最小物理主义作为认知和意识的无尺度基础

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

Abstract

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

论文提出17个预测

Prediction 1: Moving in 3d space does not require a QRF for 3d space, and hence does not require experiencing 3d space. E. coli chemotaxis provides an example. E. coli has a 1d spatial QRF: its body axis, with (mainly) anterior chemoreceptors and (mainly) posterior flagella. Directed “approach” motion is along this axis; undirected “tumbling” re-orients this axis randomly in the 3d “lab” frame of an observer equipped with a 3d QRF. E. coli has no known means of computing the relative angle between its preand post-tumbling linear motion, i.e. it has no known 3d QRF; hence tumbling appears to implement a 3d random walk (Wadhams and Armitage 2004). Colonial microbes living in planar mats, on the other hand, can potentially use differential cell–cell or cell–substrate interactions to distinguish left from right (Jauffred et al. 2017) and hence establish a 2d QRF; microbes inhabiting multi-species 3d mats with vertical division of labor may have 3d QRFs (Prieto-Barajas et al. 2018).

Differentiated cells of multicellular eukaryotes clearly employ such QRFs (e.g. Bajpai et al. 2021); interestingly, multi-axis morphology correlates with the presence of neurons and appears to be directly instructed by neural signaling (Fields et al. 2020).

How the representation of 3d space in migrating cells is coupled to the representations of cell- or extra-cellular surface characteristics, bioelectric and morphogen gradients, and other morphogenetic signals remains a central question in developmental biology (Grossberg 1978; Pezzulo and Levin 2015).

Prediction 2: Successful interaction with an object does not require a QRF that identifies that object, and hence does not require experiencing that object. E. coli mating provides an example: the mating pilus extends randomly in 3d space, and is tipped by an adhesin of unknown specificity (Cabezo ´n et al.

2015). Commonplace lateral gene transfer (LGT) between members of distant microbial lineages (Robbins et al. 2016) suggests that mating without mate detection is routine in the microbial world; viral-mediated gene transfer and direct uptake of nucleic acids from the environment provide even more extreme examples. Communal microbes such as Myxococcus xanthus that differentiate kin from nonkin even below the species level, however, appear to have sophisticated, though yet uncharacterized, QRFs for other organisms (Mu~ noz-Dorado et al. 2016; Thiery and Kaimer 2020). Fungi that engage in differential anastomosis appear similarly equipped (de la Providencia et al. 2005).

Cell-type identification QRFs appear to be implemented in part bioelectrically in multi-species microbial mats (Humphries et al.

2017; Yang et al. 2020); the ubiquitous use of bioelectric signaling in fungi suggests that this may be the case for fungal cell-type QRFs as well.

Prediction 3: Successful causation does not require a mechanism for detecting causation, and hence does not require experiencing causation. Hunting swarms of M. xanthus kill and eat other microbes, but appear to have separate, noncommunicating detection systems for prey species and edible prey components (Thiery and Kaimer 2020). Hence they have no way of causally associating the killing of prey with the subsequent availability of edible prey components.

These can be summarized by the following, which recognizes the key role of memory in enabling the experiences of space, objects, and causation:

Prediction 4: Having a memory does not require a QRF for linear time, and hence does not require experiencing time or retrievable memory. All organisms have genomes that record their phylogenetic history, but they have no mechanism for reading their previous genomic states. The genome does not, therefore, function as an internal QRF for linear time. This applies, in particular, to us, although we can employ the genomes of other organisms as external linear time QRFs (Kumar 2005).

Prediction 5: All retrievable biological memories are stigmergic. Beginning with bacteria (Gloag et al. 2015), biological systems ubiquitously employ stigmergic memories (Heylighen 2016). This is not a surprising observation to be explained, but rather an empirical confirmation in MP. The idea that the experiencing mind “extends” (Clark 1998) into the environment via stigmergic memory is a direct consequence of quantum theory.

The stigmergic nature of memory can be reconciled with the experience of memory as an internal, private phenomenon only if the agent A is compartmentalized by internalizing part of B to provide an internal boundary C on which classical information can be encoded. This internal boundary imposes a separability condition jA> ¼ jA1>jA2> as shown in Fig. 3; A1becomes part of the “environment” of A2 and vice-versa. The interaction between A1and A2can be written in the form of Equation (5); hence information flow across C is bidirectional classical communication between the components A1and A2. If we view A2 as implementing perception and A1 as implementing action, this communication, together with the environment’s response, forms a closed loop. Hence we have:

Prediction 6: Internal awareness requires internal boundaries.

Any system A capable of experiencing internal memories, in particular, has a separable internal state and positive integrated information U. Internal memories are built into all systems that qualify as conscious in IIT (Oizumi et al. 2014, see especially Fig.

19) and are the basis for such systems having U > 0. Here, we see this as a consequence of quantum information theory.

As discussed above in connection with MBs, this prediction links separability in its quantum-theoretic sense of state distinguishability with separability in its classical sense of separation by a boundary: the internal boundary C functions as an MB that encodes classical information and imposes conditional independence. It provides for a general expectation that living systems will be compartmentalized by internal boundaries on which classical information can be encoded. Because MBs limit, as well as enabling, information transfer, it also predicts a systematic inability to determine the source of a memory

Prediction 7: Compartmentalized systems cannot determine the sources of their encoded memories. Systems can be expected to behave as if memories they encode reflect their own past experience, whether they do or not. Studies of memory transplantation (Pietsch and Schneider 1969) and falsememory induction (Ramirez et al. 2013) in nonhuman animals provide mechanistic support for this prediction (see Levin 2020 for additional examples and discussion), as does the psychology (Henkel and Carbuto 2008) and neuroscience (Straube 2012) of false-memory induction in humans.

Organisms record memories as messages to their future selves. Neither the mechanism that recorded a memory, the internal or environmental events that induced recording, nor the context in which the recording occurred are discoverable, however, when the memory is later retrieved. This fundamental uncertainty about the sources of memories can be seen as a consequence of a more general uncertainty about whether QRFs are shared, either by distinct observers at a single time, or by a single observer across time. This question of QRF sharing is provably finite Turing undecidable (Fields et al. 2021).

The compartmentalization in Fig. 3 can be arbitrarily generalized:

Prediction 8: Experiential complexity scales with internal compartmentalization. Evolution “discovered” the benefits of internal compartmentalization ca. 3.5 billion years ago with the development of microbial biofilms exhibiting differential exposure to the open environment and division of metabolic labor (Stal 2012). The organelles of eukaryotic cells, including internal membrane complexes such as the Golgi apparatus, are canonical intracellular compartments. Multicellularity is the most common form of macroscopic compartmentalization, but is not required; examples such as Acetabularia (Schweiger 1969), glass sponges (Leys 2016), and syncitial fungi (Roper et al. 2015) all illustrate complex functional compartmentalization within single giant cells.

This prediction is clearly in line with the expectations of IIT, and again provides a physical basis for these expectations.

Prediction 9: Memory stability scales with the frequency of read/write cycles. The stabilization of classical bit values by repeated cycles of preparation followed by observation is called the “quantum Zeno effect” (Misra and Sudarshan 1977); the probability that the state remains stable is proportional to the frequency of observations.

Hence memory decay—“forgetting”—is predicted whenever memories are not routinely accessed. This is broadly observed across phylogeny. Repair systems for nucleic acids and degradation systems for proteins provide molecular-scale examples.

Rewriting classical information costs free energy, i.e.

requires metabolism as discussed below. This requirement for read/write cycles suggests that compartmentalization plays a key role in the implementation of linear time QRFs:

Prediction 10: Any ordered sequence of separate memories together with a comparison function can serve as a linear time QRF. Trajectories, including looming, are the simplest linear time QRFs; in the limit, they may support only sequential comparisons and hence only distinguish “then” from “now.” Insects are capable of at least short-sequence linear time perception; spiders are capable of longer sequence perception (Japyassu ánd Laland 2017). Merely executing a fixed action pattern does not require experiencing linear time, although it clearly requires an internal clock. Molecular cell-cycle clocks are as old as life, and circadian clocks are as old as cyanobacteria (Johnson 2004); both are highly conserved across phylogeny (Doherty and Kay 2010). Possessing a molecular clock does not, by itself, enable time perception.

Prediction 11: Time and object/feature identity are duals.

Perceiving a trajectory requires perceiving an object or feature executing that trajectory; conversely, motion perception is the basis of object identity (Fields 2011). Whether insects or relatively low-complexity vertebrates are aware of objects as such or only features of their environments is unclear; bees at least have robust spatial memory and feature perception, and may recognize objects as such (Chittka 2017), while fish appear to recognize conspecifics as distinct objects (Sovrano et al. 2018).

Spiders are capable of robust object perception and object-directed planning (Japyassu ánd Laland 2017), as are cephalopods (Mather 2019), birds and mammals.

Prediction 12: Organisms only require the energy needed to maintain their classical states. As seen above, these are encoded on either exterior or intercompartmental boundaries. An organism’s energy budget must, therefore, supply the free energy needed to maintain, via Zeno-effect read/write cycles, the classical states of their compartment boundaries. Nonboundary states can remain quantum, evolve unitarily, and consume no free energy. Only the inputs and outputs of these quantum computations are classically encoded, all on boundaries.

Clearly not all boundary-localized processes are classical; electron-transport processes operating in the THz range could consume a cell’s entire energy budget if fully classical. The technical difficulty of observing nonclassical behavior in such systems (e.g. Cao et al. 2020) is not surprising. All of our observational outcomes are classical by definition; observing quantum coherence requires observing expectation violations in probability distributions over recorded classical events (e.g.

Mermin 1993). This suggests that an indirect approach to quantitating nonclassicality in biological systems is required. From the above considerations, stable memories provide a quantitative lower limit on classicality, while the cellular energy budget provides an upper limit. Determining what states a cell or organism commits free energy to maintain, i.e. what the set points for homeostasis/allostasis are, may be the best approach to estimating the net classicality of biologically encoded information.

Biological encodings of classical information have lower limits of 1–2 nm in radius, e.g. the size of a typical protein active site (Liang et al. 1998) or a gap-junction channel (Sosinsky and Nicholson 2005), and about 200 fs in time, e.g. the response time of rhodopsin to photons (Wang et al. 1994). Cellular response times, even for bioelectric responses, are orders of magnitude larger and involve much larger areas. This loss in resolution is a consequence of bioenergetics:

Prediction 13: Actionable classical encodings are coarsegrained. Actionability requires irreversible encoding as in Fig. 3.

The free energy to support this encoding must come from B, and hence must consume some of the bits encoded on B as fuel.

The information encoded by these bits is irreversibly lost to A; hence all of A’s irreversible encodings are coarse-grained.

Any system A that encodes information irreversibly is, therefore, faced with a choice that its computational architecture must resolve: the tradeoff between preserving information via memory and losing information due to coarse-graining. A flexible solution to this tradeoff is to devote memory resources to only some input information, i.e. to the results computed by only some QRFs, allowing these selected results to be recorded at higher resolution while recording others either at low resolution or not at all.

Prediction 14: Living systems in complex, dynamic environments will evolve attention-switching systems. Attention has long been associated with consciousness (Engel and Singer 2001) and attention allocation is one of the main functions of competition for access to the workspace in GNW models (Dehaene and Naccache 2001), whether formulated in terms of the “rich club” (Sporns 2013), “connective core” (Shanahan 2012) or “giant component” (Wallace 2005) of the larger network.

Indeed the “self” has been described as a working model of attention allocation (Graziano and Webb 2014; see also below). In humans, attention can drive entirely illusory object perception (Ongchoco and Scholl 2019), consistent with active-inference models of Bayesian-precision allocation (Kanai et al. 2015). Here, we see a requirement for active attention as a consequence of the thermodynamic requirements of classically encoding information.

The approach/avoid switching in simple chemotactic systems such as E. coli provides a basal model of the switch between active exploration and expectation revision at the heart of active inference theory (Friston 2010, 2013). Such models suggest that every such switch is governed by a reference value set by some QRF.

Dorsal/ventral attention system switching in humans (Vossel et al.

2014) is sensitive to a vast array of expectations, the reference values of which are unknown and in at least some cases subject to considerable individual variation. Salience assignments driving both exploratory and reactive behavior are, in particular, highly dependent on individual experience, strongly coupled to the core self-representation and the reward system, and subject to variation in both prosocial and pathological directions (Uddin 2015).

Much of ethology can be viewed as the comparative study of salience. Understanding how QRFs that regulate salience vary across phylogeny will be a major step toward answering the “what is it like?” question in a systematic way.

Prediction 15: The “self” comprises three core monitoring functions, for free-energy availability, physiological status, and organismal integrity, and three core response functions, free-energy acquisition, physiological damage control, and defense against parasites and other invaders. These will be found in every organism. Indeed they are found even in E. coli, which has inducible metabolite acquisition and digestion systems (Jacob and Monod 1961), the generalized “heat shock”stress response system (Burdon 1986), and restriction enzymes that detect and destroy foreign, e.g. viral DNA (Horiuchi and Zinder 1972). All of these responses act to restore an overall homeostatic setpoint, i.e. an expected nonequilibrium state; hence they can all be viewed as acting to minimize environmental variational free energy or Bayesian expectation violation (Friston 2010; 2013).

Specialized molecular pathways for the core functions of the self are supplemented by specialized intercellular communication pathways in multicellular organisms and by inter-individual communication pathways in social organisms, e.g.

eusocial insects (Robinson 1992) or humans. Networks of neurons support feeding, locomotion (a primary stress response) and defense already in Cnidarians (Bosch et al. 2017) and at least feeding and locomotion in Ctenophores (Moroz 2015); these functions become far more complex in bilaterian animals, particularly in active animals including arthropods, cephalopods, and vertebrates (Liebeskind et al. 2017), and form the basis of “core consciousness”in Damasio’s (2000) framework.

Considerable evidence now indicates that interoception in humans, and so presumably in mammals generally, employs a predictive coding mechanism (Hohwy 2013; Seth 2013) and is strongly coupled to the core self-representation and the salience network via the insula—cingulate—orbitofrontal loop (Craig 2010; Uddin 2015; Seth and Tsakiris 2018). This predictive coding system manages homeostasis/allostasis across the scale hierarchy from cells to organ systems (Corcoran and Hohwy 2019) and couples interoception to exteroception and proprioception (Barrett and Simmons 2015; Seth 2013; Barrett 2017; Seth and Tsakiris 2018). Intriguingly, emotional and stress responses, core components of the self, are highly sensitive to gut microbiome activity in humans and other animals (Vuong et al. 2017; Sarkar et al. 2018). As all multicellular eukaryotes have obligate symbiotic microbiomes (hence are “holobionts”; Gilbert 2014), one can expect that microbial contributions to stress detection and response are universal.

The thermodynamic costs of memory impose a particular restriction on the self, which predatory eukaryotic unicells such as Paramecium, animals, and quite possibly plants solve by weighting the less-certain future lower in resource priority than the (by default assumed to be) more-certain past:

Prediction 16: Organisms bias their “cognitive light cones”(Levin 2019), their representational and computational boundaries of concern or goal-directedness, toward the past. Memory, in other words, takes precedence over planning. The extent to which food caching, cooperative hunt organization, and other future-directed activities by nonhuman animals provides evidence of deliberative, explicit planning remains controversial (Bayne et al. 2019), with some arguing that all forms of imaginative “mental time travel” are human-specific (Suddendorf and Corballis 2007). It is, however, clear that explicit planning requires classically encoded representations of future events and so competes for resources with memory. As planning also requires explicit memory, it cannot win this competition.

The free-energy costs of mental time travel in either direction constrain it to an “off-line” activity when organisms are faced with rapid change that requires high-resolution perception and action. Such constraints also apply to the real-time representational costs of the self. Hence we can predict:

Prediction 17: Increasing real-time response requirements will disrupt encoding of the self-representation. This is observed in humans in “flow” states (Csikszentmiha ´lyi 1990), in highly automated, expertise-dependent activities including most social interaction (Bargh and Ferguson 2000; Bargh et al. 2012), and in experimental manipulations in which fatigue in various modalities affects cognitive performance (Massar et al. 2018). The extent to which nonhuman animals are able to apply “theory of mind” reasoning to themselves, and hence maintain a metacognitive selfrepresentation, remains controversial (Martin and Santos 2016).

The picture that emerges from these energetic considerations is of a self-representation with critical functions and deep evolutionary roots (Levin 2020), but with only limited and transient exposure to awareness via explicit encoding.

完整内容请参考原论文