This is the text of my talk at Derrida Today in Montreal from May 2018:
What I hope to interrogate with you all, and with Derrida, today is an event. An event, whose specific history and historical development I will have to leave to the side in the interest of time, but that I hope nonetheless you will recognize. This event is one that corresponds with the rise of machine learning, data analytics, the data deluge, “the fourth paradigm” of science as Microsoft has put it, “the end of theory” as Chris Anderson described.
In order to do so I will, like Levi-Strauss’ bricoleurs, need to in due course invent a few engineers; so I offer, in advance, my apologizes to those implicated. But they all seem to now have start-ups or to be Associate Deans for research. So, I think they will be fine, but still my apologies.
Just like with structuralism the question of the event will be central and de-centered to the thought of this event. Derrida begins Sign, Structure and Play by calling the rise of structuralism an event, but warning that the very concept of event is one that structural thinking seeks to “reduce or suspect.” The event appears as a threat to structurality, especially if we think of the event in its quasi-messianicity that it takes on in Derrida’s later work, because the event, that which arrives unpredictably from elsewhere, is effective in so much as it disrupts, rearranges and reconfigures structure.
I would like to suggest that an especially productive term for this impulse or this event would be neo-structuralism. With this term I aim both to argue that these discourses take part in a certain historical amnesia of structuralism and cybernetics and yet that something has clearly taken place in the history of thought, corresponding with the increasing reliance for social and scientific knowledge production on algorithms trained on relatively large datasets. Ultimately what I would like to do, in a possibly round about way, is to articulate what precisely constitutes this neo-structuralist event, how it relates to and differs from classic structuralism and argue for the applicability of the Derridean critique of structuralism to this new form, especially as it is articulated in Sign, Structure and Play in the Discourse of the Human Sciences.
Sign, Structure and Play is an incredibly dense text with a number of elements at stake. For the sake of orientation, and with apologies to the depth and nuance of the text, let me offer a very sparse and overly simple summary of what interests me most vis-à-vis these questions: Derrida argues that the history of structure has been the history of the replacement of transcendental centers: God, man, etc. Structuralism is unique in that it de-centers structure insisting on its free play, but Derrida tells us that its attempt to erase the center always get caught back in the game and end up under the sway of a now absent center.
But to turn to our present moment, and offer an explanation of neo-structuralism, I believe there is no surer guide than the clearly hyperbolic and likely wholly and technically incorrect words of Chris Anderson. But, for all that he gets wrong, I think Anderson in his 2008 piece in Wired Magazine entitled “The End of Theory” (as Derrideans we have perhaps heard those words or that threat too many times to count) most fully encapsulates and enunciates the ultimate desire or even drive of machine learning and data analysis. He says:
Out with every theory of human behavior, from linguistics to sociology. Forget taxonomy, ontology, and psychology. Who knows why people do what they do? The point is they do it, and we can track and measure it with unprecedented fidelity. With enough data, the numbers speak for themselves. The big target here isn’t advertising, though. It’s science…faced with massive data, this approach to science — hypothesize, model, test — is becoming obsolete…There is now a better way. Petabytes allow us to say: “Correlation is enough.” We can stop looking for models.
In sum, Anderson’s dream is one in which we throw enough data into a computer running some machine learning algorithm and it offers us, instead of causal theories, predictions with associated probabilities for individual questions. It no longer matters how the economy actually functions if we can predict what it will do tomorrow. The same with medicine, physics, public policy, perhaps even literature, etc.
In this way, many of these systems go beyond structuralism in effacing any sort of a center or even a ground. We can take the example of Google’s Page Rank algorithm, whose specifics we do not entirely know, but whose basic operation was published as an academic paper by Larry Page. One computes the probability that a random web ‘surfer’ clicking on random links and visiting random pages would end up at any given page. Then this probability becomes a measure of its importance. In essence, what we end up with is a definition of importance that is wholly recursive. What makes a page important is that many other important pages point to it, for that is precisely how our ‘surfer’ will end up on a given page. In lieu of some deep theory of structure, we get simulations than run over and over until they converge or stabilize on some passable solution to the problem at hand, always deferring what we mean to some other moment.
Instead of categories defining our relationship to the world these approaches attempt to operationalize momentary correlations, what my colleague John Cheney-Lippold calls “measurable types” (e.g. ‘importance’ in the recursive way PageRank defines it). Geoffrey Bowker puts the condition quiet poetically stating: “Thus, in molecular biology, most scientists do not believe in the categories of ethnicity (Reardon, 2001)—and are content to assign genetic clusters to diseases without passing through ethnicity…recommender systems work through correlation of purchases without passing through the vapid categories of the marketers—you don’t need to know whether someone is male or female, queer or straight, you just need to know his or her patterns of purchases and find similar clusters.”
But, Bowker dealing rather succinctly with both Anderson and Bruno Latour, argues that there are two main problems with this approach. The first is that categories still matter. These categories still define our identities. We can cite Judith Butler’s work on performance here to quickly note that we all still in some capacity perform some concept of gender that we identify with—and these algorithms by showing us and guiding us relative to our previous patterns push us even further into performing that which we are already performing. Moreover, we require the categories of race and sex in order to be able to point to the existence of racism and sexism. The occlusion of these categories offered up by the high priests of machine learning result in the situation outlined so clearly in Safiya Nobel’s new book Algorithms of Oppression or Virginia Eubanks’ Automating Inequality. These algorithms can very easily reproduce the categories they claim to efface.
Second, and still with Bowker, the archive from which this data is drawn is always selective. He cites Archive Fever on this point to argue that the data of big data is never the world, but rather a very small selection of it. To compute the world in whole would require the entirety of the world; and hence every computable dataset is an archive, a finite selection of attributes and data points. Thus, this selection always requires a model or a theory—those who claim to not have one merely have bad and implicit theories. As Bowker and others have put it: “there is no raw data.” Herein lies the fundamental contradiction of neo-structuralism, while it claims to do away with models and theories, it always, perhaps at the moment if offers up a TED-talk but also at the moment it commences gathering data, falls back into theories.
In turning this work immediately into action–“who cares why these things correlate, but let us act as if they do”–neo-structuralism tries to go even one step further than structuralism in neutralizing the event that would disrupt it. Algorithms are often designed to privilege more recent data over older—even for the sake of the size of the archive deleting old data—but what this ultimately means in terms of actual function is that an event can, at least theoretically, be absorbed, be managed. To offer a personal, if banal example, from the realm of online shopping, one could suddenly lose their interest in ultra marathons and develop a new found appreciation for the nuances of diaper sizing, brands and types.
Machine learning and algorithmicity are constantly being fine-tuned to strike an exceptional balance between memory and event; to be able to act “in real time” without giving in to the Humeian concern that perhaps it is equally likely the sun will rise or not rise tomorrow. But, as these forms of knowledge production attempt to capture the event two things happen. First, they are confronted with the possibility of an event that would be more than internal to their operation (such as a change of tastes or an outbreak of a new disease), but rather an event that overflows the algorithm; an event perhaps of the kind that the financial crisis of 2008 represents; one that would break the frame of the algorithm and even if that particular event was reabsorbed into global capitalism another one may not.
Second, and most importantly for our purposes there is a tendency for something else to get caught in the capture of the event. Derrida says of structuralism:
This disruption was repetition in all of the senses of this word. From then on it became necessary to think the law which governed, as it were, the desire for the center…The absence of the transcendental signified extends the domain and the interplay of signification ad infinitum.
We find ourselves, just as under structuralism, in want of categories, in want of a center that would hold the freeplay of these algorithms together. Derrida says this of those structures that used to have a center, God, man, the empire, etc: “coherence in contradiction expresses the force of a desire.” This desire is precisely what structuralism and neo-structuralism repeat. In doing away with this center it finds itself governed by its logic as an absence, all the more powerful for having been repudiated.
Nowhere is this more clear than in a recent project by Matt Jockers, wherein he used sentiment analysis to detect plot structures in tens of thousands of books; a project that was picked up with some interest by the national media. He concluded, and this is from an interview:
“I did some distance similarity metric calculations and machine clustering to see if I could identify archetypal plot shapes,””The short answer is, yes I did, and there’s six or sometimes seven.” That little ambiguity, Jockers explained, is because the data collecting and sorting technique “involves picking at random from 50,000.” “There’s six about 90 percent of the time,” Jockers said. “Ten percent of the time, the computer says there’s a seventh [plot shape].”
Let me just note as an aside, Annie Swafford did some amazing investigation and experiments with Jockers’ code and demonstrated that most of what he was detecting was not even in the data he collected but rather what are known as ringing effects that result from the low pass filters he was using to attempt to smooth these structures. I’m happy to explain exactly what that means in the Q&A or after to anyone who is curious, but the important point is that the categories he thought he saw were largely artifacts, not of his archive, but of his tools. Moreover, we can see here precisely Bowker and Derrida’s point from above: this is an amazing reduction and occlusion of the archive—all of these novels, which are only English novels, are in essence reduced to whether words are happy or sad words.
But, what is most telling is the force of this desire that would make of the ability to perform real-time analysis a set of structural-archetypal statements; namely that there are in the end six (well maybe seven) plot structures; and we should note this probabilistic remainder of the seventh category is telling of the contradiction of these methodologies. In neutralizing the event in real-time neo-structuralism offers up the illusion that the event has been mastered beyond time, beyond history and so despite Anderson’s promise of an end of theory, we are tempted, just this once, since the numbers look so good to offer a theory that would be beyond event. And so we are given the six (well maybe seven) plot structures that explain all literature. While these structures may appear antithetical to the aims of neo-structuralism, its real-time mobile and measurable types, we see here coherence in contradiction expressing the force of a desire—for a center, for stability, for a very old mode of reading and categorizing literature. Just as old centered structures used to, it, offers up, as Derrida says, the “certitude anxiety can be mastered.”
One more quick example of this event in a slightly different field: In 1988, Strack, Martin, and Stepper published the results from a study that found that when participants moved their mouths into the shape of a smile or frown (by holding a pen between their teeth or lips) it affected their emotional response. This ‘facial feedback hypothesis’ has since become a well accepted idea in modern experimental psychology. But, a large scale replication study failed to find any significant effect.
While the replication study was being carried out at multiple independent labs, Strack co-authored an article with Stroebe, a social psychologist, arguing that replication studies do not necessarily call into question the theories they claim to debunk. They argue “There seem to be no reasons to panic the field into another crisis. Crises in psychology are not caused by methodological flaws but by the way people talk about them (67).” They argue that these replication studies do not replicate the initial conditions—e.g. the moment, culture, place, etc. I do not think it was their intent, but their argument if taken in its entirety seems to suggest that ‘studies’ that explore effects on multiple subjects do not tell us anything about human psychology. Rather, what constitutes the human psyche is an infinite variation of individual semantic and historical linkages, in essence the psychoanalytic subject. They never admit as much, but the path they lay out away from any replication crisis leads straight back to psychoanalysis and the importance of the case over aggregate experiments.
They are caught in the same trap as Levi-Strauss, as Derrida outlines: “The risk I am speaking of is always assumed by Lévi-Strauss and it is the very price of his endeavor. I have said that empiricism is the matrix of all the faults menacing a discourse which continues, as with Lévi-Strauss in particular, to elect to be scientific.” Their commitment to maintain some generality to their theory—under conditions electing to be scientific—mean that the empirical result of a replication test always threatens this work.
While the scene is wholly different and their work is not machine learning per say, they attempt to repeat a variation of Jocker’s move and, I would argue, operationalize the event of neo-structuralism for their defense. They use the infinite variability of the individual data point, the algorithm’s receptivity to an event, to neutralize the event; to say in the end: “well there is really no event, the general theory holds because the algorithm can always update itself.” There is a double movement at play here: on the one hand Anderson’s dream is accepted as reality but in order to reassert the strength of a theory (a theory which amounts to in Derrida’s words “philosophizing badly”) to have done away with the event rather than to have integrated it. There are, again to quote Derrida, “many ways of being caught in this circle.”
Neo-structuralism then presents itself as a variation of structuralism, a repetition of this earlier repetition. We can hear it clearly in what Derrida says of Levi-Strauss:
If Lévi-Strauss, better than any other, has brought to light the freeplay of repetition and the repetition of freeplay, one no less perceives in his work a sort of ethic of presence, an ethic of nostalgia for origins…As a turning toward the presence, lost or impossible, of the absent origin, this structuralist thematic of broken immediateness is thus the sad, negative, nostalgic, guilty, Rousseauist facet of the thinking of freeplay of which the Nietzschean affirmation—the joyous affirmation of the freeplay of the world and without truth, without origin, offered to an active interpretation—would be the other side…There are thus two interpretations of interpretation, of structure, of sign, of freeplay. The one seeks to decipher, dreams of deciphering, a truth or an origin which is free from freeplay and from the order of the sign, and lives like an exile the necessity of interpretation. The other, which is no longer turned toward the origin, affirms freeplay and tries to pass beyond man and humanism, the name man being the name of that being who, throughout the history of metaphysics or of onto-theology—in other words, through the history of all of his history—has dreamed of full presence, the reassuring foundation, the origin and the end of the game.
It is here, I think, that we can fully differentiate neo-structuralism from structuralism but also see its return into structuralism. What we witness is its desire for a future moment of coherence, of convergence over a certain time bound data set, all the while admitting the impossibility of origin. Everything is chaos and noise, but we will seek out some small correlation that offers some advantage to our action. Thus, if structuralism always chose the former path, and failed to do away with the origin, the arche, neo-structuralism repeats this, offering instead of an origin a telos, a future of managed free-play. We do not know what created the complexity of the world, but the neo-structuralist claim is that with enough data and the right algorithms we can manage and predict the unpredictable. But, right at the beginning of sign, structure and play, we are told: “From the basis of what we therefore call the center and which, because it can be either inside or outside, is as readily called the origin as the end, as readily arché as telos.” While neo-structuralism may substitute a telos for an arche, in the end it will have amounted to nearly the same thing.
So, perhaps we should prefer a different course, rather than repeating structuralism to instead choose the Nietzschean affirmation, to try to pass beyond man and humanism. But, as Derrida reminds us, we are caught in this game, these systems and structures compute and think us. It is far from clear that a choice is at our disposable. Derrida again:
For my part, although these two interpretations must acknowledge and accentuate their différence and define their irreducibility, I do not believe that today there is any question of choosing…Here there is a sort of question, call it historical, of which we are only glimpsing today the conception, the formation, the gestation, the labor. I employ these words, I admit, with a glance toward the business of childbearing-but also with a glance toward those who, in a company from which I do not exclude myself, turn their eyes away in the face of the as yet unnameable which is proclaiming itself and which can do so, as is necessary whenever a birth is in the offing, only under the species of the non-species, in the formless, mute, infant, and terrifying form of monstrosity.
If this is our task, we have only just begun or perhaps not even begun. In a perhaps slightly different scene and with slightly different technologies, we repeat a dream that we have dreamt before. I’m not sure that our course will be different or that it will necessarily be the same. But, my hope is in the very least to demonstrate the value of Derrida’s thought and writing to this moment and beyond that to suggest that in seeing how we are confronted not with an event that is wholly new, but rather an event that is a form of repetition we may be better prepared to await the arrival of this future, to engage the labor of its thought.