Counting with Cilia: The Role of Morphological Computation in Basal Cognition Research

As it seems reasonable to assume that there are some systems that would qualify as “morphological control without informational input” (although no examples can be found in Müller and Hoffmann [6], and I was unable to provide any as well), then we can take the sets of systems exhibiting morphological control and performing morphological computation as intersecting. None of the examples that Müller and Hoffmann provide, however, qualify as even morphological control under the criteria proposed here: while the passive–dynamic walker has been discussed extensively above, we may also recall the example of gecko feet. Müller and Hoffmann argue that gecko feet, which have a particular hierarchical morphology enabling the appearance of appropriate van der Waals forces between the feet and vertical surfaces, allowing geckos to climb even smooth surfaces, are a case of morphological control. However, the adhesion of feet to surfaces is not control, just as the fact that a car hook’s shape, which enables it to attach a looped line and tow other vehicles (to remain in the realm of automotive examples), is not a form of control on its own. Control in the case of gecko feet relates to the forces the animal applies to adhere the feet to the surface and to move—and this is not performed morphologically.

Hence, under the cybernetic–mechanistic view accepted here, the tripartite distinction collapses into a two-fold one, although in some cases the original classes provide a good taxonomy of forms and purposes of morphological computation (see Figure 1 for a diagrammatic depiction).

The study of morphogenesis can be dated back to Aristotle [42], but in its modern version, it began with the mathematical and physical descriptions proposed by [43] and Turing [44]. Morphogenesis is the process responsible for the development of the shape of biological structures (cells, tissues, and organisms). While all biological systems undergo this process during development, many are capable of it throughout their lifetime, and hence, we can distinguish between developmental and regenerative morphogenesis, although their mechanisms are speculated to be largely identical.

Importantly, quite early attempts have already cast morphogenesis in terms of computation. One of the earliest explicit statements of this perspective comes from Wolpert and Lewis [45], p. 21, which attributes the following question to Sydney Brenner: “Is the adult (organism) computable from an egg?” This line of research has been expanded by, among others, Paulien Hogeweg (e.g., [46]), and more recently by Michael Levin and his collaborators (e.g., [47]), whose work is analyzed in more detail below. However, the computational nature of this process was already implicit in the reaction–diffusion system proposed by Turing, as it has been shown since that such systems can constitute analogue computers [48]. Hence, the mechanisms implementing morphogenesis in biological systems are at least partially computational in nature (according to the definition of computation accepted here, or “intrinsically computational” in terms of Müller and Hoffmann [6]).

There is, however, a discontinuity in this research tradition. Brenner, Wolpert, and Lewis, when they asked whether an organism is computable from an egg [45], p. 21, were interested if a researcher, given the genetic information in the egg, would be able to “compute”, i.e., predict, an adult organism phenotype from it. They proposed a model of the morphogenetic process in terms of cellular automata governed by a genetic program implemented by binary gene switching networks. However, their account is metaphysically modest and makes no claims for the “computational nature” of the process thus described. Similarly in the work of Hogeweg: she has proposed computational models and simulations of morphogenetic processes (e.g., [46,49]) thath ave been described as attempts at computing an organism [50]. However, whenever she references the word “computing” or calls the morphogenetic processes “informatic”, she uses “scare quotes” around these words, differentiating computing from ‘computing’. These implicit assumptions can be interpreted in instrumentalist terms: Hogeweg uses the language of ‘computation’, computer simulations in particular, as it offers a good explanatory framework. She does not presuppose, however, that these explanations translate into actual computations in nature but rather into mechanisms that can be partially described as “processing information”.

Levin and collaborators’ work (e.g., [47,51,52,53]) introduces a major shift in this regard. Although he states that his work on morphogenesis follows the tradition discussed in the previous paragraph, he does not merely use the computational language for explanation but rather claims that the processes of morphogenesis are literally computational and process information. This is clearly visible in the way they update the aforementioned quote originally due Brenner: instead of asking how can we, scientists, compute an organism from an egg, Manicka and Levin ask “How does an embryo compute its own pattern?” [47], (p. 1; emphasis added). They further strengthen this perspective by exploring the possibility of using insights from the study of morphogenesis in the development of artificial computational platforms. Contrasted with Hogeweg’s instrumentalist view, Levin’s metaphysical perspective on scientific concepts can be best accounted for by the model-based philosophy of science (e.g., [54,55,56]), and his claims about computations can be understood as identifying robust properties [56] of the natural world.

Development and regeneration require information processing to meet several intertwined goals: patterning requires the cell to process (relative) locational information to move to an appropriate part of the growing organism; cytodifferentiation also requires input of positional information together with a model of the general body plan for developing appropriate cell types in appropriate places; finally, the dividing cells need to solve the most difficult problem of tissue growth: they require information about the structure of the organism attained thus far to be able to stop divisions when the correct shape is reached. While some of this information is encoded genetically (e.g., available cell types), existing experimental work shows that genetic information is hardly the main input into this computation process. Levin’s experimental work provides two important lines of argument in this regard.

First, while traditionally DNA is taken to provide information about shape, Levin’s work in planaria, which have astonishing regenerative capabilities and can regenerate a whole organism from even the tiniest slice, proves that much of the information is, in fact, encoded in bioelectric signals between cells (e.g., [57,58]; see also the review in [53]). Without altering the genome, Levin and collaborators induced bioelectrical changes by modifying ion channel expression in regenerating worms. This intervention modified the patterning so that the organism regenerated a second head instead of a tail (a body plan with 0, 1, or 4 heads can be thus induced, or even a head morphology of a different species, i.e., a flat head in a round-head worm species or vice versa). Furthermore, this new body plan became encoded in the bioelectrical activity, so that after further incisions, the new two-heads-no-tail body plan was retained without any further intervention [58].

Second, DNA cannot be set to encode a set of morphogenetic movements, or steps, but rather an “error minimization process” or a homeostatic loop [53], p. 15 (see also the work exploring connections of this approach to the free energy principle framework: [59]; [60]). The developmental and regenerative processes are highly robust, in that they are able to attain the correct result from a variety of abnormal starting conditions. Levin’s lab has shown this by modifying developing tadpoles. After scrambling the position of their facial organs, the developing frog alters its developmental ‘program’ to achieve the correct face shape [61]. Even more astonishingly, a tadpole created with an eye in an incorrect position (i.e., located on the tail) is still able to develop the eye correctly, together with the optic nerve that connects in a novel location to the spinal cord, allowing the tadpole to see quite well [62].

Both these examples show that morphogenesis is not a ‘blind’ feedforward process but rather an adaptive process that achieves “fixed ends with varying means”—what William James has proposed as a minimal criterion for mentality [53,63]. Hence, this research casts morphogenesis as a complex information processing and control process implemented by—and over—an ever-changing medium. This interaction of structure and information processing clearly puts morphogenesis in the domain of morphological computation, involving both morphological perception and morphological control. At the same time, this perspective introduces restrictions regarding the admissible concept of computation that it can be understood with: interpreting it through the lens of the semantic approach, particularly in the flavor proposed by Müller and Hoffmann [6], would require identifying semantic representation of abstract computational states in cell states and specifying a user of those computations. This is not the case: the computations are performed online by individual cells, and they are used by those very cells. However, the computations cannot be individuated on this level. Consider the idea of a “morphogenetic field” [64] on which Levin builds (e.g., [65]): this model takes morphogenesis to be guided by non-local interactions of cells (physiological or chemical gradients, steady-state bioelectrical properties, etc.), suggesting that morphogenesis is best understood not on the level of a cell, but rather at higher levels of organization. As a result, each cell does not need to represent its position, but rather it responds to available information, giving rise to the computational process at a higher level of organization.

The “computational entity” or “user”, if we were to identify one in the case of morphogenesis, would need to be considered an emergent property of the process, appearing on that higher level of interacting cells. As such, it is impossible to establish a relevant encoding/decoding relation between physical and abstract states. The criteria of computation are not met under the semantic view.

However, under the mechanistic perspective, morphogenesis can be considered an example of computation: the cells process available information (in the form of available physiological or chemical gradients, bioelectrical properties, etc.) and use them to guide their actions. Further, as we have seen from the above discussion of the research on this biological phenomenon, the complexity of these processes exceeds what can be accounted for on the level of physicochemical interactions, requiring scientists to switch to the “informational stance” (cf. [33] and Section 4). Hence, under the mechanistic view, the more realistic approach of Levin and collaborators is, in fact, warranted.

In the introduction, I briefly summarized the history of the concept of “morphological computation”, highlighting its roots in the embodied traditionally anti-computational view of the mind. The emerging research field of basal cognition is an intellectual descendant of this tradition. Basal cognition focuses on “the fundamental processes and mechanisms that enabled organisms to track some environmental states and act appropriately to ensure survival (finding food, avoiding danger) and reproduction long before nervous systems […] evolved” [66], p. 4. As such, it “describes a toolkit of biological capacities involved in becoming familiar with, valuing and exploring, exploiting or evading environmental features in the furtherance of existential goals” [66], p. 5. As a field, it aims to strengthen the inclusion of life science and biology—in particular, developmental biology and regenerative medicine—into the cognitive sciences. Basal cognition, similarly to the embodied framework in the late 20th century, aims to expand our understanding of cognition and of mechanisms that may be responsible for cognitive faculties by pointing not only to unconventional cognitive media but also to hitherto overlooked phenomena. Developmental and regenerative morphogenesis is one of the primary examples proponents of this field refer to. As already mentioned in the previous section, the adaptive nature of morphogenetic processes seems to meet classical criteria of mentality. In more contemporary terms, Levin [53], p. 2 expands James’ notion by proposing to understand intelligence as the degree of competence in navigating spaces.

When thinking of the physical, 3D space, this definition includes processes we normally conceive of as intelligent (e.g., rats’ ability to navigate mazes) or that are more controversial (e.g., slime molds’ and oil droplets’ ability to navigate mazes; see [67,68]). However, this notion can also be expanded to other “spaces”: Levin mentions primarily “morphospace”—the space of possible morphological configurations of an organism explored in the process of morphogenesis, which makes morphogenesis conform to this definition of intelligence (other spaces he lists include transcriptional spaces explored by mRNA, physiological spaces explored, e.g., by ion channels, etc.). Importantly, Levin regards intelligence as a continuum in this regard so that we can discuss systems possessing less or more intelligence, and hence need not treat rats and slime moulds as equivalent (as rats can be regarded as more competent in navigating spaces).

Morphological computation can play the role of a central conceptual hub relating the research in basal cognition to the investigation of human brains in computational neuroscience. In a way, I want to suggest that it can play a key role in understanding the principles of “scaling up” cognition in its “endless forms most beautiful” and complex (for further suggestions about “scaling up”, see [69]).

While the brain does not necessarily involve “morphological” computation, the achievements of computational neuroscience (see, e.g., [70]) as well as significant developments of artificial intelligence, in particular in machine learning and neural network architectures (e.g., [71,72,73]) secured a central position in cognitive science for the computational theory of mind. Hence, showing that basal organisms not only employ equivalent mechanisms (e.g., as is the case with ion channels in bacteria; see [74]) but also are capable of performing complex computational functions supports the analogies between human and basal cognitive capacities that basal cognition research ultimately tries to draw.

Furthermore, basal cognition draws heavily on the investigations within synthetic biology and soft robotics (e.g., [75,76,77,78]), which is also one of the domains that supported the emergence of the ideas of morphological computation (e.g., studies of octopus arm control; see [79,80]). This concept, then, appears to permeate the research tradition, even if hitherto its central role was not appreciated.

Importantly, despite the continued tensions between proponents of computationalist and anti-computationalist theories of mind, recently, a middle way has been considered by some researchers. Researchers such as Villalobos and Dewhurst [9] and Korbak [8] have attempted to reconcile the radically anti-computationalist, anti-representationalist view of the enactive approach to the mind with computationalist perspectives. This attempt resulted in a view that Korbak calls “computational enactivism”. Morphological computation allows us to associate basal cognition with this idea. Computational enactivism argues that the autopoietic requirements set by enactivism are compatible with, or in a stronger form—co-dependent on, the computations performed by certain physical systems, in particular living organisms. If we consider morphological computation in the context of processes such as morphogenesis, where it is tasked with implementing a particular homeostatic loop and overseeing this autonomous, self-organizing — autopoietic — process, it provides significant support for the “computational enactivist” position. Consider further the example of the Ashby’s Homeostat, mentioned earlier, which has served as the source of inspiration for the theory of autopoiesis [31]. In virtue of this, we can regard the computational nature of morphogenetic processes as a proof-of-principle of a simple cognitive system conforming to the propositions introduced within computational enactivism and hence offering a way out of the existing binary.

An argument similar in spirit may be found in a paper by Beer and Williams [81]. The authors develop computer simulations of simple agents that are evolved to perform a categorization task. They analyze these simulations using the tools of information theory and dynamical systems theory—tools that are traditionally connected with the deeply opposed theoretical strands in philosophy of mind and cognitive science: computationalism and the embodied mind theory, respectively. They conclude that these two forms of mathematical descriptions are precisely that—different mathematical languages that can be used to elucidate different properties of the systems under study. These languages are not at odds, and although they have different purposes, their descriptions are often consistent and complementary. Crucially, this arises from a methodological approach similar to the one suggested here, namely, from focusing on a “minimal” system that exhibits sufficient complexity of behavior to be considered under the cognitive lens but at the same time is simple enough to be amenable to fine-grained analysis, including empirical studies in the lab.

While my argument takes the conclusions of Beer and Williams’s work for granted and builds on the commonalities between information theory and dynamical systems theory, the analysis provided here aimed to show that two properties taken to be at odds, autopoiesis and computation, can, in fact, be identified in the same system as interacting and co-supporting. In this way, I suggest regarding morphogenesis—the metaphorical “counting with cilia”—as an initial example of how computational descriptions may be of use for accounting for autopoietic and self-organizing processes.

This paper discusses the concept of “morphological computation”, pointing out that the existing critical analysis of Müller and Hoffmann [6] relies on a particular definition of computation that implicitly falls under the scope of the semantic view of computational implementation. Accepting instead the alternative, mechanistic perspective on physical computation offers a significantly different classification of processes that can be considered to be performing morphological computation. In fact, under the mechanistic view, including some insights provided by cybernetics, the tripartite distinction Müller and Hoffmann introduce between morphological control, perception, and computation proper collapses into a simpler view that distinguishes a limited number of very trivial cases of non-computational morphological control from morphological information processing that is “really” computational and can involve control and perception as well as human-designed computation.

I further substantiate these claims by analyzing an example of morphological computation that has been overlooked so far in this literature. I refer to studies of developmental and regenerative morphogenesis that have been increasingly regarded as computational in nature, and I point out that the computations in question are, in fact, performed morphologically. This example connects the analysis back to the debates within cognitive science, philosophy of mind, and philosophy of biology, as it is one of the central cases studied under the label of “basal cognition”. Its importance can hardly be overstated, as it puts basal cognition at the center of the stalled debate between computationalists and anti-computationalists, providing important support for a recently proposed “middle-way” approach—computational enactivism—aiming to reconcile insights from these opposite traditions.

Hence, this discussion of morphological computation offers the potential of significantly influencing our understanding of cognitive systems and intelligence, helping the project of basal cognition and allowing it to make novel connections to established fields, such as computational neuroscience in particular.

I want to thank David Harrison, Urte Laukaityte, and Marcin Miłkowski for discussions, suggestions, and comments that allowed me to write this paper and helped me with the final draft. I want to also thank the audience at the Conference of the German Society for Philosophy of Science held in Berlin in August 2022, where a draft version of this work was presented, in particular my co-symposiasts—Krys Dołęga and Hajo Greif—for helpful questions and comments. Finally, I wish to thank two reviewers for Entropy for constructive and supportive suggestions.

Publication of this paper was funded by an internal grant at the University of Warsaw as part of the program Excellence Initiative—Research University.