The hierarchically mechanistic mind: A free-energy formulation of the human psyche

Abstract

This article presents a unifying theory of the embodied, situated human brain called the Hierarchically Mechanistic Mind (HMM). The HMM describes the brain as a complex adaptive system that actively minimises the decay of our sensory and physical states by producing self-fulfilling action-perception cycles via dynamical interactions between hierarchically organised neurocognitive mechanisms. This theory synthesises the free-energy principle (FEP) in neuroscience with an evolutionary systems theory of psychology that explains our brains, minds, and behaviour by appealing to Tinbergen's four questions: adaptation, phylogeny, ontogeny, and mechanism. After leveraging the FEP to formally define the HMM across different spatiotemporal scales, we conclude by exploring its implications for theorising and research in the sciences of the mind and behaviour.

Keywords: Active inference; Evolutionary systems theory; Free-energy principle; Hierarchically mechanistic mind; Neuroscience; Psychology.

Fig. 1

The hierarchical organisation of neural networks. Global brain function (i.e., cognition) can be described as the global integration of local (i.e., segregated) neuronal operations that underpin hierarchical message passing among cortical regions. Global integration is greatly facilitated by the hierarchical organisation of neural networks into (relatively modular) neurocognitive mechanisms. In network neuroscience, a neural network is modelled in terms of nodes and their connections, which are called edges. A node is defined as an integrated unit within a network. In a fractal or modular hierarchy, each node also comprises a smaller network of nodes that interact among themselves at a lower nested level. In the brain, this fractal, encapsulated hierarchy extends from neurons and macrocolumns, through to macroscopic brain regions and distributed neural networks. According to predictive coding theory, superficial pyramidal cells compare expectations at each level with top-down predictions from deep pyramidal cells at higher levels, while neuromodulatory gating or gain control of superficial pyramidal cells determines their influence on the implicit belief updating in higher hierarchical levels. Reproduced from .

Fig. 2

The free-energy principle. (A) The quantities that define variational free-energy. These quantities reflect a partition of the system into its internal states, μ, (e.g., states of the brain) and the quantities that describe its exchanges with the environment; namely, sensory input, s = g(η,a)+ω, and action, a, which alters the ways in which the organism samples its environment. The environment itself is specified by equations of motion, $\dot{\eta }=f(\eta ,a)+\omega$, which describe the dynamics of (hidden) states of the world, η. The term ω denotes random fluctuations. Both internal and active states change synergistically to minimise variational free-energy. This free-energy is a function of sensory input and a probabilistic representation of hidden environmental causes (i.e., variational density), q(η:μ), which is encoded by the system's internal states. (B): Alternative expressions for variational free-energy, which show what its minimisation entails. With respect to action, free-energy can only be suppressed by increasing the accuracy of sensory data (i.e., selectively sampling data that are predicted). Conversely, the optimisation of internal states makes the representation (i.e., variational density) an approximate conditional density over the causes of sensory input (i.e., perception, which minimises the divergence between the variational and true posterior density). This optimisation allows the variational free-energy to impose a tighter bound on surprise and enables the system to act upon the world to avoid surprising sensory and physiological states. Reproduced from .

Fig. 3

The evolutionary systems theory of psychology. Human phenotypes, cognition and behaviour emerge from circular interactions between (general and natural) selection and self-organisation operating within and across Tinbergen's four domains of biological dynamics (i.e., adaptation, phylogeny, ontogeny, and mechanism). The various fields of psychological inquiry explain this process by formulating models of human phenomena according to four intersecting levels of analysis: evolutionary hypotheses to explain species-typical, adaptive traits (i.e., evolutionary psychology); explanations for intergenerational, between-group differences (i.e., evolutionary developmental biology and psychology); ontogenetic explanations for individual similarities and differences (i.e., developmental psychology); and mechanistic explanations for real-time biobehavioural phenomena (i.e., the sub-disciplines). These levels of analysis are commensurate and complementary: evolutionary theories tackle the ultimate questions of psychology by explaining the adaptive properties of human cognition and biobehaviour; dynamic systems approaches address its proximate questions by shedding light on the intergenerational, developmental, and real-time mechanisms responsible for producing such phenomena. This perspective encapsulates and synthesises the various paradigms and sub-disciplines of psychology: the recursive informational exchange between different fields of inquiry allows researchers in each subfield to constrain their research in light of advances in others, and to integrate findings across different levels of psychological analysis to develop unique, substantive hypotheses. Importantly, the non-substantive meta-theory of EST, which formalises the interaction between (both general and natural) selection and self-organisation, permeates all four explanatory levels and imposes distinct inclusion criteria upon any derivative of the meta-theory itself: any multi-level hypothesis derived from this EST must conform to these two fundamental principles. Adapted from .

Fig. 4

The hierarchically mechanistic mind. $F(\tilde{s}(a),{\mu }^{(i)}|m^{(i)})$ denotes the variational free-energy of sensory data (and its temporal derivatives), $\tilde{s}(a)$, as well as the states, μ, of an agent, $m^{\left(i\right)}\in s$, that belongs to a subgroup, s ∈ c, of a given class, c. Action, a, regulates the sampling of sensory data; while the internal states of the organism, μ, encode expectations and predictions (i.e., Bayesian beliefs) about the mean of a probability distribution. Under this formalism, neurocognition entails two dynamically coupled processes. The first optimises neuronal and effector dynamics (i.e., perception and action) to attune the organism to its environment – by minimising prediction errors (resp. free-energy) based on a generative model of the hidden causes of sensory data. The second process optimises synaptic strength and efficacy – over seconds to hours – to encode causal structure in the sensorium and the precision of prediction errors (i.e., learning and attention). Neurodevelopment optimises human generative models through activity-dependent pruning and the maintenance of neural structures and connections, which are transmitted epigenetically. Neural microevolution optimises average free-energy over generations of individuals belonging to a subgroup (e.g., kin) of a given class (i.e., conspecifics) via the (exo- and epi-)genetic transmission of generative models. Neural evolution optimises average free-energy over time and individuals of a given class (i.e., conspecifics) through the effects of selective pressure on their generative models or priors. Reproduced from .

Fig. 5

The evolutionary systems model of depression. Under active inference, motor and autonomic reflexes mediate action and are driven by descending (proprioceptive and interoceptive) prediction errors (e.g., reflexes that resolve sensory prediction errors). Action entails the attenuation of ascending prediction errors (i.e., the down-regulation of precision). Prediction errors cannot always be resolved through action; in which case, the attenuation of sensory precision is suspended. This suspension enables ascending prediction errors to revise posterior beliefs, which improves the accuracy of top-down predictions. Here we apply active inference to depressed mood states. Under this model, when depression is adaptive, it engenders an increase in the precision of (bottom-up) social (interoceptive and affiliative) prediction errors when an individual is faced with the threat of aversive interpersonal outcomes (e.g., exclusion). This increased precision improves perceptual inference and learning about the probable causes of social stimuli: it heightens sensitivity and directs attention to socio-environmental cues, while reducing confidence in (top-down) social predictions. Cognitively, this is reflected by the inhibition or suspension of goal directed behaviour (e.g., anhedonia), along with an attentional bias toward social cues and increased rumination about self-other relations. However, depression becomes pathological when there is a pervasive failure of sensory attenuation, which induces aberrant beliefs about the likelihood of social rewards and engenders negative expectations about interactions with others (e.g., pessimism, low self-esteem). These expectations of negative social outcomes can become self-fulfilling, because they can lead the individual to search for sensory evidence that social rewards are improbable and suppress exploratory or acquisitive interpersonal behaviours (i.e., those with uncertain outcomes). Behaviourally, both adaptive and pathological depressed states reduce uncertainty within the social world by down-regulating reward-approach behaviours (e.g., anhedonia, social withdrawal), and by generating signalling behaviours that elicit interpersonal support (e.g., reassurance seeking) and defuse potential conflict (e.g., submissive behaviours). Reproduced from .