#3: Myth and Meaning

Myth and Meaning

The following are preparatory notes for the first session of the structuralism reading group, which focuses on texts by Claude Lévi-Strauss.

Brief note on definitions of AI: For now, I am going to take a leaf out of Ellen Broad’s book, Made by Humans (2018), and say there is ‘no agreed definition of artificial intelligence. AI covers lots of things: machine learning, drones, robotics, virtual reality and big data’ (p.16); and indeed echo her further comment: ‘When I say AI in the book, I’m usually talking about machine learning, specifically with respect to automated decision-making. Sometimes I switch between ‘automated system’ and ‘machine learning’ and ‘AI’, which can often be used interchangeably. ‘Automated system’ is most accurate to me, but ‘AI’ is cool and zeitgeisty and I want you to keep reading’ (p.16). She also gives various definitions in her recent talk on responsible AI design (for the RLUK’s Digital Shift Forum), which include AI as a ‘constellation of technologies’ (AI Now Institute); as processes or tasks undertaken by computers that previously only humans could undertake; or ‘whatever a computer can’t do yet, because when it can we give it a different name; we give it a name like search engine or chatbot, and it is no longer as mysterious sounding’.

Constituent Units
In the opening pages of his essay, ’The Structural Study of Myth’ (originally published in Journal of American Folklore, in 1955), Lévi-Strauss offers a précis of the problem (or lack of attention) to the study of myth, alongside a concise account of structural linguistics, from which he develops his method.

Myths, he suggests, ‘are still widely interpreted in conflicting ways’, and generally always reduced ‘either to idle play or to a crude kind of philosophic speculation’ (p.207). The problem he notes, is that ‘anything is likely to happen’ in myth, there is ‘no logic, no continuity’:

With myth, everything becomes possible. But on the other hand, this apparent arbitrariness is belied by the astounding similarity between myths collected in widely different regions. Therefore the problem: if the content of a myth is contingent, how are we going to explain the fact that myths throughout the world are so similar?
Lévi-Strauss, ’The Structural Study of Myth’, p.208

Having set out the problem, the very next paragraph offers a neat explanation of structural linguistics, noting the key point that the same ‘sounds’ used by numerous languages can convey entirely different meanings. A contradiction that is ‘surmounted only by the discovery that it is the combination of sounds, not the sounds themselves, which provides the significant data’ (p.208). It is worth noting the use of the word ‘data’ here, which crops up numerous times, and for the purposes of the reading group marks a bridge with contemporary questions of data (and arguably needs holding onto contra to the more fashionable post-structuralist critique). As Lévi-Strauss puts it: ‘Whatever emendations the original formulation may now call for, everybody will agree that the Saussurean principle of the arbitrary character of linguistic signs was a prerequisite for the accession of linguistics to the scientific level’ (p.209).

As will be noted below, there is a parallel to be drawn here with how more recent advances in AI programming really took off once a shift was made to the statistical, probability modelling of language (away from expert, rule based systems). This maps to Lévi-Strauss’ characterisation of Saussure’s distinction between langue and parole, ‘one being the structural side of language, the other the statistical aspect of it, langue belonging to a reversible time, parole being non-reversible’ (p.209). However, for Lévi-Strauss there is an important additional concern, since myths are told through these temporalities of language. He asserts a third referent of myth: ‘On the one hand, a myth always refers to events alleged to have taken place long ago. But what gives the myth an operational value is that the specific pattern described is timeless; it explains the present and the past as well as the future.’ (p.209). By way of explanation he contrasts the historian and the politician. The historian, for example, will refer to the French Revolution as a sequence of past, non-reversible events, while the politician will suggest of ‘both a sequence belonging to the past … and a timeless pattern which can be detected in the contemporary French social structure and which provides a clue for its interpretation, a lead from which to infer future developments’ (p.209). In this sense, ‘[m]yth is language, functioning on an especially high-level where meaning succeeds practically at “taking-off” from the linguistic ground on which it keeps on rolling’ (p.210).

Myth is built upon language (or, we might say, to follow Barthes’ later formulation in Mythologies, myth is of a second order system). As Lévi-Strauss argues, myth is like language in being ‘made up of constituent units’, but differs in complexity; he refers to myth as ‘gross constituent units’ (p.210-211). And, having defined the ‘unit’ of analysis, he offers a pithy, working definition of structural analysis:

…the principles which serve as a basis for any kind of structural analysis: economy of explanation; unity of solution; and ability to reconstruct the whole from a fragment, as well as later stages from previous ones.
The Structural Study of Myth’, p.211

Perhaps this same definition offers something towards a definition for AI programming? The ‘economy’ suggestive of the need to function compactly (otherwise any interface would elide its usefulness); the need for ‘unity’ becomes significant for AI to make ‘sense’ (its datasets need to be ‘complete’ in some way); and the reconstruction again takes us to the statistical modelling of AI. However, Lévi-Strauss’ characterisation of myth as operating multifariously through time (through the synchronic and diachronic) suggests potentially of a complexity that to date alludes AI modelling. (Later, for the Jakobson reading, I will reference to a presentation from Elon Musk and his various engineers, where they are asked if a self-driving car would notice if a trompe-l’œil tunnel were painted on a brick wall. ‘All in good time’ is perhaps an apt summary of the answer given).

Structuralism: The Search for Invariance
In 1977, Lévi-Strauss presented five lectures for the Massey Lecture series (broadcast on Canada’s CBC radio programme, Ideas) ^{Listen to recordings →}. The lectures are available online (and are definitely worthwhile listening, not least to hear Lévi-Strauss’ very convivial style of speaking). They provide a lucid and engaging summary and late reflection by Lévi-Strauss himself on his work and approach.

The opening lecture (‘The meeting of Myth and Science’) opens with a position that largely echoes Michel Foucault’s account in The Order of Things, regards a shift of episteme (ways of thinking about truth and discourse, as underpinning fields of knowlege) from Renaissance to Classical, to Modern thought, which includes the shift to the Enlightenment, and crucially to the compartmentalising of knowledge and inquiry. So, for example, in Foucault’s account, the episteme of the Renaissance is characterised through resemblance, whereas for the Classical era, there is a turn to matters of representation, ordering, differences, classification and taxonomy. With this in mind: a few key points to telegraph from the opening of Lévi-Strauss’ lectures:

Lévi-Strauss is a keen reader of scientific literatures (‘eager to be informed as possible of everything that takes place in modern science and its new developments). As we know, similar to others in his generation, there is a specific interest in cybernetics, but more than that (and/or along with that interest) he makes frequent reference to physics and biology.

There is an argument that modern science has ‘lost’ some things (due to classification and dissection), but that there are efforts to regain them. E.g. ‘the world of smells … accustomed to think that this was entirely subjective …Now the chemists are able to tell us that each smell or each taste has a certain chemical composition…’

In introducing the term ‘structure’, reference is given to the physical structure of the retina and neurological pathway: ‘some cells are sensitive only to straight direction, in the vertical sense, others in the horizontal’, so, while accepting he is making a simplification, Lévi-Strauss nonetheless makes the point that the ‘whole problem of experience versus mind seems to have a solution in the structure of the nervous system, not in the structure of the mind or in experience, but somewhere between mind and experience…’

Drawing upon a childhood memory of ‘reading’ signs (before language acquisition), the structuralist approach is explained as ‘the quest for the invariant, or for the invariant elements among superficial differences’.

The ‘invariant’ offers a connection to AI computation principles, and so opens perhaps a question about a potential relationship between structural analysis and contemporary AI methods. The following passage from Kate Crawford’s The Atlas of AI (2021), is pertinent:

In the 1970s, artificial intelligence researchers were mainly exploring what’s called an expert systems approach: rules-based programming that aims to reduce the field of possible actions by articulating forms of logical reasoning. But it quickly became evident that this approach was fragile and impractical in real-world settings, where a rule set was rarely able to handle uncertainty and complexity. New approaches were needed. By the mid-1980s, research labs were turning toward probabilistic or brute force approaches. In short, they were using lots of computing cycles to calculate as many options as possible to find the optimal result.
One significant example was the speech recognition group at IBM Research. The problem of speech recognition had primarily been dealt with using linguistic methods, but then information theorists Fred Jelinek and Lalit Bahl formed a new group, which included Peter Brown and Robert Mercer (long before Mercer became a billionaire, associated with funding Cambridge Analytica, Breitbart News, and Donald Trump’s 2016 presidential campaign). They tried something different. Their techniques ultimately produced precursors for the speech recognition systems underlying Siri and Dragon Dictate, as well as machine translation systems like Google Translate and Microsoft Translator.
They started using statistical methods that focused more on how often words appeared in relation to one another, rather than trying to teach computers a rules-based approach using grammatical principles or linguistic features. Making this statistical approach work required an enormous amount of real speech and text data, or training data. The result, as media scholar Xiaochang Li writes, was that it required “a radical reduction of speech to merely data, which could be modeled and interpreted in the absence of linguistic knowledge or understanding. Speech as such ceased to matter.”
Kate Crawford, The Atlas of AI, p.79

Arguably, there is a parallel here in terms of a relational, statistical approach (with structural linguistics also shifting from a philological account of language to a structuralist account, centring on the arbitrary nature of the sign and its relational effects). And for Lévi-Strauss it is not just the isolated relations of specific units of meaning, but ‘bundles of such relations’, as he outlines in his approach to large ‘data’ sets of myths (initially accessed at the New York State Library, through the introduction of Franz Boas during the 1940s; and where he was also a founding member of the École Libre des Hautes Études, along with Roman Jakobson). This account of the relational (and bundling) is set out in the session’s reading, ‘The Structural Study of Myth’, in Structural Anthropology (pp. 206-231). Here Lévi-Strauss offers an analogy (and arguably more than an analogy) to musical scores, whereby we need to understand melody, harmony and rhythm by looking across the stave, but also up and down the score. (Structuralism and music are the subject of his final 1977 Massey Lecture.). Lévi-Strauss ‘arranges’ narrative components in the form of a score, which breaks from the linear structure of storytelling to a relational, and even ‘sedimentary’ one. He draws analogy with the seams of rock in geology and archeology (which connects to Foucault’s ‘archeology of knowledge’, and also, for Lévi-Strauss, relates to the recovery of latent meaning, akin to the analysis of dreams in psychoanalysis).

While we would not usually refer to language as probabilistic, since we become so familiar with it, when we observe a young child or those learning a new language, there are frequent attempts to test out possible (and ideally probable) uses. In Albert Doja’s review of Lévi-Strauss at his centenary (for Theory, Culture & Society, 2008, Vol 25, No. 7-8, pp. 321–340), reference is made to the ‘algebraic mind’, whereby on a theoretical level, ‘analyses of computational and artificial intelligence models suggest that the brain inherently must function algebraically’, to which Doja further remarks ‘that the findings derived from empirical research using a binary computational model can be seen as supporting Lévi-Strauss’s algebraic model of mind’ (pp.329-332). However, any suggestion of a ‘brute force’ approach (to simply try something until it works!), is perhaps less fitting (or maybe takes place at neurologically deep levels where it is too fast for us to distinguish?). Importantly, we begin to reach toward a distinction, but that, if a connection between structuralism and AI programming is not too far fetched, is also where we might find renewed value in thinking about structuralism again. At stake, I want to suggest, is the hermeneutics of structuralism that might inform (reform?) the blunt, statistical approach of current AI.

Structuralism and the Subject
The case of natural language processing and speech recognition is pertinent. In certain ways we can say the statistical approach has proven to work. Not all the time, but sufficient at least to now market all number of virtual assistants and chatbots. While using vastly more ‘data’ (reducing speech to mere data, as Crawford puts it), there is a corollary with Lévi-Strauss diligently reducing, or finding the invariance in thousands of myths from around the world. The difference however is in the ability to find meaning (or at least to be in search of meaning). The point, then, is that there is another step or level of engagement in structural analysis that goes beyond statistical handling. A case can be made with Lévi-Strauss’s explanation of a myth of the skate (or stingray) fish, which is recounted in his second Massey Lecture (‘“Primitive” Thinking and the “Civilised” Mind’). The myth (from western Canada) tells of the skate trying to master the South Wind, which is especially harsh and which needs to be overcome to ensure the ability to hunt and ultimately to survive . Without going into detail, the story tells of how a group of animals, including the skate capture the wind, which then promises, once set free, not to blow all of the time. The role of the skate in the story is important due to its shape or dimensions. If seen from above it is a large, flat shape. If seen from the side it is hardly visible (and from this point of view almost impossible to hunt). The skate comes to resemble the point of the fluctuations of the wind, to remind that while it is not possible to hunt on many occasions, there still remain opportunities (just as on occasion we might see the skate fish from above, when it is a large, flat target). For the current purposes, the specifics of the story are not so important. Instead, it is Lévi-Strauss’ structuralist reading, and cybernetic account of the fish that is worth keeping in mind. He notes:

…it is an animal which, considered from either one point of view or the other, is capable of giving—let’s say in terms of cybernetics—only a ‘yes’ or ‘no’ answer. It is capable of two states which are discontinuous, and one is positive, and one is negative. The use the skate is put to in the myth—though, of course, I would not like to strain the simile too far—like the elements in modern computers which can be used to solve very difficult problems by adding a series of ‘yes’ or ‘no’ answers.
Myth and Meaning, p.17

What I am interested to explore here is how Lévi-Strauss leads us towards an understanding of the structures of thought that are on the one hand cybernetic, relational, structural, yet also equally meaningful, or at least include the need for us to establish meaning. It is such a step that is in contrast to the brute force of AI processing. And there is an important principle at stake here that arguably ‘to mean’ = order (which poses interesting challenges to how we might understand AI, whether it can find order, or only replicate it). Lévi-Strauss writes:

It is, I think, absolutely impossible to conceive of meaning without order. There is something very curious in semantics, that the word ‘meaning’ is probably, in the whole language, the word the meaning of which is the most difficult to find. What does ‘to mean’ mean? It seems to me that the only answer we can give is that ‘to mean’ means the ability of any kind of data to be translated in a different language. I do not mean a different language like French or German, but different words on a different level. After all, this translation is what a dictionary is expected to give you—the meaning of the word in different words, which on a slightly different level are isomorphic to the word or expression you are trying to understand. Now, what would a translation be without rules? It would be absolutely impossible to understand. Because you cannot replace any word by any other word or any sentence by any other sentence, you have to have rules of translation. To speak of rules and to speak of meaning is to speak of the same thing; and if we look at all the intellectual undertakings of mankind, as far as they have been recorded all over the world, the common denominator is always to introduce some kind of order. If this represents a basic need for order in the human mind and since, after all, the human mind is only part of the universe, the need probably exists because there is some order in the universe and the universe is not a chaos.
Myth and Meaning, p.9

Overall, I am tempted to echo the epistemic shift (or, as I’ll suggest an oscillation of episteme) that Foucault charts in The Order of Things, from resemblance to representation. The more obvious account of AI would be to say it is based upon the latter, representational mode; of the order of classifications, taxonomies and sequencing. Yet, equally, I would argue AI, currently, is based upon resemblance, as it works by emulating ‘ground truth’ data sets. The parallel I want to draw here is with the closed, yet holistic ‘system’ of thought that is found with magic (in Foucault’s account of the pre-modern era) and myth (in Lévi-Strauss’s work). AI systems are frequently hailed to have ‘worked’ when they achieve similitude. The fact that AI is so reliant upon the quality of the available datasets is of course why there are so many critical, ethical concerns raised about the bias and norms of current AI systems and applications (see Kate Crawford’s chapters on ‘Data’ and ‘Classification’, in The Atlas of AI).

So, there is perhaps an oscillation between the resemblance and representational episteme (in Foucault’s terms), with the latter key to how AI processing is contingent upon and performs the overly rationalist, compartmentisation and classification of ‘things’. Kate Crawford, for example, explains how one of the key datasets for training systems on image processing, ImageNet, is premised upon everything being a noun. And, borrowing from George Lakoff, notes how ‘not all nouns are created equal’, whereby, for example:

…the concept of an “apple” is a more nouny noun than the concept of “light,” which in turn is more nouny than a concept such as “health.” Nouns occupy various places on an axis from the concrete to the abstract, from the descriptive to the judgmental. These gradients have been erased in the logic of ImageNet. Everything is flattened out and pinned to a label, like taxidermy butterflies in a display case. While this approach has the aesthetics of objectivity, it is nonetheless a profoundly ideological exercise.
Kate Crawford, The Atlas of AI, p.104-105

A key problem in amongst all of this classification are of course the elements left uncategorised. In semiotic parlance, we might ask of the ‘sub-semiotic’ (those ‘units’ of meaning yet to form), or the supra-semiotics (akin to discourse, the larger constellations of meaning; or myth)? Again, there are questions lurking here about meaning and order, and crucially who gets to say there is meaning and order. On the one hand this evokes questions of power and regulation etc., but also there is an existential point: the ‘who’ in the system, the subject of thought. And part of which, I am interested in Lévi-Strauss’ reference to the ‘two-dimensional time referent’ (synchronic and diachronic) of myth (‘The Structural Study of Myth’, p.212), and the bundling of sets of relations, which perhaps hints at a pertinent model for AI. Here, Crawford’s reference to Umberto Eco is perhaps worth noting:

…classificatory infrastructures contain gaps and contradictions: they necessarily reduce complexity, and they remove significant context, in order to make the world more computable. But they also proliferate in machine learning platforms in what Umberto Eco called “chaotic enumeration.” At a certain level of granularity, like and unlike things become sufficiently commensurate so that their similarities and differences are machine readable, yet in actuality their characteristics are uncontainable.
Kate Crawford, The Atlas of AI, p.110

I’m keen to pick up on Eco’s idea here of ‘chaotic enumeration’, and which will perhaps be pertinent when turning to the Barthes reading in the third session. For now, however, in keeping to the reading of Lévi-Strauss, I wonder, I propose, we might be back at a similar moment as described by Lévi-Strauss in 1977, with the opening of his Massey Lectures, whereby ‘there are some things we have lost’ from ‘mythical’ to scientific thought (which he places between the seventeenth and the eighteenth century, or we might place earlier according to Foucault).

At that time, with Bacon, Descartes, Newton, and the others, it was necessary for science to build itself up against the old generations of mythical and mystical thought, and it was thought that science could only exist by turning its back upon the world of the senses, the world we see, smell, taste, and perceive; the sensory was a delusive world, whereas the real world was a world of mathematical properties which could only be grasped by the intellect and which was entirely at odds with the false testimony of the senses. This was probably a necessary move, for experience shows us that thanks to this separation—this schism if you like—scientific thought was able to constitute itself.
Myth and Meaning, p.4

More so than ever, AI has now been able to constitute itself, but if we consider it marked by the oscillation I refer to (of both resemblance and representation), where does it now need to (reach out/back to) in order to offer a more integrated approach? At the time of speaking, in 1977, Lévi-Strauss had the impression that contemporary science was tending to overcome the earlier schism, particularly in that ‘more and more the sense data [was] being reintegrated into scientific explanation as something which has a meaning, which has a truth, and which can be explained’. In his account, myth contained the data of the senses. How might we locate the missing ‘data’ of current work in AI? My turn to the subject of structuralism has been prompted by this very question.