Let’s fight some more about the digital humanities!

Nan Z Da’s “Computational Case Against Computational Literary Studies” (CLS) in the latest Critical Inquiry has been making the rounds on my social media feed. It’s a thorough and inventive argument and I am impressed by its doggedness, cross-field erudition and commitment to its idea: she re-did studies, chased down data sets, and reconstructed analyses. My long critique below is simply a result of my being impressed enough to care about some subtleties of the argument. Because I have seen disagreements turn into blood sport in literary studies before, let me be crystal clear: nothing I say below should indicate anything other than admiration the author or her work.*

So, what’s my concern?

I think the critique overshoots its mark in claiming that because there are errors in the data science, data science should be greeted with suspicion by literary critics. I am about to publish a co-authored paper on some of the inflated claims regarding machine learning and audio mastering, so I am sympathetic to Da’s skepticism as a general stance, but I’m concerned about how it works out in this case.

For those who haven’t read it, Da’s article proceeds by careful readings of a few CLS texts in order to argue with their modes of statistical interpretation and their relevance for literary criticism. I’m not going to dispute any of the statistical criticisms offered in the essay, because for me the main issue is how humanists should think about computation, quantification, and truth standards. (And I expect that the CLS crowd will offer its own response, and leave it to them to defend themselves.)

I also think Da is asking the right question, which is to be posed to any new movement in scholarship: what does it contribute to the conversation beyond itself? This is especially true if a field claims to displace another. In other words, your burden of proof is higher if you argue that quantification should displace other modes of literary interpretation than if you argue that quantification can be useful along side other modes of literary interpretation. I am firmly in the latter camp.

So, my issues with the piece really come down to two places:

1. What are the standards to which we want to hold humanities work? The warrant behind the main arguments of the piece: the claims of CLS do not stand up to statistical scrutiny, or are artifacts of data mining, or if the results are true, they are banal. The problem is that no humanistic hermeneutic enterprise, apart from maybe some species of philology and bibliography, could actually withstand the burdens of proof implied by Da’s critique. Da’s suggestions for reviewing CLS work at the end of the appendix also suggest a kind of double standard for quantitative and qualitative work in literary studies:

1.) Request Complete Datawork: names of databases, raw texts if non-proprietary, scraping scripts, data tables and matrices, scripts used for statistical analysis and those results. Indication of whether codes are original/proprietary or not. 
2.) Request detailed documentation for replicating the results either on the original or on a similar dataset. Authors should be able to demonstrate the power of their model/test under good faith attempts at replicating similar circumstances. 
3.) Enlist someone to make sure the authors’ scripts run.
4.) Enlist a statistician to a.) check for presence of naturally occurring, “data-mining” results, implementation errors, forward looking bias, scaling errors, Type I/II/III errors, faulty modeling, straw man null hypothesis, etc; b.) see if datawork is actually robust or over-sensitive to authors’ own filters/parameters/culling methods; c.) see if insights/patterns are actually produced by something mechanical/definitional, d.) apply Occam’s Razor Test—-would a simpler method work just as well?
5.) Enlist a non-DH literary critic or literary historian to see if the statistical results actually lead to the broader inferences/claims that matter to literary criticism/history or only do so wishfully/willfully/suggestively/by weak analogy; and to see if, in papers that seek only to perform basic tasks, human classification & human description would not actually be faster (and far more accurate) for both the dataset in question and new ones. 
6.) Apply “smell test”— is the work, minus the computational component, well-written, original, and insightful as a literary-critical or literary-historical argument? Would the argument be published without the computational elements? The “benefit of the doubt for fledgling field” should not apply.

Again, few works of literary criticism could survive this level of scrutiny: imagine every literary historical claim having to please a history-department historian, or every psychoanalytic claim having to satisfy a psychology professor. Imagine having every interpretative claim having to satisfy someone hostile to that mode of interpretation (in fact, these kind of political concerns are already an issue with reviewing, where doxa often prevents good work from surfacing in certain places). Or imagine every work of speculative hermeneutics having to pass muster with a statistical analysis.

I won’t speak for what CLS has given literary study, but at least in my field, quantitative scholarship has produced lots of important work. To give but one example: before the media theory types (like me) got into writing about media and disability, quantitative scholars were already churning out articles documenting a range of issues around access, power, procedure and policy. They identified and laid out a problem that the so-called critical and interpretive people were slow to identify and acknowledge. I have my theories about why that is, but the key thing here is that whatever limitations a “critical” humanist might attribute to quantitative analysis, those epistemologies helped a group of scholars to identify a problem systematically before the self-described “more critical” people did. Each method brings with it biases and limitations but also produces openness to important questions.

I bristle when I hear people ignorant of the humanities or interpretive social science refer to work that doesn’t use numbers as “not empirical.” But given literary criticism’s own fraught histories, well documented and unearthed by its own practitioners (Williams, Said, Sedgwick, etc), I am equally uncomfortable with the opposite bias.

Onto technology.

On page 620, Da argues: “This is not at all to argue that literary analysis must have utility—in fact I believe otherwise—but if we are employing tools whose express purpose is functional rather than metaphorical, then we must use them in accordance with their true functions.” And later: “Text mining is ethically neutral.”

I agree that applying the measure of utility to literary analysis would be a bad thing. But both the other claims seem bizarre to me. First, the humanities routinely adopt and abuse methods from other fields to their ends, from Art Historians’ adaptation of slide projectors and slideware designed for business presentations, to literary critics’ adoption of psychoanalytic theories that are mostly discounted in contemporary psychology. That the use of a data science method is different in literary studies than in its home field is not itself an issue; in fact, I would hope that it is different. The claim about ethical neutrality is contravened by the examples: investing is not ethically neutral, so data mining to invest better is not ethically neutral. Ditto for legal discovery, given how the law works in practice. In fact, I know of no serious scholar of technology who would claim that any technology or technical process is inherently “neutral.”

Da claims CLS “has developed literary metaphors for what coding and statistics actually are and involve, turning elementary coding decisions and statistical mechanics into metanarratives about interpretive choice.” The implication is that they’re wrong to do this, but they are actually correct to do this. There is a long tradition in the history of statistics of thinking about the politics of statistics in terms of interpretation (see Ted Porter or Ken Danzinger or Mara Mills’ forthcoming discussion of the statistical construction of normal hearing); this is readily acknowledged by people who work in signal processing (eg, how to represent the behaviour of a spring reverb in an algorithm); and many authors in New Media Studies and Science and Technology Studies have shown that interpretation is heavily bound up with quantification and coding (some of the most important work in that are right now is appearing through Data & Society and Artificial Intelligence Now).

Finally, a small correction. The grant monies Da calculates from Andrew Piper’s CV in footnote 5 may seem almost impossible to American readers. I read the number as intended to indicate that it’s a waste of money to spend so much on data analytics, as a swipe at CLS. But while Piper is no doubt quite successful at getting grants, one can find equally successful Canadian humanists who have brought in similar or even greater amounts for no computational work whatsoever. One can peruse the online CVs of other humanities profs in Canada for comparison. In fact, most of this money winds up in the hands to students (or sometimes postdocs) who are hired as researchers (or whose degrees are funded). A relatively small part of it is for equipment or database access.** It becomes part of graduate funding, and I and my students have benefitted tremendously from this system (I also coauthor with my students frequently as a result). While DH may get the lion’s share of humanities funding in the US (I don’t know), this is not actually the case in Canada or at McGill University. Canada’s investment in humanities research isn’t perfect and it raises some issues I don’t like, but Critical Inquiry ought to be celebrating the use of public money to pay humanists to do research.

Do my concerns here detract from the force of Da’s critique of CLS? That depends on how you read them. But as we critique stuff, we also have to think carefully about what we are arguing for. I would like to see humanities that are heterodox, open to experimentation, and curious about approaches that are wildly different from their own.


*Discloures and disclaimers: Andrew Piper is a friend and a colleague, and Richard Jean So is a new hire at McGill who I’m excited and pleased to have. I am much less fired up about the quantitative turn in literary studies than either of them, though I’m always happy to hear about their work. Most of my work is historical, interpretative, philosophical and ethnographic. I am, however, the child of two quantitative social scientists and am “good at statistics,” and I work in a field that is home to both quantitative and qualitative researchers.

**The exception here is Canada Fund for Innovation Grants. But those are a special kind of hell. My failed CFI application to build a crossdisciplinary multimodal multimedia lab ranks as the second worst experience of my career as a scholar, from undergrad to the present day. Anyone who actually wins one has surely earned it.

The Waiting Game

Well, I’m still not ON the drugs. I’d expected it this week (because last week the doctor told me to expect it this week). But when you are taking a relatively new, hard core cancer drug, there are papers to file and bureaucratic procedures.

First, there’s insurance. I have insurance through Quebec, but private insurance through McGill. As I understand it, if you have private insurance, the province requires the private insurer to cover things like cancer drugs first. There was a form to fill out, which I submitted, and my oncologist, Dr. T. (I guess I’m going to need to give these people names), told me that they can’t refuse me coverage. Whether it’ll be 100% or somewhat less, I don’t know.

Then there’s getting the drug. This isn’t the kind of thing that’s in stock at the corner pharmacy. I was referred to a company called Eisai. They appear to be some kind of medical intermediary business that helps drug companies get their drugs approved in various countries (Canada being one of them), and at the same time provides a set of patient services. I have a nurse, who I’ll name Nurse H., and a pharmacist. They will deliver me the drug and a blood pressure monitor. They were the ones that sent me the glossy patient pamphlet as well. I actually appreciate the level of service, and it’ll be nice to be able to get through to a nurse if I have questions about side effects, etc. I can also email my oncologist but he’s constantly assaulted by demands on his attention, so the nurse is a better first call.

Then there’s the cost of the drug. The form I completed for Manulife, my insurer, lists the price at $2500CAD a month. Nurse H. said $5000CAD, and someone on my Facebook group (more on that in a future post) said in the US it’s over $22,000US a month. I think the stratospheric cost is what probably triggers this level of bureaucracy. I’m sure it’s also why I get a personal nurse who checks in on me, but for that level of cost, I’ll take it.

In the meantime, I’m in this weird space, sort of waiting for the next thing to happen. Maybe it’ll be next week. I am writing, reading, seeing people, playing music, and getting my modular synth ready for a big project that I hope to do while I’m getting settled on the drugs. Every time I’ve gone through something like this, I have found signal processing to be strangely meditative and centering. As is writing.

Back in the cancer saddle

Headlines: I’m going on a new cancer drug. It’s called lenvatinib (the brand name is Lenvima). It’s meant to be permanent, but there are side effects, and they need to see which ones I get and how I’m affected. So I’m cancelling all my travel for the next 3 months (my hope/goal is to be able to go to Berlin in June 2019) and in the meantime will more or less undergo a science experiment on myself. The hope is that I am going to be able to go on with my life indefinitely once we get everything stable, which is the whole point of the drug. My oncologist specifically said his goal was that I be on the highest possible dose while still being able to live my life. But that means experimenting.

More detail:

When I started blogging again I did not plan to return to my cancer patient hobby, but here we are, and the timing–blogwise and even lifewise with me on leave from teaching this term–is actually good in a way.

In case you missed the previous episode, here is some background:

I have metastatic thyroid cancer in my lungs. It was found when I was diagnosed with aggressive papillary thyroid cancer in 2009. My team of doctors been “watching and waiting” ever since, punctuated by occasional freakouts and one more round of radioactive iodine.

The big nodules in my lungs have been growing 1-2mm (with a margin of error) per year. And little ones keep appearing. So we’ve been in a “do nothing” (aka “watch and wait”) phase for almost a decade now. I have never said I am in remission, but on the advice of a good therapist, outside my test periods, I have operated in denial. You can’t spend a decade running around with your hair on fire saying “holy shit I have slow growing cancer in my lungs” every day. Also, my endocrinologist told me he expects me to die from something else.

/end background.

Now it’s almost a decade later. 1mm a year for 10 years is 1cm. And my three biggest spots are now close to an inch in diameter because the were already around 1cm when we started watching them. And there are more of them. Carrie came with me for my last CT results and was visibly shocked; I see the insides of my lungs every 6-12 months so am used to it. Later our “earth at night screensaver” came on the TV and she joked “that looks like your lungs.” Sort of like this but definitely not in the shape of the United States and remember, there are only 3 big spots:

US and a bit of Canada, Mexico and the Caribbean at night, image by NASA.

The original plan was to watch and wait: do something if my spots started multiplying more rapidly, or approaching the pleura, since that would be an escape route out of the lungs. I am still not in any immanent danger, and we don’t think I am symptomatic. (Though I was rediagnosed with asthma in the fall, which could theoretically be related.)

But things change. There’s a new drug, and my team of doctors has started taking a slightly more aggressive attitude. And while the spots aren’t at my pleura, they could get there at some point in the future.

The drug is called lenvantinib. Brand name Lenvima. The key thing to know is that it is part of a new approach to cancer called “targeted therapy” that targets cell behaviours other than rapid division, which is what normal chemo does. Also “targeted therapy” is a nicer name for the treatment than the old name for long term approaches like this–“soft chemo”–which evokes images of soft rock and soft boiled eggs, neither of which I find appealing. Also? It is very expensive, but I believe my insurance will cover it. More on that in a future post.

Levantinib is part of a class of drugs called “tyrosine kinase inhibitors.” It is not actually known for certain how they work. Even Lenvima’s own website hedges its bets:

LENVIMA is believed to block the signals that allow the cells (tumor and normal) to survive and multiply.

LENVIMA is believed to block signals that help blood vessels grow. Blood vessels support the tumor’s survival and growth.

“It is believed” is not a phrase you often hear in Western medicine, but at least I’m on the cutting edge here.

I will have lots more to say about how I feel, what I think about the drug, changes in the online culture of people with weird thyroid cancers from 2009 when I started, drug companies, Canadian medicine, and on and on.

More blog fodder in coming days and weeks. If you want to know how I’m doing the important information will be findable here.

Logic Functions Bonus Round for Synth Nerds

I learned about the XOR function, and pretty much everything I know about logic functions, from modular synthesis. Modular synthesis, like AI or any other media technology, works on a set of conventions ensconced in a set of standards. A modular synthesizer is basically an analog computer (this is a whole other post, which I will at some point write up) that separates sound from control (yes you can mix them up but let’s not worry about that for now), and works according to a set of standard voltages. So in my synthesizer, if I’m controlling pitch, a pitch will rise one octave with one volt. This is purely an agreed upon convention. A media standard. If you’re controlling a gate, let’s say to hear a pitched sound or not, it generally looks for the difference between 0 and some other number–maybe 1, 3 or 5 volts. So if the threshold is 3 volts, every time it receives 5 volts, it will make a sound. Every time it receives 2 volts, nothing happens. Yet of course the numbers 5 and 2 are different, as are those voltages.

Now, one can imagine controlling our synth sound with an XOR logic gate. Send continuously varying voltages, let’s say a pair of varying sine waves of different phase — one into input A and one into input B — and our logic module compares them. If it’s set to an XOR comparison, every time they are different, no matter what the amount or difference is, the gate outputs a 1 and you get sound. Every time they are the same, the gate outputs a zero and you get no sound. With this kind of XOR setup, most of the time you’ll have sound, with an occasional silence. Switch it to XAND, which outputs a 1 only when the voltages are the same, and you have the opposite scenario, only sound once in awhile. But again, the voltages could be any number, and could be varying quite wildly in different ways.

So in essence, the whole point of a binary logic gate is to reduce the blooming buzzing confusion of reality to two states: same or different. This is not a problem in modular synthesis: reduction and quantization are useful for all sorts of things. For instance, turning a set of continuously rising and dropping tones into a melody that makes musical sense–like a double bassist knowing where to put their finger on the fingerboard to play a note in tune. Quantization in audio is also incredibly useful and important, and the sampling theorem means that we can reconstruct continuous waves from discrete data points, as well as store big sounds in small places.

But when those binary operations are judgments about people, processes, or things that matter to people, the issue becomes something else entirely.

Sound or no sound is very different from qualified or not qualified, threat or no threat, human or gorilla. This is why a critique of quantification–or quantization as such is never enough.

A Few Random Thoughts on the Politics of the Logic Functions

Burç Kostem pointed me me to this wonderful piece by Matteo Pasquinelli on the history of neural networks. In the middle, there’s a small historical detail that I never quite grasped before:

In 1969 Marvin Minsky and Seymour Papert’s book, titled Perceptrons, attacked Rosenblatt’s neural network model by wrongly claiming that a Perceptron (although a simple single-layer one) could not learn the XOR function and solve classifications in higher dimensions. This recalcitrant book had a devastating impact, also because of Rosenblatt’s premature death in 1971, and blocked funds to neural network research for decades. What is termed as the first ‘winter of Artificial Intelligence’ would be better described as the ‘winter of neural networks,’ which lasted until 1986 when the two volumes Parallel Distributed Processing clarified that (multilayer) Perceptrons can actually learn complex logic functions.

In terms of the emerging historiography of machine learning, this is a place where the whig historians (aka, internalist histories by computer scientists) and the critical historians seem to agree.

But XOR as a test of neural networks, or intelligence, is a very interesting reduction of pattern recognition to a binary option. It is basically a binary calculation that compares two inputs. It answers “yes” if the two inputs are the different, and “no” if the two inputs are the same. Here’s a picture from Wikipedia to illustrate:

XOR Truth Table from Wikipedia

The whole political argument around pattern recognition in AI is basically reducible to what counts as a 1 and what counts as a 0 in that B column: this is why Google had to block its image recognition algorithms from identifying gorillas. Solving for XOR (are they different) or XAND (are they the same?) is never quite enough in real cultural contexts.

Here’s why. Let’s consider this against a famous statement of morphological resemblance from the history of social thought:

Is it surprising that prisons resemble factories, schools, barracks, hospitals, which all resemble prisons?

(Michel Foucault, Disipline and Punish)

This is an argument about general morphology. It is not that prisons and schools are exactly the same but that they share some meaningful aspects. Can an AI be trained to understand whether or not a given social arrangement fits a “panoptic diagram”? Only if you can parameterize every element of the description of a social milieu.

So we now have a media environment where neural nets can and do solve for binary logic functions all the time. The question is what trips that switch from 0 to 1 in either column. It is, in other words, built around a politics of classification.

For all the talk about process fairness and ethics in AI, we know from the history, anthropology, and STS study of classification that classifications are always tied to power.

Dynamic range compression isn’t “the problem” with music

Writers like Milner and a music business that still focuses on Top 40 charts are.

In a recent New York Times piece, Greg Milner argues that the “loudness wars” and “listener fatigue” are the reasons that music isn’t as good as it used to be. Here are some reasons why he’s wrong:

  1. Milner’s greatest offence is devaluating the work of thousands or hundreds of thousands of musicians making incredible new music, with less financial support and backing than ever. We are living through a “golden age” in more genres than I can possibly name or count. Go out and find the good music. There is nothing stopping you.
  2. Any analysis of any cultural phenomenon that explains it with a single causal factor is always wrong.
  3. If you were to actually look at the pop music charts from any of the years of the golden oldie hits references by Milner, you would find most of the music didn’t have the staying power or meaning he ascribes to the songs he uses as reference points.
  4. Any study of popular music today that uses top 40 charts and ignores the vast swaths of creativity one can find elsewhere–bandcamp, soundcloud, YouTube–has no right to say anything about the state or quality of music.
  5. An analogous argument about harmonic distortion and excessive studio production could have been made by fans of 1940s big band music, Mambo or Sinatra decrying the rock and pop that Milner celebrates.
  6. “I don’t understand hip hop” is not a convincing argument. Neither is “get off my lawn.”
  7. The listener fatigue argument is pseudo-scientific at best. Commercial success does not mean aesthetic merit; there are exactly zero reproducible psychology studies that connect musical pleasure in all people with specific sonic effects. Musical sound is only meaningful in context. The perceptual situation of someone listening on earbuds on the train is wholly different from that of someone paying attention in their living room with better speakers or headphones. That today’s most capitalized recommendation engines–Spotify, Google, etc–mostly rely on factors other than musical sound for their recommendations should tell you that little inherent meaning can be gleaned from tone or timbre alone.
  8. “There are millions of people in the basements, waiting to blow your mind.” — Vernon Reid.

Rupert Cox

Another new feature. When I meet cool or interesting people, or hear a talk that I find particularly engaging in one way or another, I will make mention of them here with a link to their work. I may or may not be loquacious.

So, first up is Rupert Cox, social anthropologist, who’s been doing interesting work on listening, sound, aircraft and the memory of war in Japan. He’s working on a book for Bloomsbury, but in the meantime, here’s a link to a short snippet of his work (you will need university library access for this one, sorry).

One of the things we talked about was the experience of sound outside the sound sounding in a physical way — remembered or imagined sounds, for instance. This seems especially important in writing about war and trauma, but has also become increasingly important for other areas: not only does this point come up a couple times in Remapping Sound Studies, but it’s important for thinking about sound, impairment and disability, as in Mack Hagood’s discussion of tinnitus.