Minimal physicalism as a scale-free substrate for cognition and consciousness

Theories of consciousness and cognition that assume a neural substrate automatically regard phylogenetically basal, nonneural systems as nonconscious and noncognitive. Here, we advance a scale-free characterization of consciousness and cognition that regards basal systems, including synthetic constructs, as not only informative about the structure and function of experience in more complex systems but also as offering distinct advantages for experimental manipulation. Our “minimal physicalist” approach makes no assumptions beyond those of quantum information theory, and hence is applicable from the molecular scale upwards. We show that standard concepts including integrated information, state broadcasting via small-world networks, and hierarchical Bayesian inference emerge naturally in this setting, and that common phenomena including stigmergic memory, perceptual coarse-graining, and attention switching follow directly from the thermodynamic requirements of classical computation. We show that the self-representation that lies at the heart of human autonoetic awareness can be traced as far back as, and serves the same basic functions as, the stress response in bacteria and other basal systems.

Keywords: active inference, aneural systems, basal cognition, classical computation, integrated information, memory, quantum computation, self-representation, state broadcasting

The starting point for sciences of consciousness and cognition has traditionally been human consciousness and cognition. The Cartesian presumption against nonhuman consciousness and cognition remained strong enough, even just two decades ago, that prominent researchers found it necessary to argue in print that nonhuman animals and even human infants experienced perception and interoception and engaged in intentional actions (e.g. Panksepp 2005; Trevarthen 2010; Rochat 2012). Those days are thankfully behind us. While researchers in comparative cognition may debate the levels of complexity, compositionality, or stimulus-independence an information-processing system must have to be regarded as “cognitive” (see Bayne et al. 2019 for a recent snapshot), big-brained creatures that learn readily and display flexible behavior in complex environments are now regularly regarded as cognitive systems that are aware of both their environments and their own bodily states, even if they are birds (e.g. Güntürkün and Bugnyar 2016; Nieder et al. 2020) or cephalopods (e.g. Mather 2019; Schnell et al. 2020). To insist that such creatures altogether lack phenomenal consciousness, that they are incapable of experiencing pain and likewise experience nothing whatsoever while hunting, improvising tools, or engaging in mating displays as a doctrinaire Cartesian presumably would, seems by current standards chauvinistic or even perverse.

Characterizing birds and cephalopods, and for that matter other mammals and even other people, as phenomenally conscious begs, of course, the philosophical questions of what phenomenal consciousness is ontologically, and of how it relates to observable structures, functions or behaviors. These questions constitute the “hard problem” (HP) of consciousness (Chalmers 1996; cf. Dietrich and Hardcastle 2005; Dietrich et al. 2021). We agree with Klein and Barron (2020) that the HP is not a problem to be solved, but rather a set of intuitions to be overcome. The questions of interest regarding birds, cephalopods, other mammals, and other people are not, as far as we are concerned, whether or why, in some philosophical sense, they are capable of experience, but rather: (1) what any particular organism or other system of interest is aware of, and (2) how its awareness drives its cognition. Our interest here is not, moreover, in what any particular experience is like for any given system, but in whether it is like something or other; to employ a common example, it is not the particular characters of red or green qualia that are of interest (see e.g. Hacker 2002 for an argument that no such particular characters even exist), but whether some difference between what we, as 3rd party observers, call “red” and “green” is experienced by the system of interest, and whether this difference has some effect on what the system does. Hence we regard consciousness as a (potential) property that a system may have, independently of any details about what it experiences: it is enough, in our view, that it experiences something, that some experience occurs (see Lamme 2018 for a discussion of some of the subtleties that this “binary” view elides).

We will use the terms “awareness” and “consciousness” to mean the capacity for or capability of having phenomenal experiences, however basic or minimally structured, and will focus exclusively in what follows on the “what” and “how” questions above.

A well-established literature extends the concepts of consciousness—the capability of having phenomenal experiences, however basic or minimally structured—and cognition to phylogenetically basal systems, including free-living or facultatively communal unicells, whether pro- or eukaryotic (Maturana and Varela 1980; Pattee 1982; Stewart 1996; di Primio et al. 2000; Lyon 2015; Baluška and Levin 2016; Baluška and Reber 2019, 2021; Levin 2019, 2020; Fields and Levin 2020a; Lyon 2020; Baluška et al. 2021), plants (Calvo and Kiejzer 2009; Gagliano and Grimonprez 2015; Baluška et al. 2018; Debono and Souza 2019; Calvo et al. 2020), and aneural or lower (than mammals, birds, or cephalopods) complexity neural metazoa, particularly flatworms (Shomrat and Levin 2013; Inoue et al. 2015; Levin 2020) and insects (Menzel 2012; Perry et al. 2017; Pfeffer and Wolf 2020). Like the extension of these concepts from humans to nonhuman mammals and then to big-brained nonmammals, this extension to more basal organisms was initially motivated by observations of communication, learning, and behavioral flexibility, and by functional similarities between the mechanisms supporting information processing and learning in basal systems and in more complex systems such as mammals. Both molecular and bioelectric mechanisms of cellular information processing, memory, communication, and error correction are, in particular, evolutionarily ancient and conserved across phylogeny (Fields and Levin 2018, 2020b; Fields et al. 2020; Levin 2020). Neither the specialization of some epithelial cells for communication, i.e. as neurons (Nickel 2010; Arendt et al. 2015; Burkhardt and Sprecher 2017) nor the gradual elaboration of brains (Holland et al. 2013; Liebeskind et al. 2017) introduce fundamentally novel molecular or bioelectric mechanisms. Hence while claims of specific kinds of experience, e.g.

pain in plants (Draguhn et al. 2020), are controversial, there is no evident biological discontinuity below which organisms can clearly be viewed as completely unaware of their bodily states or their environments, i.e. as utterly lacking phenomenal consciousness. Nor is it obvious that supra-organismal systems utterly lack phenomenal consciousness (Friedman and Søvik 2021). Indeed, the criteria for “having experiences” may be as vague, general, and extensible to non-Terrestrial or even artificial systems as plausible criteria for “life” are (e.g. Bartlett and Wong 2020; but see also Mariscal and Doolittle 2020 for an argument that such “definitionism” is scientifically pointless).

With the development of more sophisticated, biologically motivated theoretical models of consciousness and cognition, the functional similarities between information processing in basal and in more elaborated systems have become even more evident. It is now well-known that signal transduction and gene regulatory networks in single cells share the scale-free, small-world topology (Agrawal 2002; Barabási and Oltvai 2004; Schlitt and Brazma 2007) found ubiquitously in nervous systems (Sporns and Honey 2006; Hagmann et al. 2008; Shanahan 2012; Sporns 2013) and even in social and technological networks (Watts 1999; Barabási 2016; but see Broido and Clauset 2019 for evidence that some biological and social networks are small-world but not strictly scale-free). Networks with high fan-in/high fan-out “bow-tie” topology (Wang et al. 2016; Niss et al. 2020), in particular, implement state-broadcasting functions analogous, at the cellular level, to the broadcasting functions of long-distance connections between network hubs in global neuronal workspace (GNW) models (Dehaene and Naccache 2001; Baars and Franklin 2003; Wallace 2005; Baars et al. 2013; Dehaene et al. 2014; Mashour et al. 2020). Signal transduction and gene regulatory networks are, even when they contain bow-tie nodes, characterized by multiple, typically cross-modulating feedback loops that enable both behavioral plasticity and learning (e.g. Brun-Usan et al. 2020); hence they have positive integrated information Φ and render their minimal containing systems conscious by the criteria of Integrated Information Theory (IIT; Tononi 2008; Oizumi et al. 2014; Tononi and Koch 2015).

As the locus of molecular, thermodynamic, and bioelectric exchange with the environment, the cell membrane implements a Markov Blanket (MB) that renders its interior conditionally independent of its exterior (Pearl 1988; Clark 2017); this allows the cell to be described as a Bayesian active inference system (Friston 2010, 2013; see also Cooke 2020 for a variation on this approach). The utility of this Bayesian approach has been demonstrated in simulation models of cell–cell communication driving morphoghenesis (Friston et al. 2015; Kuchling et al. 2020). These cross-scale similarities motivate a hypothesis that consciousness and cognition are scale-free phenomena that characterize all living systems (Levin 2019, 2020; Fields and Levin 2020a).

If consciousness and cognition are scale-free phenomena, we can expect them to be supported by common, scalable mechanisms that can be investigated in whatever systems permit the most straightforward theoretical and experimental approaches. Phylogenetically basal organisms, in vitro preparations, and synthetic constructs (e.g. Kriegman et al. 2020) provide obvious advantages of manipulability and environmental control. Studies of basal systems are, moreover, especially effective in overcoming the intuitions that give rise to the HP, as they allow the mechanisms via which single cells and relatively simple multicellular organisms navigate their environments—mechanisms that they share with most of our cells, and with us as organisms—to be investigated in detail. A theoretical framework suitable for such systems should be scalable, provide a full suite of formal capabilities, and make as few assumptions as possible. It should, in particular, make no scale-dependent architectural assumptions.

Here, we provide a straightforward construction of fundamental, scale-free features of consciousness and cognition within a generic description of system-environment information exchange as bipartite physical interaction (Fields and Glazebrook 2020a; Fields and Marcianò 2020a,b). We term this description “minimal physicalism” (MP) as it makes no assumptions about classical computational architecture, in particular, no assumptions about network architecture (Broido and Clauset 2019), and no physical assumptions beyond those of quantum information theory (Nielsen and Chuang 2000 provide a standard introduction). At the level of molecular interactions at the Angstrom, femtosecond (Å, fs) scale of molecular dynamics calculations (Zwier and Chong 2010; Vlachakis et al. 2014), biological systems are quantum systems, and biological information processing is quantum computation: cellular energy budgets of both prokaryotes and eukaryotes fall orders of magnitude short of the power required to maintain classical states of just protein conformation and localization at this scale (Fields and Levin 2021), despite the massive consumption of ATP by big-brained eukaryotes such as humans (Ueno et al. 2005; Okuno et al. 2011). Hence cellular information processing cannot be entirely, or even primarily classical, the experimental difficulties (Cao et al. 2020) of unambiguously observing quantum coherence in biological systems notwithstanding.

As we show below, the thermodynamic requirements of classical computation by quantum systems have profound consequences for cellular, and hence for organismal, awareness and information processing. In general, basing MP on the physical assumptions of quantum information theory has significant empirical consequences that extend far beyond a mere rejection of dualism. These include:

MP is completely scale-free, applying in the same form to interactions between molecules, cells, tissues, organisms, social groups, or ecosystems and their respective environments. Hence it predicts common mechanisms that can be probed empirically at any scale. It makes no assumptions about the structure or dynamics of the environment, at any scale, beyond its being a physical system.

MP requires every property of either itself or the environment to which an organism or other living system is differentially sensitive to be specified explicitly in terms of the information processing employed by the organism or system to detect and respond to that property. This includes properties often taken for granted, such as the existence of external objects or their embedding in three-dimensional (3d) space, and applies regardless of whether detection and/or response is recorded to memory or “reportable” via any specific assay.

MP treats information and energy as formally equivalent, explicitly enforces conservation of energy, and hence requires the thermodynamics of classical computation (Landauer 1961; Bennett 1982) to be represented explicitly. It requires the thermodynamic cost of memory to be accounted for at every scale, and automatically enforces a metabolism—cognition tradeoff.

Information processing functions, including writing to or reading from a classical memory, are specified in this framework using the scale- and implementation-independent, category-theoretic language of Barwise-Seligman (“token” – “type”) classifiers and their associated infomorphisms (Barwise and Seligman 1997; see Fields and Glazebrook 2019a for review and example applications). This generic specification language has been previously employed to characterize human object and event categorization using both abstraction and mereological hierarchies (Fields and Glazebrook 2019b) and human problem solving within a GNW context (Fields and Glazebrook 2020b). As it requires the reference frames employed to give internal, e.g. perceptual or motor, representations operational meaning to be explicitly specified, this language requires all semantic hypotheses to be formulated in explicit, experimentally accessible molecular or network-theoretic terms.

We outline the main components of MP in the next section, keeping formalism to the minimum and relying on previous work for details. We show that MBs and a notion of actionable or “meaningful” (Bateson 1972; Roederer 2005) information emerge naturally in an MP framework. We then discuss predictions of MP regarding perception, memory, and attention, providing examples primarily from basal systems for illustration. We show that requirements for coarse-graining and attention switching emerge naturally from the thermodynamics of classical computation. We then consider resource-usage monitoring as a requirement of homeo- or allostasis, and the emergence of a “self” as a representation of resource usage.

The re-interpretation of physics as an information theory was initiated by Feynman (1982) and Wheeler (1990), but has its roots in the work of Boltzmann (1995), who first recognized that reducing uncertainty has an energetic cost, and Shannon (1948), who linked uncertainty reduction with communication. It is deeply rooted in Wheeler’s (1983) idea of “observer-participancy”: that information can only be obtained by intervention, i.e. by actively posing a question (see Mermin 2018; Müller 2020 for recent, somewhat philosophical discussions). Physics is intrinsically quantum due to the quantization of obtainable information into bits: answers to yes/no questions (Wheeler 1983). The basic formalism of this quantum information theory is surprisingly simple, and rests on only the two assumptions specified below.

The simplest physical interactions are bipartite: some physical, i.e. quantum system A (conventionally called “Alice”) interacts with some other system B (“Bob”). We can therefore write:

where U is the joint system (“universe”) comprising A and B and ⊗ is the Hilbert-space tensor product. The dynamics of the joint system is completely described by a Hamiltonian operator:

where HA and HB are the internal or “self” interactions of A and B respectively and HAB is their mutual interaction. The operator HU has units of energy and drives state change via the Schrödinger equation:

where i = √-1, ℏ = h/2π, h is Planck’s constant, and |U(t)> is the time-dependent state of U. The mutual interaction HAB describes Alice’s observations of and actions on Bob, and Bob’s observations of and actions on Alice. The internal interactions HA and HB describe Alice’s and Bob’s internal state changes.

The above descriptions are completely generic and characterize any bipartite interaction between quantum systems (Nielsen and Chuang 2000). As a quantum system is simply a collection of physical degrees of freedom, this formalism makes, in particular, no assumptions about implementation: it applies equally to matter (e.g. molecules) or fields (e.g. electromagnetic fields as in McFadden 2020). We now impose two assumptions:

We require the dimension of the joint system U to be finite. This limits both A and B to finite energy resources and HAB to finite resolution.

We require the joint state |U> = |AB> to be separable, i.e. |U> = |AB> = |A>|B>. Equivalently, A and B are not entangled. In this case, HAB specifies the entire dependence of A on B and vice-versa, and we can talk about the state |A> of A and the state |B> of B independently.

With these two assumptions, the interaction HAB can be written, without loss of generality, as:

where k = A or B, i = 1 … N for finite N, kB is Boltzmann’s constant, Tk is k’s temperature, βk ≥ ln 2 is an inverse measure of k’s average thermodynamic efficiency that depends on the internal dynamics Hk, the αki(t) are N real functions such that, for any macroscopic time interval Δt:

and the Mki are N Hermitian operators with binary eigenvalues +1 or -1 that can be regarded as “questions to Nature” with yes/no answers [Wheeler 1983; see Fields and Glazebrook 2020a; Fields and Marcianò 2019, 2020b for details on Equations (5) and (6) and their interpretation]. The requirement that βk ≥ ln 2 in Equation (5) enforces Landauer’s principle of the finite per-bit cost of irreversibly acquiring an observational outcome (Landauer 1961; Bennett 1982), while Equation (6) imposes a normalization condition, equivalent to the requirement that the probability of acquiring some outcome whenever a measurement is made is unity. As we discuss below, Equations (5) and (6) are the only physical requirements needed to describe both free-energy exchange and information transfer in a generic bipartite interaction between finite quantum systems in a separable joint state. We can write Equation (5) in English as:

This is the conceptual heart of quantum information theory. It, like Equations (5) and (6), applies at any scale.

The complete generality of Equation (5) allows us to view any interaction between two finite quantum systems in a separable joint state as classical communication: the exchange of some finite number of finite bit strings during any finite time interval Δt. The interaction can be depicted as in Fig. 1: A and B can be regarded as alternately preparing, and then measuring N independent, mutually noninteracting quantum bits (qubits), e.g. photon polarizations, electron spins, or any other systems having two clearly distinguishable values of a manipulable state variable. In each cycle of interaction, A prepares the N qubits and then B measures them; the roles then reverse, with B preparing and A measuring. As N becomes large, the interaction may appear “continuous” at macroscopic scales, but remains constrained to finite resolution, and hence a finite bit rate, by the restriction to finite thermodynamic resources built into Equation (5). The time required for A to prepare, and then B to measure (or vice versa), the N systems can be regarded as the smallest possible “macroscopic” and hence effectively classical time interval (see Fields and Glazebrook 2020 for an explicit construction of this interval). The collection of all N qubits encodes, at each such minimal macroscopic time, one N-bit string specifying one of the 2N eigenvalues of HAB; each cycle of interaction is thus an exchange of N-bit strings, each one encoding an (in general different) eigenvalue.