The term coding along with processing and computation forms a trio of concepts whose ubiquitous use in neuroscience is underwritten by a usage so loose that often they can replace one another without appreciably affecting the sense of what is being conveyed, as in “We study how the brain codes (or processes, or computes) emotions.” As placeholders for “something unknown or yet to be determined,” these terms might seem innocuous, were it not for the fact that they carry with them semantic baggage from their nonneural provenance. This technical or formal sense figures in some, but far from all, of their uses in neuroscience discourse, a situation that promotes conceptual and communicative confusion.
Regarding the technical and formal sense of “code” and “coding” specifically, the core of the construct as used in the disciplines from which it was imported is that of rule-governed relations of correspondence between two domains with arbitrary correspondence assignments in the sense that alternative assignments would work. This is the only common denominator of the two principal codes to be found in nature: (1) the nucleotide-triplet code by which transfer RNA uses the base sequence of messenger RNA to string amino acids into proteins (other genetic operations being template-based), and (2) the two-level combinatorics of phonological and lexical elements by which the sound strings of human languages code the arbitrary (in this case conventional) mapping between the form and meaning of utterances. The arbitrary/conventional element is conspicuous also in the various artificial codes created in reliance on human language: ciphers in cryptography, Morse code in telegraphy, and the many coding schemes employed in the design of digital computers and their programs.
Bona fide codes represent a vanishingly small portion of the myriad lawful relationships that make up the natural world. Likewise for the nervous system: All neural operations of which we have actual knowledge are lawful ones lacking the arbitrary aspect of the correspondence rules of a coding scheme. There is no dearth of appropriate terms for such noncoding lawful relations: “function,” “transfer function,” “transduction,” “mapping,” “transform,” “representation,” and more. Yet the technically incorrect “coding” often substitutes for more informative terms, especially when incompletely known aspects of function, and issues of significance or meaning in particular, are being addressed.
The tacit analogy appears to be the message passing made possible by human language: Neurons “communicate,” and somehow the nervous system generates meaning, so perhaps neurons send language-like messages to one another, coding significance or meaning in the temporal sequence of their spike trains. This would supply neurons with an extra-local “code for information” or common language.
But consider the roughly 8000 synapses impinging on a single cortical pyramidal cell: They originate in many hundreds, if not thousands, of other pyramidal cells and subcortical neurons. The blending at the axon initial segment of the graded potentials induced by the synaptic activity of all these sources, excitatory and inhibitory, jointly determines whether that cell will reach threshold to release its all-or-none action potential to its audience of hundreds or thousands of other cells. “Messages” do not survive such treatment anymore than it is possible to monitor the conversations of a cocktail party by a single decibel meter rigged to issue a spike at a fixed sound pressure level. The fact that close to threshold a situation can arise in which the spike output of a neuron replicates the firing of one or a few afferents does not change the basic irrecoverability of the pattern of afferent input from a neuron's output.
Regarding significance, more than half a century of arduous mapping of the response properties of single neurons throughout the nervous system tells us that the principal determinant of what the activity of a neuron signifies is where it is located. This “where” ranges from the gross subdivisions of the brain down to a cell's precise connective position in the synaptic network it inhabits (Passingham et al. Reference Passingham, Stephan and Kötter2002). Moving a microelectrode to a neighboring cell typically discloses slightly different, but related, response properties, in the aggregate generating the familiar functional maps that abound at every level of the neuraxis.
The reason for this parallel and analog mode of representing significance in the brain is not far to seek. The cortical signal propagation speed of 1 to 6 meters per second (based on Pascual-Leone et al. Reference Pascual-Leone, Walsh and Rothwell2000; Schmolesky et al. Reference Schmolesky, Wang, Hanes, Thompson, Leutgeb, Schall and Leventhal1998) is some 180 million to 30 million times slower than that of electronic circuitry. Operations in this sluggish medium must deliver their global results some three to four times per second (the frequency of gaze movements, the leading edge of most behavior; Merker Reference Merker2013b). The brain is therefore always strapped for time, and as computer programmers know, when time is short, you don't compute, you use lookup tables.
To perform sophisticated functions with sluggish components in real time, the brain arrays them in complex concatenations of innate as well as acquired (learned) maps (“lookup tables” for “priors”) interfaced with one another via a variety of connectivity-based functional transformations. Together they form an analog “computer,” not a digital one, for which the analog inner workings of so-called neural networks supply toy models. In analog computers there is no program running on the hardware: The hardware itself is the program, and that hardware, moreover, is modifiable by its own activity, generating the learned content of maps.
With no program to run, and no messages to send, there is nothing to code, because significance is not represented in propositional or symbolic form in the brain, but positionally: where activity flares is what it signifies or means. But what about the signal discretization of the action potential? A neuron's work takes place by analog blending of graded potentials on its soma-dendrite membrane, and spikes serve only to transmit with fidelity a running record of the upshot of that work to other locations, in keeping with Shannon's insight regarding the utility of discretization for faithful signal transmission. No neural work is being performed by discrete spikes except that of bridging distance.
In sum, the answer to the question posed in Romain Brette's title is a resounding “No!”
You have
Access
The term coding along with processing and computation forms a trio of concepts whose ubiquitous use in neuroscience is underwritten by a usage so loose that often they can replace one another without appreciably affecting the sense of what is being conveyed, as in “We study how the brain codes (or processes, or computes) emotions.” As placeholders for “something unknown or yet to be determined,” these terms might seem innocuous, were it not for the fact that they carry with them semantic baggage from their nonneural provenance. This technical or formal sense figures in some, but far from all, of their uses in neuroscience discourse, a situation that promotes conceptual and communicative confusion.
Regarding the technical and formal sense of “code” and “coding” specifically, the core of the construct as used in the disciplines from which it was imported is that of rule-governed relations of correspondence between two domains with arbitrary correspondence assignments in the sense that alternative assignments would work. This is the only common denominator of the two principal codes to be found in nature: (1) the nucleotide-triplet code by which transfer RNA uses the base sequence of messenger RNA to string amino acids into proteins (other genetic operations being template-based), and (2) the two-level combinatorics of phonological and lexical elements by which the sound strings of human languages code the arbitrary (in this case conventional) mapping between the form and meaning of utterances. The arbitrary/conventional element is conspicuous also in the various artificial codes created in reliance on human language: ciphers in cryptography, Morse code in telegraphy, and the many coding schemes employed in the design of digital computers and their programs.
Bona fide codes represent a vanishingly small portion of the myriad lawful relationships that make up the natural world. Likewise for the nervous system: All neural operations of which we have actual knowledge are lawful ones lacking the arbitrary aspect of the correspondence rules of a coding scheme. There is no dearth of appropriate terms for such noncoding lawful relations: “function,” “transfer function,” “transduction,” “mapping,” “transform,” “representation,” and more. Yet the technically incorrect “coding” often substitutes for more informative terms, especially when incompletely known aspects of function, and issues of significance or meaning in particular, are being addressed.
The tacit analogy appears to be the message passing made possible by human language: Neurons “communicate,” and somehow the nervous system generates meaning, so perhaps neurons send language-like messages to one another, coding significance or meaning in the temporal sequence of their spike trains. This would supply neurons with an extra-local “code for information” or common language.
But consider the roughly 8000 synapses impinging on a single cortical pyramidal cell: They originate in many hundreds, if not thousands, of other pyramidal cells and subcortical neurons. The blending at the axon initial segment of the graded potentials induced by the synaptic activity of all these sources, excitatory and inhibitory, jointly determines whether that cell will reach threshold to release its all-or-none action potential to its audience of hundreds or thousands of other cells. “Messages” do not survive such treatment anymore than it is possible to monitor the conversations of a cocktail party by a single decibel meter rigged to issue a spike at a fixed sound pressure level. The fact that close to threshold a situation can arise in which the spike output of a neuron replicates the firing of one or a few afferents does not change the basic irrecoverability of the pattern of afferent input from a neuron's output.
Regarding significance, more than half a century of arduous mapping of the response properties of single neurons throughout the nervous system tells us that the principal determinant of what the activity of a neuron signifies is where it is located. This “where” ranges from the gross subdivisions of the brain down to a cell's precise connective position in the synaptic network it inhabits (Passingham et al. Reference Passingham, Stephan and Kötter2002). Moving a microelectrode to a neighboring cell typically discloses slightly different, but related, response properties, in the aggregate generating the familiar functional maps that abound at every level of the neuraxis.
The reason for this parallel and analog mode of representing significance in the brain is not far to seek. The cortical signal propagation speed of 1 to 6 meters per second (based on Pascual-Leone et al. Reference Pascual-Leone, Walsh and Rothwell2000; Schmolesky et al. Reference Schmolesky, Wang, Hanes, Thompson, Leutgeb, Schall and Leventhal1998) is some 180 million to 30 million times slower than that of electronic circuitry. Operations in this sluggish medium must deliver their global results some three to four times per second (the frequency of gaze movements, the leading edge of most behavior; Merker Reference Merker2013b). The brain is therefore always strapped for time, and as computer programmers know, when time is short, you don't compute, you use lookup tables.
To perform sophisticated functions with sluggish components in real time, the brain arrays them in complex concatenations of innate as well as acquired (learned) maps (“lookup tables” for “priors”) interfaced with one another via a variety of connectivity-based functional transformations. Together they form an analog “computer,” not a digital one, for which the analog inner workings of so-called neural networks supply toy models. In analog computers there is no program running on the hardware: The hardware itself is the program, and that hardware, moreover, is modifiable by its own activity, generating the learned content of maps.
With no program to run, and no messages to send, there is nothing to code, because significance is not represented in propositional or symbolic form in the brain, but positionally: where activity flares is what it signifies or means. But what about the signal discretization of the action potential? A neuron's work takes place by analog blending of graded potentials on its soma-dendrite membrane, and spikes serve only to transmit with fidelity a running record of the upshot of that work to other locations, in keeping with Shannon's insight regarding the utility of discretization for faithful signal transmission. No neural work is being performed by discrete spikes except that of bridging distance.
In sum, the answer to the question posed in Romain Brette's title is a resounding “No!”