy that is encapsulated by complementary paradigms in psychology: evolutionary psychology; evolutionary developmental approaches; developmental psychology; and the psychological subdisciplines [7], [70]. (For interpretation of the references to colour in this figure, the reader is referred to the web version of this article.)Fig. 4

Following EST, the hierarchical brain also exemplifies the complementary relationship between evolution and development—selection has canalized early sensorimotor regions that serve as neurodevelopmental anchors, allowing for the protracted, activity-dependent self-organisation of ‘domain-general’ association cortices throughout development that enhance evolvability by responding flexibly to environmental change [75], [76]. Consistent with this, simulation studies of evolving networks have shown that a hierarchical neural structure enhances evolvability by adapting faster to new environments than non-hierarchical structures [77]. The hierarchical segregation of neural networks into distributed neighbourhoods has also been found to stretch the parameter range for self-organized criticality by allowing subcritical and supercritical dynamics to co-exist simultaneously [78], which optimizes information processing and is therefore likely to be favoured by selection [79]. Maturational studies of neural networks throughout childhood and adolescence have further revealed that human cortical development mirrors phylogeny, progressing from the sensorimotor hierarchies found in all mammals through to recent association areas shared by humans and other primates [80]. The phylogeny of the brain is also reflected across nested levels of neurophysiological organisation, ranging from genes and transcription factors through to synaptic epigenesis and the long-range connectivity that is thought to underpin consciousness [81]. It is unsurprising, then, that a hierarchical neural structure has been found in every mammal examined to date [82].

The HMM rests on two assumptions. First, it subsumes the FEP as a biologically plausible EST of the evolution, development, form, and function of the brain. The FEP suggests that instead of just encoding a model of the world, the brain itself manifests a (hierarchical) phenotypic transcription of causal structure in its environment (i.e., an embodied statistical or generative model) that is optimized by evolution, development, and learning. Here, selection is simply nature's way of performing Bayesian model optimisation by minimising the (variational) free energy of human phenotypes across different (Tinbergian) timescales (Fig. 4, panel C). Emphasis is placed on adaptive priors, which are (epi)genetically-specified expectations that have been shaped by selection to guide action-perception cycles toward adaptive or unsurprising states [26].

The second assumption follows EST in psychology by recognising that the brain is an emergent sub-system produced by a hierarchy of causal mechanisms that act on the brain-body-environment system over time (i.e., adaptation, phylogeny, ontogeny, and mechanism; [7], [53], [83]). This causal hierarchy is recapitulated by major paradigms in psychology—which focus differentially on Tinbergen's [69] four questions (Fig. 4, panel C)—and calls for sophisticated, multiscale hypotheses that synthesise the FEP with different research programs in psychology and anthropology to explain both why a particular trait is adaptive, along with how it emerges from intergenerational, developmental, and real-time processes [7], [70]. The FEP and EST in psychology can therefore be regarded as different sides of the same coin—the FEP supplies a fully generalizable mechanistic theory of neural structure and function that applies to species generally, while an evolutionary systems approach to psychology builds upon the FEP by providing a substantive, evolutionary framework that is capable of explaining how the FEP manifests in Homo sapiens [84].

The HMM is not just a second-order theory derived from variational neuroethology; it can also be translated into a multidisciplinary research heuristic that promotes a tripartite approach to scientific inquiry [70]. First, it calls for integrative, multilevel hypotheses in psychology that address all four of Tinbergen's questions by synthesising research across evolutionary psychology, biological and cognitive anthropology; evolutionary developmental biology and psychology; developmental psychology; and psychological sub-disciplines such as cognitive, social, biological and personality psychology (Fig. 4, panel C). Second, it requires an analysis of how the phenomenon of interest can be explained in terms of the FEP. Finally, it calls for empirical research into the ways in which the phenomenon is neurophysiologically instantiated in a dynamic, hierarchical manner and manifests behaviourally via active inference [85]. Although this synergistic approach to developing hypotheses is undoubtedly complex and remains in its infancy, it has already furnished important new insights into depression [70], and yields distinct implications for neuroscientists and psychologists alike, while creating new avenues for collaborative research. This variational neuroethology of human systems has the major methodological advantage of conferring bona fide predictive power to the biological, cognitive, and social sciences. We can use simulation studies to predict which solutions will be adopted to respond adaptively to a given ecological problem, and then compare these computational models with empirical data.

For neuroscientists, the HMM lends itself to multimethod approaches that explore how detectable regularities in measurements of the brain can be explained by the psychological factors responsible for different patterns of neural activity in different contexts [86]. One way to isolate these factors is to use large databases of task-based fMRI activation studies to characterize the functional fingerprints of specific neural regions across different task demands [86], while another is to use such datasets or combine functional activation studies with neuropsychological lesion-deficit models to derive cognitive ontologies that systematically map relationships between specific cognitive functions and hierarchical neural dynamics [87], [88]. More broadly, the HMM also encourages the uptake of traditional methods in psychology—such as observational data collection and interviews—to explore the intersections between mind, brain, and behaviour [70]. Approaches in developmental psychology can be used to explore the dynamic ways in which human development differentiates (error-minimising) neurocognitive and biobehavioural patterns between individuals [70], while comparative, cross-cultural, computational and dynamical approaches in evolutionary psychology and cognitive anthropology can elucidate the (epi)genetic mechanisms that underlie our species-typical adaptive priors [29], [84], [89], [90]. Finally, dynamical methods such as computer simulations and computational models enable us to directly examine how such levels of causation interact [91], [92], [93], allowing neuroscientists to discover how the phenomena highlighted by psychologists reflect adaptive free-energy minimisation under different (evolutionary, developmental and ecobiopsychosocial) conditions.

The outcomes of such analyses might then be confirmed through experimental research, potentiating a fruitful dialectic between computational analyses and real-world observations.

On the other hand, the FEP offers a neurobiologically plausible EST of the mind and biobehaviour to psychologists. It has already been fruitfully applied to a wide range of psychological phenomena [68], [94], extending from emotion (e.g., [95]), anxiety (e.g., [96]) and depression (e.g., [70]), through to autism (e.g., [97]), illusions (e.g., [98]) and psychosis (e.g., [99]). The FEP also lends itself to methods that are highly familiar to psychologists, such as the P300; a psychophysiological measure that captures temporal fluctuations in surprise [100]. Finally, since it can be equally applied across all four of Tinbergen's research questions, the FEP casts new light on the reciprocal interplay of ultimate and proximate processes; stands to benefit both evolutionary and developmental psychologists; and proffers a common, transdisciplinary language to unite psychology's sub-disciplines [7], [70].

However, if human phenotypes can only be understood by analysing the ecology from which they emerge, how might the HMM incorporate sociocultural influences—arguably our most influential selection pressure to date [89], [90], [101]? A promising way to address this question is to incorporate recent work on cultural affordances. According to this perspective, cultural ensembles minimise free energy by enculturing their members so that they share common sets of precision-weighting priors [29]. Human beings—with our specific forms of neural organisation, phenotypes, evolved behavioural tendencies and sociocultural patterns—minimise more free energy across spatial and temporal scales than any other species. Arguably, this is because we have crossed the ‘evolutionary Rubicon’: our survival depends on our ability to access and leverage cultural information and immersively participate in culturally adapted practices [89]. This is just another coordinated set of nested spatial and temporal scales in the greater Markov blanket of Homo sapiens. Another scale to consider—which we share with other animals like beavers and bees—is the free energy minimisation accomplished by designer environments [68]. Although the ways in which the HMM might be extended to incorporate cultural affordances remains an open question, a promising avenue would be to explore approaches in scientometrics. Using science as the subject of its inquiry, this discipline incorporates a wealth of models and quantitative techniques (e.g., citation and text analyses) that can be used to analyse how general selection and self-organisation act upon theorizing and research to reduce uncertainty about the environment [102], [103].

This affords a promising means to explore how the sociocultural generative models instantiated by different disciplines optimise (scientific) model selection by minimising free-energy (i.e., scientific prediction errors) over time [7], [34].

Ultimately, the HMM and theory of cultural affordances both rely on the idea that the Markov blanket of Homo sapiens possesses hierarchically nested temporal and spatial dimensions. These models both appeal to a scientific meta-theory (i.e., variational neuroethology) that defines Homo sapiens as each individual (phenotype) throughout the course of evolution, along with every individual that has either existed in the past or occupies the present (i.e., our species)—who may, in turn, alter the course of our evolution and our characteristically adapted niches. This is just another way of saying that the Markov blanket constituted by a single phenotype is dynamically nested within the global Markov blanket of our species (extending across all of Tinbergen's temporal scales). Conveniently, this approach also allows researchers to navigate the recursive complexities of the spatial axis (from biological macromolecules all the way up). By placing our Markov blanket around Homo sapiens, we necessarily encapsulate all of the dynamic, lower-level processes responsible for producing every phenotype, while imposing a clear upper limit on the complex adaptive system under scrutiny. Although the human Markov blanket is nested within the broader dynamics of other global Markov blankets that extend out into the universe, these lie beyond the limits of the system that this ecobiopsychosocial framework endeavours to explain.

The purpose of this article was to answer Schrödinger's question – ‘what is life?’ – by presenting a hierarchical multiscale free energy formulation—called variational neuroethology—that offers the sciences of life, mind, behaviour and society with a principled, computationally tractable guide to discovery. Arguably, the FEP affords a unifying perspective that explains the dynamics of living systems across spatial and temporal scales. The ontology for biological systems discussed here entails a multidimensional phase space populated by events (i.e., living systems as adaptive, self-organising patterns) that depend on the temporal dynamics of free energy minimisation. These events can be described as spatially and temporally nested Markov blankets, the dynamics of which are summarised by the Lagrangian of the FEP. Synthesising this framework with Tinbergen's four questions allows us to frame our scientific investigations with clearly defined temporal and spatial bounds, depending on the specific questions we wish to answer. We have described an empirically supported exemplar of this approach in the cognitive and behavioural sciences—the HMM—and shown how this theory yields unique and promising avenues for interdisciplinary research. The challenge, of course, lies in translating theory into productive scientific practice, and in testing the current limits of the FEP by applying it to species without a brain, like E. coli, fungi and flora.