Sentience and the Origins of Consciousness: From Cartesian Duality to Markovian Monism

Dual-aspect monism is the position that reality has two aspects: a mental and a physical aspect. Dual-aspect monism is very similar to neutral monism. Depending on how it is defined, it may even collapse into neutral monism (or into panpsychism, see [104], p. 366). For instance, if dual-aspect monism is defined as the view that reality is, at a fundamental level, both physical and mental, then this comes extremely close to the view that basic entities are intrinsically both mental and physical—and hence to a version of neutral monism (see [101], sec. 8.3).

Furthermore, if the aspect in ‘dual-aspect’ is interpreted in terms of properties, such that basic entities have both mental and physical properties (as suggested by [105], p. 46), then dual-aspect theory becomes a form of property dualism—which we ruled out above.

There are versions of dual-aspect theory that explicitly refrain from defining the dual aspect in terms of property dualism (see, e.g., [106], pp. 339,342). Markovian monism is similar to dual-aspect monism (cf. [107], pp. 220–221), in that it entails that one and the same thing (i.e., internal states of a system possessing a Markov blanket) can be viewed from two perspectives. Internal states can either be regarded as states of a random dynamical system; or they can be viewed as the parameters of a probability distribution (i.e., probabilistic belief). In order to count as a dual-aspect monism, these two perspectives would have to be mutually irreducible (see [106], p. 46; [105], p. 341), we are sceptical that this would be a coherent interpretation of the dual information geometry.

As with the other two interpretations discussed above, an interpretation in terms of a dual-aspect monism would presuppose a realist view on descriptions in terms of the extrinsic information geometry. Furthermore, just as the interpretation in terms of neutral monism, it would entail that single-cell organisms have a mind. In what follows, we will sketch how Markovian monism can ground versions of reductive materialism. This physicalist interpretation of Markovian monism is the one we favour—although we admit that other interpretations cannot conclusively be ruled out.

Here is what we believe is the most coherent interpretation of the formal treatment. The fact that one can associate two information geometries with systems possessing a Markov blanket reveals a continuity between simple, non-conscious systems and more complex, conscious systems such as human beings: due to the extrinsic information geometry, simple systems can be described as if they had beliefs about external states. For conscious systems, the perspective—afforded by such ‘as if’ descriptions—acquires a special status, because it will typically abstract away from many of the details inherent in the mechanistic perspective. For instance, the probabilistic beliefs ascribed to explain cognitive phenomena are typically assumed to be represented by the average activities of neuronal populations, which means that any differences between populations with the same average properties will be irrelevant from the perspective of the extrinsic information geometry. This squares well with the idea that causation (including mental causation) is a macroscopic phenomenon [108,109]. At the same time, these macrostates are always grounded in more fine-grained physical states, and their properties can be reductively explained in terms of physical properties.

Furthermore, the computational properties ascribed to conscious systems will be more numerous and more complex than those ascribed to non-conscious systems. There are no additional, non-reducible properties, which are necessary to explain the mind and consciousness; between some non-conscious and conscious systems, there is only a gradual difference. This entails that consciousness is a vague concept; i.e., there will be borderline cases in which the concept cannot unequivocally be applied.

In particular, this proposal rests upon a distinction between temporally deep and shallow generative models, that accompanies the distinction between conscious and unconscious inference. This distinction is vague, in the sense that any generative model of dynamics has, to a certain extent, temporal depth. For example, predictive coding, homoeostasis and thermostats can all be articulated in terms of perceptual control [46,110,111] and a reflexive form of active inference using generative models based upon differential equations. The fact that a generative model entails differential equations means that there is some inference over time. The distinction between deep and shallow then becomes a quantitative issue. Perhaps a better distinction would be between generative models that entertain a single trajectory into the future, versus multiple (counterfactual action dependent) trajectories that incur a selection problem; namely, choosing an action or planning.

Construing consciousness as a vague concept may even have relevance for the meta-problem of consciousness [112]; i.e., the problem of explaining why it seems (to many) that a physical duplicate of a conscious creature could be non-conscious. Although solving the meta-problem is not the aim of this paper,19 we can at least contribute to an explanation: as noted above, our interpretation of Markovian monism entails that there is only a gradual difference between some non-conscious and conscious systems, and that consciousness is a vague concept. So, when people claim they can imagine a physical duplicate that is unconscious, they may in fact imagine not a complete duplicate, but a system that differs in (seemingly non-significant ways) from a conscious system. As an analogy, consider a heap of sand. A heap of sand is constituted by grains of sand. But, one could object, a heap of sand cannot be just a collection of grains of sand, because I can imagine a collection of grains of sand (say, three grains) that does not count as a heap. Hence, there seems to be a crucial difference between collections of grains of sand and heaps of sand—just adding a grain of sand to something that is not a heap does not turn it into a heap. Similarly, just adding a bit more structure and function to a non-conscious system does not turn it into a conscious system. Hence, it would seem as if consciousness requires more than just the right structure and functions, and the hard problem arises. But if consciousness is a vague concept (as suggested by our interpretation of Markovian monism), then the right structure and functions can be metaphysically sufficient for consciousness, even if adding just a bit of structure and function to any uncontroversially non-conscious system does not make it conscious.

Furthermore, the very existence of the meta-problem implies a certain kind of Bayesian belief that entails some puzzlement about ‘our capacity to have subjective experiences of a quantitative sort’. But ‘qualia’ and accompanying ‘puzzlement’ are just Bayesian beliefs that imply an extrinsic information geometry. So, is there anything special about Bayesian beliefs about Bayesian beliefs? The answer is yes: beliefs about beliefs (in a mathematical sense) require a hierarchical generative model. But, a hierarchical generative model requires hierarchically deployed Markov blankets to introduce the necessary conditional independencies (which make it hierarchical). We therefore conclude that phenomenally conscious systems for which a hard problem exists must possess a certain kind of statistical structure; namely, Markov blankets within Markov blankets [4].

Although we believe that there are only gradual differences between non-conscious and conscious systems, if one merely considers the probabilistic beliefs that can be ascribed to such systems, there are still categorical differences that can be described in terms of more high-level properties, such as intentionality and computation (note that this does not imply a “phase transition” between unconscious and conscious systems).

In particular, one can make a threefold distinction between (i) systems that behave only ‘as if’ they implemented computations over probabilistic beliefs, (ii) systems for which the “computational stance” [114] provides added explanatory value, and (iii) systems that can usefully be described as not only computational, but also as representational systems.20 While not speaking against a continuity between life and mind [103], this threefold distinction could be used to establish a discontinuity between life and consciousness.

Specifying the difference between (i) and (ii) would require defending a particular account of computation, which is beyond the scope of this paper. The step from (ii) to (iii)—i.e., from a computational to a representational system—requires ascribing content to internal states of the system. Representationalist interpretations of the free-energy principle refer to computations that are implemented (or approximated) by systems that minimize free energy (see, e.g., [117], pp. 571–572). Such computations are defined with respect to exactly the types of probabilistic beliefs encoded by systems with an extrinsic information geometry.

Although Markovian monism, interpreted as a form of reductive materialism, is not a theory of consciousness, it refers to properties that may ground mental properties (including phenomenal properties). As such, it provides a foundation for various physicalist approaches to consciousness and the mind, most notably representationalism and (computational) functionalism.21

There have been previous attempts to use information theory to describe conscious processing. Perhaps the most notable is integrated information theory [118,119]. One might ask about the relationship between the free energy principle (FEP) and integrated information theory (IIT)? At the time of writing, there is a gap between these theoretical approaches. First, the FEP is a ‘first principle’ account that uses variational principles to build upon the Langevin formulation of random dynamical systems. In contrast, IIT is an ‘axiomatic’ approach that starts with some assumptions about what information processing must look like to be a contender for explaining conscious experience. The formal distinction between the FEP and IIT is that the free energy principle is articulated in terms of probabilistic beliefs about some (external) thing, while integrated information theory deals with probability distributions over the states of some system. In other words, IIT does not commit to an extrinsic information geometry (the “geometry of integrated information” is an intrinsic information geometry, see [120]). This is not necessarily a problem, in so far as IIT offers a normative (i.e., measurable, in principle) description of systems that comply with axioms, which inherit pre-theoretical notions of consciousness. On the other hand, both the free energy FEP and IIT can be cast in terms of information theory and in particular functionals (e.g., variational free energy and ‘phi’). Furthermore, they both rest upon partitions (e.g., Markov blankets that separate internal from external states and complexes that constitute conscious entities and can be distinguished from other entities). This speaks to the possibility of, at least, numerical analyses that show that minimising variational free energy maximises ‘phi’ and vice versa.

Although integrated information theory does not commit to a Markovian information geometry of experience (i.e., conscious or unconscious inference about something), it is possible to establish some kind of construct validity between the FEP and IIT in terms of the axioms upon which IIT is predicated. In other words, one can establish—at least heuristically—that the FEP features the essential properties of experience that constitute the axiomatic basis of IIT. There are five axioms; namely, intrinsic existence, composition, information, integration and exclusion. In brief:

Intrinsic existence—consciousness exists: each experience is actual and exists from its own intrinsic perspective. This is a necessary consequence of Bayesian mechanics under the free energy principle because the dynamics underlying inference are physically realised and are, by construction, intrinsic in the sense of pertaining to internal states.

Composition—consciousness is structured: with multiple phenomenal distinctions. Again, this is a necessary aspect of Bayesian mechanics, which is defined in terms of the structure implicit in conditional independencies. Indeed, from a statistical perspective, minimising variational free energy is synonymous with structure learning [59,121,122].

Information—consciousness is unique: each experience is the particular way it is, thereby differing from other possible experiences (i.e., differentiation). Again, this is a fundament of Bayesian mechanics under the free energy principle; in the sense that any information geometry implies a particular point on a statistical manifold (of internal or intrinsic states) maps to a particular probability or belief state with phenomenal support (i.e., an extrinsic belief distribution over the external states).

Integration—consciousness is unified: each experience is irreducible to disjoint subsets of phenomenal distinctions (i.e., integration). Again, this is a necessary aspect of the information geometry that underwrites the free energy principle. This follows because for each point on the internal statistical manifold, there is a single probabilistic belief (i.e., variational density). In other words, although this density could be very high dimensional, it is just one probabilistic belief that cannot be dissembled or reduced. Another aspect of the axiom of integration is that “every part of the system has both causes and effects within the rest of the system” ([123], p. 3). This is true for systems possessing a Markov blanket, because the gradient flows of internal states (and associated belief updating) are, by definition, conditionally dependent.

Exclusion—consciousness is definite: each experience is characterised by what it is (neither less no more than) and flows at the speed it flows (neither faster nor slower). Again, this is a necessary consequence of the density dynamics that underwrites the free energy principle. In other words, flows on the extrinsic (statistical) manifold are unique and entail particular probabilistic beliefs about external states, i.e., precise beliefs about being in a particular (external) state but not another. Furthermore, each probabilistic belief has its own sufficient statistics that exclude the possibility of other sufficient statistics. For example, beliefs about my temperature can be stipulated with an expectation that my temperature is such and such. This precludes the possibility that I expect to my temperature to be anything else. In contrast to the exclusion axiom, however, the existence of a Markov blanket at one spatiotemporal scale does not exclude the existence of (e.g., nested) Markov blankets at other spatiotemporal scales.

In summary, on an informal review, the information geometry and density dynamics implied by Markov blankets appear to possess the qualities—or conform to the essential criteria—that constitute the axiomatic basis of integrated information theory.

The important result of this section, from our perspective, is that at least some properties associated with consciousness are already entailed by Bayesian mechanics under the free energy principle. This supports the (speculative) hypothesis that adding further constraints on generative models—entailed by systems possessing a Markov blanket—might enable us to say which systems are conscious, and which are not. Unconscious systems do not perform active inference in a way that entails that characteristic features of consciousness are instantiated, whereas conscious systems do. Specifying the constraints on generative models that underpin active inference of the sort that entails characteristic features of consciousness can lead to a unitary concept of consciousness ([124], as opposed to a bundle of feature descriptions; see [125]). In other words, a sufficiently specified sort of active inference may describe computational processes that account for clusters of features that are characteristic for consciousness—and thereby show why these features cluster together (cf. the natural kind approach sketched in [126], p. 7).

To recap, the information geometry above—and attending free energy principle—rest upon a separation of external from internal states by blanket states. This move is crucial for elaborating a physics of sentience, in which physical dynamics entail probabilistic beliefs about something. In this sense, it takes us beyond existing formalisms in the physical and philosophical sciences—revealing some key issues. For example, quantum treatments generally rely upon some specification of a Schrödinger potential. But where did this potential come from? Similarly, for statistical thermodynamics, where did the ‘heat bath’ (i.e., thermal reservoir) come from and what contains the heat bath? In short, there would be no quantum or statistical mechanics in the absence of Markov blankets (i.e., Schrödinger potentials and heat baths). The same questions can be posed to things like integrated information theory: what is this information about, in the absence of a Markovian (belief-based) information geometry? What principles explain the emergence and maintenance of partitions induced by complexes?

The point here is that a Markovian monism (or information geometry) necessarily requires some notion of duality or conjugacy, here afforded by the Markov blanket. On this reading of self-evidencing to nonequilibrium steady state, some pressing questions arise. For example, what would happen if internal and external states were statistically sequestered. In other words, is there a sentient physics for isolated systems, such as those considered in statistical mechanics. From a neurobiological perspective, this speaks to altered states of consciousness that ensue with physiological or pharmacological quenching of blanket states. There are many examples that we could pursue here; including states of consciousness associated with psychedelic and psychomimetic drugs, or, indeed, the false inference associated with psychopathology (e.g., hallucinations and delusions). However, we will focus on a canonical example; namely, sleep and dreaming. So, what does sleep physiology tell us about conscious or unconscious inference?

If, for simplicity, we assume that the state of sleep corresponds to a sequestering of internal states from blanket states, we have an interesting preparation of a neuronal system that is temporarily—and repeatedly—isolated from the sensorium. This simplification is easily substantiated by many neurophysiological and neurochemical aspects of sleep physiology [127]. For us, the key question is: what happens to the Markovian information geometry and Bayesian mechanics of the internal (neuronal) states? At first glance, the notion of self-evidencing as an explanation for internal dynamics simply goes away. This is because the Lyapunov or potential function driving dynamics ceases to exist in the absence of blanket states (technically, the gradients that underwrite gradient flows disappear). However, at nonequilibrium steady-state, periods of disconnection from blanket states must themselves be transient and repetitive; i.e., be part of the itinerant dynamics that have a pullback attractor. This means Bayesian mechanics must still apply, even during the suspension of any coupling with blanket states. We will consider a physiological case (of sleep) in which in autonomous dynamics are still in play. In this setting, the variational free energy gradients are driven by the part of free energy that does not depend upon blanket states. This part is the complexity term of Equation (5), where removing blanket states discloses a description of complexity (i.e., redundancy) resolving internal dynamics:

In other words, neuronal dynamics during sleep will appear to minimise the complexity of the generative model (i.e., minimise the divergence between the posterior beliefs and prior beliefs—in the absence of sensory evidence). This is precisely the argument put forward in statistics for optimising models in the absence of new statistical data—by removing redundant model parameters [128]. In neurophysiology, this is the argument that we have made previously to explain the very existence of sleep phenomenology—and in particular, the role of dreaming [129,130,131]. In short, physiological states of altered consciousness, such as sleep, may offer an important empirical handle on theoretical notions—notions that arise from the variational principles of sentience.

In summary, an extrinsic information geometry can exist in the (temporary) absence of blanket states, in virtue of prior beliefs held by internal states. These prior beliefs underwrite proto-consciousness [127] and are necessary to generate virtual or fictive realities [132] in states such as dreaming [129,130,131]. There are many fascinating issues here; for example, the complexity term of the free energy functional above provides a compelling metaphor for the housekeeping that we may enjoy during sleep [71,133]. This complexity minimisation itself has formal links with both machine learning [128] and universal computation [88,90,134]—and physiology in the form of synaptic homoeostasis [133,135].

In conclusion, we have rehearsed some of the cornerstones of statistical physics and information theory to show how the very existence of things (i.e., Markov blankets) necessarily induces an information geometry with two aspects. First, the dynamics of physical (internal) states of any sentient particle or creature is equipped with an information geometry, in terms of time dependent changes in probability distributions over internal states. We have called this an intrinsic information geometry. At the same time, there is a conjugate information geometry, which pertains to probability densities over external states parameterised by internal states. We have called this an extrinsic information geometry (because it is predicated upon probabilistic beliefs about external states). Crucially, the two are formally and fundamentally linked—in that the dynamics of internal states can always be expressed as a gradient flow on a variational free energy functional of belief (protophenomenal) states. This construction is entirely consistent with forces cast in terms of stochastic thermodynamics, with the appropriate constant of proportionality (i.e., Boltzmann’s constant and the temperature).

Second, we have considered the time it takes for a particle or creature to return to its attracting manifold (i.e., nonequilibrium steady state) from an initial state. When treated in the form of a path integral or fluctuation theorem, this temporal aspect may distinguish among different kinds of creatures; depending on how deeply their generative model (entailed by internal states) considers the future; c.f., counterfactual depth [111,136]. This is functionally equivalent to the temporal depth or extent of policies; namely, courses of action, and internally consistent with the notion of planning as inference [85,87,137].

Another technical formulation of information-processing—that is closely related to information geometry—is the use of gauge theories (e.g., the celebrated theory of general relativity). Our own work in this area [138] focused on gauge theories associated with information geometry and the Fisher information metric. Recall that the Fisher information metric that equips the belief space or statistical manifold (here, afforded by internal states) with a geometry has a number of revealing interpretations. First, the Fisher information metric is simply the curvature of the variational free energy as one moves on the internal (statistical) manifold. This is the same as the conditional precision or confidence placed in beliefs about external states. From a psychological perspective, this curvature or precision plays a key role in predictive processing (i.e., Bayesian brain) accounts of attentional selection and, a particularly important role in interoceptive inference [67,68,76,139,140]. We emphasise this seamless connection from gauge theories—through information geometry and variational inference—to precision for a special reason. The central role of precision and confidence in mediating consciousness is exactly the endpoint of the phenomenological and neuropsychological analysis of conscious processing and selfhood offered by Mark Solms [107]. Furthermore, the ‘paper trail’ from gauge theory to attention endorses pre-theoretical notions about their intimate relationship [141,142]. One could develop this story even further, in terms of the predictive processing of precision per se—and how this may underwrite mental action and a sense of agency [111,139].