Although some commentators may argue that the code metaphor has been set up as a strawman by Brette, it takes little effort to catalogue its ubiquitous use in the neurosciences over the past half-century. The influence of this conceptual framing has been reinforced by the highly successful technique of recording the spike trains of individual neurons in response to stimulus presentations. Thus, a correlation between specific stimulus features and rapid spike production by a specific neuron is presumed to license the claim that this neuronal activity in some way encodes that stimulus feature. Yet correlated neural activity of any of the potentially large number of neurons located anywhere along the path linking the initial registration of the stimulus to a specific neuron thereby caused to become highly active could likewise be understood as encoding that same stimulus.
It must indeed be the case that signal transduction from neuron to neuron is in some way necessary for brain processes to be about anything. But the problem with treating correlations or covariant dynamics as the sufficient basis for explaining cognitive or even perceptual functions is that the designated aboutness is only in the eye of the experimenter, not an intrinsic property of neural processes. But is this even a useful heuristic fiction? We agree with Brette's assessment that framing the problem in these terms has become more problematic than helpful, precisely because it is often accepted as an explanation when instead it is merely descriptive.
But having exposed the dangers of employing the code metaphor in cognitive neuroscience, we need a more appropriate alternative. One standard response is to abandon the concept of representation altogether and use only the language of covarying coupled dynamics. But this merely replaces an atomistic mapping relation with a dynamical mapping relation. In this commentary we argue that joining the insights provided by a dynamical systems perspective with the semiotic framework of C. S. Peirce (Hartshorne & Weiss Reference Hartshorne and Weiss1931–1963) can provide a middle path between the atomistic reductionism of an encoding paradigm and the “dynamics only” approaches. This is because each of these frameworks addresses key weaknesses in the other. A dynamical systems perspective can help to ground the notion of interpretation that is framed only in formal terms within semiotic theories, whereas a semiotic perspective can help to disentangle the distinctive roles of dynamics and form in dynamical systems theories.
A key concept necessary to bridge these frameworks is the concept of constraint. This is a fundamental concept both in dynamical systems theories and in information theory, which provides a precise analytical tool for characterizing relations of form. Two seminal thinkers in theoretical biology – Howard Pattee (Reference Pattee, Pattee and Braziller1973) and Michael Polanyi (Reference Polanyi1968) – have argued that focusing on the relationship between constraints and dynamics is essential if we are to understand the fundamental logic of living processes. They each stress that these complementary aspects of a living process must at the same time be functionally interdependent but physically independent.
Thus, Polanyi (Reference Polanyi1968) argues that the constraints that organize the dynamical processes of life are “Irreducible higher principles [that] are additional to the laws of physics and chemistry” (p. 160). And similarly, Pattee (Reference Pattee, Coombs and Sulcoski1997) points out that “Physical laws and semiotic controls require disjoint, complementary modes of conceptualization and description. Laws are global and inexorable. Controls are local and conditional. Life originated with semiotic controls” (p. 9).
For Pattee, “semiotic controls” exist in the form of “non-integrable constraints” that involve a “necessary epistemic cut between the coherent physical dynamics and its rate-independent semiotic description.” Importantly, constraints are off-loadable onto artifacts and their structure/form. This facilitates and is necessary for the preservation of constraints across potential changes of dynamics.
Although the Peircean semiotic framework is often treated as though it is a structuralist typology of sign types, it is also compatible with a dynamical framework. This is because Peirce understood the interpretation of signs to be constituted by the production of signs (interpretant production), and because he also understood mental processes to be sign production in this same sense and not the locus of some intrinsically meaningful mental token. In this framework, there is no simple mapping between a sign vehicle (signifier) and what it refers to (signified). Rather, sign vehicles are physical forms that mark phases of a process of forms-modifying-the-production-of-other-forms. There is no final form that marks the terminus of the process. Indeed, for Peirce, what he calls the “final interpretant” is in effect a habit of interpretant generation, that is, a process organization.
To bring this semiotic analysis into alignment with dynamical theories it is necessary to understand signs as sources of constraint on dynamics and, as Pattee and Polanyi independently recognized, to understand that semiotic constraints must necessarily be distinct from the dynamics that they control. Both the code metaphor and the conception of cognition as mere correlated dynamics ignore this distinction. But a dynamical semiotic approach that treats sign vehicles (whether words or constraints on neural activity) as information structures that control the dynamics of the production of other sign vehicles can preserve the concept of representation, which satisfies the requirements specified by Brette (see also Bickhard Reference Bickhard2009), without reducing it to a mapping relationship. It also provides a context for distinguishing modes of semiotic relations and semiotic differentiation processes in terms of different modes of constraint.
This reframing also illuminates the relationships of this debate to debates in information theory on the one hand and linguistic theory on the other.
Information theory, following the pioneering insights of Claude Shannon (Reference Shannon1948), notoriously avoids any effort to deal with representational content or normative properties, such as accuracy and truth. And yet it provides a precise formalization for measuring the information content of a medium or a message within that medium, for optimally encoding a signal, and for compensating for noise (though both signal and noise are normative distinctions). Shannon's measure of information in a message is assessed in terms of the uncertainty that is thereby reduced by virtue of the constraint on its possible entropy. Thus, implicitly, it treats whatever semiotic value can be provided in a message as a function of this constraint.
Linguistics has also struggled with the code metaphor (as reflected in the famous signifier-signified relationship described by Ferdinand de Saussure [1959]), and critics of this conception have likewise attempted to reframe “languaging” in purely correlated dynamical terms. The assumption of a completely unstructured “arbitrary” correspondence between word sounds and meanings has led to a dilemma called the symbol grounding problem by Steven Harnad (Reference Harnad1990b), on one hand, and has motivated theories of innate or culturally imposed grammatical principles and syntactic rules to explain the complex structure of languages, on the other. But a reframing of the linguistic structures in terms of constraints on dynamically grounded communicative processes also resolves these dilemmas (Deacon Reference Deacon, Hayes and Wilson2018; Rączaszek-Leonardi Reference Raczaszek-Leonardi2009; Rączaszek-Leonardi et al. Reference Raczaszek-Leonardi, Nomikou, Rohlfing and Deacon2018).
Although some commentators may argue that the code metaphor has been set up as a strawman by Brette, it takes little effort to catalogue its ubiquitous use in the neurosciences over the past half-century. The influence of this conceptual framing has been reinforced by the highly successful technique of recording the spike trains of individual neurons in response to stimulus presentations. Thus, a correlation between specific stimulus features and rapid spike production by a specific neuron is presumed to license the claim that this neuronal activity in some way encodes that stimulus feature. Yet correlated neural activity of any of the potentially large number of neurons located anywhere along the path linking the initial registration of the stimulus to a specific neuron thereby caused to become highly active could likewise be understood as encoding that same stimulus.
It must indeed be the case that signal transduction from neuron to neuron is in some way necessary for brain processes to be about anything. But the problem with treating correlations or covariant dynamics as the sufficient basis for explaining cognitive or even perceptual functions is that the designated aboutness is only in the eye of the experimenter, not an intrinsic property of neural processes. But is this even a useful heuristic fiction? We agree with Brette's assessment that framing the problem in these terms has become more problematic than helpful, precisely because it is often accepted as an explanation when instead it is merely descriptive.
But having exposed the dangers of employing the code metaphor in cognitive neuroscience, we need a more appropriate alternative. One standard response is to abandon the concept of representation altogether and use only the language of covarying coupled dynamics. But this merely replaces an atomistic mapping relation with a dynamical mapping relation. In this commentary we argue that joining the insights provided by a dynamical systems perspective with the semiotic framework of C. S. Peirce (Hartshorne & Weiss Reference Hartshorne and Weiss1931–1963) can provide a middle path between the atomistic reductionism of an encoding paradigm and the “dynamics only” approaches. This is because each of these frameworks addresses key weaknesses in the other. A dynamical systems perspective can help to ground the notion of interpretation that is framed only in formal terms within semiotic theories, whereas a semiotic perspective can help to disentangle the distinctive roles of dynamics and form in dynamical systems theories.
A key concept necessary to bridge these frameworks is the concept of constraint. This is a fundamental concept both in dynamical systems theories and in information theory, which provides a precise analytical tool for characterizing relations of form. Two seminal thinkers in theoretical biology – Howard Pattee (Reference Pattee, Pattee and Braziller1973) and Michael Polanyi (Reference Polanyi1968) – have argued that focusing on the relationship between constraints and dynamics is essential if we are to understand the fundamental logic of living processes. They each stress that these complementary aspects of a living process must at the same time be functionally interdependent but physically independent.
Thus, Polanyi (Reference Polanyi1968) argues that the constraints that organize the dynamical processes of life are “Irreducible higher principles [that] are additional to the laws of physics and chemistry” (p. 160). And similarly, Pattee (Reference Pattee, Coombs and Sulcoski1997) points out that “Physical laws and semiotic controls require disjoint, complementary modes of conceptualization and description. Laws are global and inexorable. Controls are local and conditional. Life originated with semiotic controls” (p. 9).
For Pattee, “semiotic controls” exist in the form of “non-integrable constraints” that involve a “necessary epistemic cut between the coherent physical dynamics and its rate-independent semiotic description.” Importantly, constraints are off-loadable onto artifacts and their structure/form. This facilitates and is necessary for the preservation of constraints across potential changes of dynamics.
Although the Peircean semiotic framework is often treated as though it is a structuralist typology of sign types, it is also compatible with a dynamical framework. This is because Peirce understood the interpretation of signs to be constituted by the production of signs (interpretant production), and because he also understood mental processes to be sign production in this same sense and not the locus of some intrinsically meaningful mental token. In this framework, there is no simple mapping between a sign vehicle (signifier) and what it refers to (signified). Rather, sign vehicles are physical forms that mark phases of a process of forms-modifying-the-production-of-other-forms. There is no final form that marks the terminus of the process. Indeed, for Peirce, what he calls the “final interpretant” is in effect a habit of interpretant generation, that is, a process organization.
To bring this semiotic analysis into alignment with dynamical theories it is necessary to understand signs as sources of constraint on dynamics and, as Pattee and Polanyi independently recognized, to understand that semiotic constraints must necessarily be distinct from the dynamics that they control. Both the code metaphor and the conception of cognition as mere correlated dynamics ignore this distinction. But a dynamical semiotic approach that treats sign vehicles (whether words or constraints on neural activity) as information structures that control the dynamics of the production of other sign vehicles can preserve the concept of representation, which satisfies the requirements specified by Brette (see also Bickhard Reference Bickhard2009), without reducing it to a mapping relationship. It also provides a context for distinguishing modes of semiotic relations and semiotic differentiation processes in terms of different modes of constraint.
This reframing also illuminates the relationships of this debate to debates in information theory on the one hand and linguistic theory on the other.
Information theory, following the pioneering insights of Claude Shannon (Reference Shannon1948), notoriously avoids any effort to deal with representational content or normative properties, such as accuracy and truth. And yet it provides a precise formalization for measuring the information content of a medium or a message within that medium, for optimally encoding a signal, and for compensating for noise (though both signal and noise are normative distinctions). Shannon's measure of information in a message is assessed in terms of the uncertainty that is thereby reduced by virtue of the constraint on its possible entropy. Thus, implicitly, it treats whatever semiotic value can be provided in a message as a function of this constraint.
Linguistics has also struggled with the code metaphor (as reflected in the famous signifier-signified relationship described by Ferdinand de Saussure [1959]), and critics of this conception have likewise attempted to reframe “languaging” in purely correlated dynamical terms. The assumption of a completely unstructured “arbitrary” correspondence between word sounds and meanings has led to a dilemma called the symbol grounding problem by Steven Harnad (Reference Harnad1990b), on one hand, and has motivated theories of innate or culturally imposed grammatical principles and syntactic rules to explain the complex structure of languages, on the other. But a reframing of the linguistic structures in terms of constraints on dynamically grounded communicative processes also resolves these dilemmas (Deacon Reference Deacon, Hayes and Wilson2018; Rączaszek-Leonardi Reference Raczaszek-Leonardi2009; Rączaszek-Leonardi et al. Reference Raczaszek-Leonardi, Nomikou, Rohlfing and Deacon2018).
Acknowledgments
This work was supported by NCN OPUS 2018/29/B/HS1/00884.