Via this post by John Elder Robison I just read a new paper “Autism as a disorder of prediction” (the full text is available but took me a while to find; click on the red Adobe Acrobat icon).

Statistical vs. Logical Prediction

The paper talks about a particular kind of prediction, which they call “Markov chain” (you predict Y after X, if Y is statistically common after X). Markov chains are often used in computer programs to generate nonsense text that matches a “genre.” A Markov gibberish-generator analyzes how words tend to group together in a bunch of sample text, then generates text with those same statistical probabilities. The result is humorous nonsense that sounds like a parody of the genre used to “train” the gibberish-generator.

In the fields of linguistic anthropology and sociolinguistics, these statistical properties of text are thought of as largely-unconscious ways to convey meaning, often social meaning. For example, the way we talk (word choice, pronunciation, gestures, sentence length, etc.) is often enough to let other people place us socially: race, class, gender, even occupation. This social meaning may have nothing to do with the actual “denotational content” (dictionary definition) of what we’re saying; it’s a sort of overlay. And it’s based on statistical probability (stereotype, in essence) rather than logic. The same mechanisms that define a literary genre such as the “romance novel” also define genres of everyday speech and genres of identity.

A famous on-purpose example of Markov genre parody was Alan Sokal’s hoax in the journal Social Text. A famous not-on-purpose example is anything written by Thomas Friedman. These are texts that appear coherent and match a genre, but their logical content (i.e. what the words mean in a dictionary sense) makes no sense.

I’ve also known several people in real life who were like this, very charismatic, sounded very convincing in their social role, successful even, but the words meant very little beyond “I am talking like a CEO” or “I am talking like a minister.” These people were all-genre, no-content. Finally there’s an interesting condition called Williams Syndrome which leads to intellectual disability paired with the ability to convincingly inhabit social roles.

The point here is that the ability to make statistical language chains does appear  to be decoupled in some sense from the ability to think logically.

With that background, let’s get back to the paper.

Problem: Attempted unified account of autism

Let’s note that the paper is on a fool’s errand trying to unify and explain all autism, and these authors should take Lynn Waterhouse’s advice on that front .  The paper would still be interesting if it led to an account of some conditions underlying autism, and that should be our default interpretation of it.

Problem: Assumption that if it correlates with autism, it must be bad

And let’s note that the paper takes whatever thing correlates with autism and assumes it must be a bad thing, which is simply unscientific as usual. Statistical Markov-style prediction creates many of the known cognitive biases, not to mention racism, sexism, and a lot of nonsensical speaking and writing. Unquestionably, this kind of thinking is also necessary for humans to get things done — it isn’t effective to think everything out explicitly all the time — but there’s a reason we have the ability to both “pattern match” and also think things through step-by-step.

Possible meaning and future direction: intellect and instinct again

In the past I listed 20+ word pairs used in everyday language and various academic fields to refer to the distinction between “intellect” and “instinct.” Here’s the post listing those and talking about how I feel it relates to autism.

To me this prediction paper is getting at the same dichotomy again, this time calling “instinct” “Markov prediction.”

Generations of academics from all kinds of fields have wrestled with this distinction. The theoretical framework is weak; what does this distinction really mean? What does it mean biologically? What does it mean in everyday experience? At the same time, there’s a lot of data and previous work that could be learned from. Sociolinguistics and linguistic anthropology have a lot to offer here, among other fields. There’s no need to start from scratch.

Rather than unifying autism, can we unify some of the past work on this intellect vs. instinct dimension of human experience? Can we figure out if this is a real thing or just a folk theory? Can we decide on ways to measure this dimension?

I think there’s interesting work to be done, somewhere in the vicinity of this paper.

Alternative explanations for prediction failure

Before concluding that failure to Markov-predict is due to deficits in prediction ability (due to our unscientific bias against autism), here are two examples of alternative explanations we could also consider.

1. Attention and interest.  That is, interest in instinctive prediction vs. a preference for other kinds of cognitive activity. For example, I typically feel pretty good (even superior) at pattern matching in areas of intellectual interest. John Elder Robison brings up the example of autistic kids playing a video game they’re really into.
2. Interference from a “thinking slow” filter. Introspectively, I feel that step-by-step intellectual thinking builds on a “pattern matching” substrate; each “step” is a pattern match, and you learn to apply a sequence of pattern matches to get to a conclusion. The difference between this and a Markov chain is that you aren’t allowed to use “false” pattern matches (stereotypical associations or genre features), you can only use those that make sense on the literal/factual/denotational/logical level. This is an extra level of filtering which may hinder automatic, instinctive pattern matching. Framed this way, the “problem” is that someone is doing more on top of and beyond Markov prediction, rather than that their Markov-predictor is broken.

I don’t know if these are right; the point is, we’d consider them if we didn’t jump straight to “if it’s autism, it’s broken.”

Avoiding fuzzy thinking

This paper is a great example of how our biases (toward autism as a unified condition, and toward autism as an unqualified deficit) can cloud our insights.

If we open our minds to the possibility that we can understand some autism, and to the possibility that some traits may be both strengths and deficits, depending on context, we can explore a lot of ideas that we might otherwise miss.

To find the truth researchers must make things more complicated.