Open threads at the Open Thread tab every Sunday and Wednesday

It’s Bayes All The Way Up

[Epistemic status: Very speculative. I am not a neuroscientist and apologize for any misinterpretation of the papers involved. Thanks to the people who posted these papers in r/slatestarcodex. See also Mysticism and Pattern-Matching and Bayes For Schizophrenics]

Bayes’ Theorem is an equation for calculating certain kinds of conditional probabilities. For something so obscure, it’s attracted a surprisingly wide fanbase, including doctors, environmental scientists, economists, bodybuilders, fen-dwellers, and international smugglers. Eventually the hype reached the point where there was both a Bayesian cabaret and a Bayesian choir, popular books using Bayes’ Theorem to prove both the existence and the nonexistence of God, and even Bayesian dating advice. Eventually everyone agreed to dial down their exuberance a little, and accept that Bayes’ Theorem might not literally explain absolutely everything.

So – did you know that the neurotransmitters in the brain might represent different terms in Bayes’ Theorem?

First things first: Bayes’ Theorem is a mathematical framework for integrating new evidence with prior beliefs. For example, suppose you’re sitting in your quiet suburban home and you hear something that sounds like a lion roaring. You have some prior beliefs that lions are unlikely to be near your house, so you figure that it’s probably not a lion. Probably it’s some weird machine of your neighbor’s that just happens to sound like a lion, or some kids pranking you by playing lion noises, or something. You end up believing that there’s probably no lion nearby, but you do have a slightly higher probability of there being a lion nearby than you had before you heard the roaring noise. Bayes’ Theorem is just this kind of reasoning converted to math. You can find the long version here.

This is what the brain does too: integrate new evidence with prior beliefs. Here are some examples I’ve used on this blog before:

All three of these are examples of top-down processing. Bottom-up processing is when you build perceptions into a model of the the world. Top-down processing is when you let your models of the world influence your perceptions. In the first image, you view the center letter of the the first word as an H and the second as an A, even though they’re the the same character; your model of the world tells you that THE CAT is more likely than TAE CHT. In the second image, you read “PARIS IN THE SPRINGTIME”, skimming over the duplication of the word “the”; your model of the world tells you that the phrase should probably only have one “the” in it (just as you’ve probably skimmed over it the three times I’ve duplicated “the” in this paragraph alone!). The third image might look meaningless until you realize it’s a cow’s head; once you see the cow’s head your model of the world informs your perception and it’s almost impossible to see it as anything else.

(Teh fcat taht you can siltl raed wrods wtih all the itroneir ltretrs rgraneanrd is ahonter empxlae of top-dwon pssirocneg mkinag nsioy btotom-up dtaa sanp itno pacle)

But top-down processing is much more omnipresent than even these examples would suggest. Even something as simple as looking out the window and seeing a tree requires top-down processing; it may be too dark or foggy to see the tree one hundred percent clearly, the exact pattern of light and darkness on the tree might be something you’ve never seen before – but because you know what trees are and expect them to be around, the image “snaps” into the schema “tree” and you see a tree there. As usual, this process is most obvious when it goes wrong; for example, when random patterns on a wall or ceiling “snap” into the image of a face, or when the whistling of the wind “snaps” into a voice calling your name.

Corlett, Frith & Fletcher (2009) (henceforth CFF) expand on this idea and speculate on the biochemical substrates of each part of the process. They view perception as a “handshake” between top-down and bottom-up processing. Top-down models predict what we’re going to see, bottom-up models perceive the real world, then they meet in the middle and compare notes to calculate a prediction error. When the prediction error is low enough, it gets smoothed over into a consensus view of reality. When the prediction error is too high, it registers as salience/surprise, and we focus our attention on the stimulus involved to try to reconcile the models. If it turns out that bottom-up was right and top-down was wrong, then we adjust our priors (ie the models used by the top-down systems) and so learning occurs.

In their model, bottom-up sensory processing involves glutamate via the AMPA receptor, and top-down sensory processing involves glutamate via the NMDA receptor. Dopamine codes for prediction error, and seem to represent the level of certainty or the “confidence interval” of a given prediction or perception. Serotonin, acetylcholine, and the others seem to modulate these systems, where “modulate” is a generic neuroscientist weasel word. They provide a lot of neurological and radiologic evidence for these correspondences, for which I highly recommend reading the paper but which I’m not going to get into here. What I found interesting was their attempts to match this system to known pharmacological and psychological processes.

CFF discuss a couple of possible disruptions of their system. Consider increased AMPA signaling combined with decreased NMDA signaling. Bottom-up processing would become more powerful, unrestrained by top-down models. The world would seem to become “noisier”, as sensory inputs took on a life of their own and failed to snap into existing categories. In extreme cases, the “handshake” between exuberant bottom-up processes and overly timid top-down processes would fail completely, which would take the form of the sudden assignment of salience to a random stimulus.

Schizophrenics are famous for “delusions of reference”, where they think a random object or phrase is deeply important for reasons they have trouble explaining. Wikipedia gives as examples:

– A feeling that people on television or radio are talking about or talking directly to them

– Believing that headlines or stories in newspapers are written especially for them

– Seeing objects or events as being set up deliberately to convey a special or particular meaning to themselves

– Thinking ‘that the slightest careless movement on the part of another person had great personal meaning…increased significance’

In CFF, these are perceptual handshake failures; even though “there’s a story about the economy in today’s newspaper” should be perfectly predictable, noisy AMPA signaling registers it as an extreme prediction failure, and it fails its perceptual handshake with overly-weak priors. Then it gets flagged as shocking and deeply important. If you’re unlucky enough to have your brain flag a random newspaper article as shocking and deeply important, maybe phenomenologically that feels like it’s a secret message for you.

And this pattern – increased AMPA signaling combined with decreased NMDA signaling – is pretty much the effect profile of the drug ketamine, and ketamine does cause a paranoid psychosis mixed with delusions of reference.

Organic psychosis like schizophrenia might involve a similar process. There’s a test called the binocular depth inversion illusion, which looks like this:


The mask in the picture is concave, ie the nose is furthest away from the camera. But most viewers interpret it as convex, with the nose closest to the camera. This makes sense in terms of Bayesian perception; we see right-side-in faces a whole lot more often than inside-out faces.

Schizophrenics (and people stoned on marijuana!) are more likely to properly identify the face as concave than everyone else. In CFF’s system, something about schizophrenia and marijuana messes with NMDA, impairs priors, and reduces the power of top-down processing. This predicts that schizophrenics and potheads would both have paranoia and delusions of reference, which seems about right.

Consider a slightly different distortion: increased AMPA signaling combined with increased NMDA signaling. You’ve still got a lot of sensory noise. But you’ve also got stronger priors to try to make sense of them. CFF argue these are the perfect conditions to create hallucinations. The increase in sensory noise means there’s a lot of data to be explained; the increased top-down pattern-matching means that the brain is very keen to fit all of it into some grand narrative. The result is vivid, convincing hallucinations of things that are totally not there at all.

LSD is mostly serotonergic, but most things that happen in the brain bottom out in glutamate eventually, and LSD bottoms out in exactly the pattern of increased AMPA and increased NMDA that we would expect to produce hallucinations. CFF don’t mention this, but I would also like to add my theory of pattern-matching based mysticism. Make the top-down prior-using NMDA system strong enough, and the entire world collapses into a single narrative, a divine grand plan in which everything makes sense and you understand all of it. This is also something I associate with LSD.

If dopamine represents a confidence interval, then increased dopaminergic signaling should mean narrowed confidence intervals and increased certainty. Perceptually, this would correspond to increased sensory acuity. More abstractly, it might increase “self-confidence” as usually described. Amphetamines, which act as dopamine agonists, do both. Amphetamine users report increased visual acuity (weirdly, they also report blurred vision sometimes; I don’t understand exactly what’s going on here). They also create an elevated mood and grandiose delusions, making users more sure of themselves and making them feel like they can do anything.

(something I remain confused about: elevated mood and grandiose delusions are also typical of bipolar mania. People on amphetamines and other dopamine agonists act pretty much exactly like manic people. Antidopaminergic drugs like olanzapine are very effective acute antimanics. But people don’t generally think of mania as primarily dopaminergic. Why not?)

CFF end their paper with a discussion of sensory deprivation. If perception is a handshake between bottom-up sense-data and top-down priors, what happens when we turn the sense-data off entirely? Psychologists note that most people go a little crazy when placed in total sensory deprivation, but that schizophrenics actually seem to do better under sense-deprivation conditions. Why?

The brain filters sense-data to adjust for ambient conditions. For example, when it’s very dark, your eyes gradually adjust until you can see by whatever light is present. When it’s perfectly silent, you can hear the proverbial pin drop. In a state of total sensory deprivation, any attempt to adjust to a threshold where you can detect the nonexistent signal is actually just going to bring you down below the point where you’re picking up noise. As with LSD, when there’s too much noise the top-down systems do their best to impose structure on it, leading to hallucinations; when they fail, you get delusions. If schizophrenics have inherently noisy perceptual systems, such that all perception comes with noise the same way a bad microphone gives off bursts of static whenever anyone tries to speak into it, then their brains will actually become less noisy as sense-data disappears.

(this might be a good time to remember that no congentally blind people ever develop schizophrenia and no one knows why)


Lawson, Rees, and Friston (2014) offer a Bayesian link to autism.

(there are probably a lot of links between Bayesians and autism, but this is the only one that needs a journal article)

They argue that autism is a form of aberrant precision. That is, confidence intervals are too low; bottom-up sense-data cannot handshake with top-down models unless they’re almost-exactly the same. Since they rarely are, top-down models lose their ability to “smooth over” bottom-up information. The world is full of random noise that fails to cohere into any more general plan.

Right now I’m sitting in a room writing on a computer. A white noise machine produces white noise. A fluorescent lamp flickers overhead. My body is doing all sorts of body stuff like digesting food and pumping blood. There are a few things I need to concentrate on: this essay I’m writing, my pager if it goes off, any sorts of sudden dramatic pains in my body that might indicate a life-threatening illness. But I don’t need to worry about the feeling of my back against the back fo the chair, or the occasional flickers of the fluorescent light, or the feeling of my shirt on my skin.

A well-functioning perceptual system gates out those things I don’t need to worry about. Since my shirt always feels more or less similar on my skin, my top-down model learns to predict that feeling. When the top-down model predicts the shirt on my skin, and my bottom-up sensation reports the shirt on my skin, they handshake and agree that all is well. Even if a slight change in posture makes a different part of my shirt brush against my skin than usual, the confidence intervals are wide: it is still an instance of the class “shirt on skin”, it “snaps” into my shirt-on-skin schema, and the perceptual handshake goes off successfully, and all remains well. If something dramatic happens – for example my pager starts beeping really loudly – then my top-down model, which has thus far predicted silence – is rudely surprised by the sudden burst of noise. The perceptual handshake fails, and I am startled, upset, and instantly stop writing my essay as I try to figure out what to do next (hopefully answer my pager). The system works.

The autistic version works differently. The top-down model tries to predict the feeling of the shirt on my skin, but tiny changes in the position of the shirt change the feeling somewhat; bottom-up data does not quite match top-down prediction. In a neurotypical with wide confidence intervals, the brain would shrug off such a tiny difference, declare it good enough for government work, and (correctly) ignore it. In an autistic person, the confidence intervals are very narrow; the top-down systems expect the feeling of shirt-on-skin, but the bottom-up systems report a slightly different feeling of shirt-on-skin. These fail to snap together, the perceptual handshake fails, and the brain flags it as important; the autistic person is startled, upset, and feels like stopping what they’re doing in order to attend to it.

(in fact, I think the paper might be claiming that “attention” just means a localized narrowing of confidence intervals in a certain direction; for example, if I pay attention to the feeling of my shirt on my skin, then I can feel every little fold and micromovement. This seems like an important point with a lot of implications.)

Such handshake failures match some of the sensory symptoms of autism pretty well. Autistic people dislike environments that are (literally or metaphorically) noisy. Small sensory imperfections bother them. They literally get annoyed by scratchy clothing. They tend to seek routine, make sure everything is maximally predictable, and act as if even tiny deviations from normal are worthy of alarm.

They also stim. LRF interpret stimming as an attempt to control sensory predictive environment. If you’re moving your arms in a rhythmic motion, the overwhelming majority of sensory input from your arm is from that rhythmic motion; tiny deviations get lost in the larger signal, the same way a firefly would disappear when seen against the blaze of a searchlight. The rhythmic signal which you yourself are creating and keeping maximally rhythmic is the most predictable thing possible. Even something like head-banging serves to create extremely strong sensory data – sensory data whose production the head-banger is themselves in complete control of. If the brain is in some sense minimizing predictive error, and there’s no reasonable way to minimize prediction error because your predictive system is messed up and registering everything as a dangerous error – then sometimes you have to take things into your own hands, bang your head against a metal wall, and say “I totally predicted all that pain”.

(the paper doesn’t mention this, but it wouldn’t surprise me if weighted blankets work the same way. A bunch of weights placed on top of you will predictably stay there; if they’re heavy enough this is one of the strongest sensory signals you’re receiving and it might “raise your average” in terms of having low predictive error)

What about all the non-sensory-gating-related symptoms of autism? LRF think that autistic people dislike social interaction because it’s “the greatest uncertainty”; other people are the hardest-to-predict things we encounter. Neurotypical people are able to smooth social interaction into general categories: this person seems friendly, that person probably doesn’t like me. Autistic people get the same bottom-up data: an eye-twitch here, a weird half-smile there – but it never snaps into recognizable models; it just stays weird uninterpretable clues. So:

This provides a simple explanation for the pronounced social-communication difficulties in autism; given that other agents are arguably the most difficult things to predict. In the complex world of social interactions, the many-to-one mappings between causes and sensory input are dramatically increased and difficult to learn; especially if one cannot contextualize the prediction errors that drive that learning.

They don’t really address differences between autists and neurotypicals in terms of personality or skills. But a lot of people have come up with stories about how autistic people are better at tasks that require a lot of precision and less good at tasks that require central coherence, which seems like sort of what this theory would predict.

LRF ends by discussing biochemical bases. They agree with CFF that top-down processing is probably related to NMDA receptors, and so suspect this is damaged in autism. Transgenic mice who lack an important NMDA receptor component seem to behave kind of like autistic humans, which they take as support for their model – although obviously a lot more research is needed. They agree that acetylcholine “modulates” all of this and suggest it might be a promising pathway for future research. They agree with CFF that dopamine may represent precision/confidence, but despite their whole spiel being that precision/confidence is messed up in autism, they don’t have much to say about dopamine except that it probably modulates something, just like everything else.


All of this is fascinating and elegant. But is it elegant enough?

I notice that I am confused about the relative role of NMDA and AMPA in producing hallucinations and delusions. CFF say that enhanced NMDA signaling results in hallucinations as the brain tries to add excess order to experience and “overfits” the visual data. Fine. So maybe you get a tiny bit of visual noise and think you’re seeing the Devil. But shouldn’t NMDA and top-down processing also be the system that tells you there is a high prior against the Devil being in any particular visual region?

Also, once psychotics develop a delusion, that delusion usually sticks around. It might be that a stray word in a newspaper makes someone think that the FBI is after them, but once they think the FBI is after them, they fit everything into this new paradigm – for example, they might think their psychiatrist is an FBI agent sent to poison them. This sounds a lot like a new, very strong prior! Their doctor presumably isn’t doing much that seems FBI-agent-ish, but because they’re working off a narrative of the FBI coming to get them, they fit everything, including their doctor, into that story. But if psychosis is a case of attenuated priors, why should that be?

(maybe they would answer that because psychotic people also have increased dopamine, they believe in the FBI with absolute certainty? But then how come most psychotics don’t seem to be manic – that is, why aren’t they overconfident in anything except their delusions?)

LRF discuss prediction error in terms of mild surprise and annoyance; you didn’t expect a beeping noise, the beeping noise happened, so you become startled. CFF discuss prediction error as sudden surprising salience, but then say that the attribution of salience to an odd stimulus creates a delusion of reference, a belief that it’s somehow pregnant with secret messages. These are two very different views of prediction error; an autist wearing uncomfortable clothes might be constantly focusing on their itchiness rather than on whatever she’s trying to do at the time, but she’s not going to start thinking they’re a sign from God. What’s the difference?

Finally, although they highlighted a selection of drugs that make sense within their model, others seem not to. For example, there’s some discussion of ampakines for schizophrenia. But this is the opposite of what you’d want if psychosis involved overactive AMPA signaling! I’m not saying that the ampakines for schizophrenia definitely work, but they don’t seem to make the schizophrenia noticeably worse either.

Probably this will end the same way most things in psychiatry end – hopelessly bogged down in complexity. Probably AMPA does one thing in one part of the brain, the opposite in other parts of the brain, and it’s all nonlinear and different amounts of AMPA will have totally different effects and maybe downregulate itself somewhere else.

Still, it’s neat to have at least a vague high-level overview of what might be going on.

This entry was posted in Uncategorized and tagged , . Bookmark the permalink.

215 Responses to It’s Bayes All The Way Up

  1. William O. B'Livion says:

    > Eventually…Bayesian cabarets

    But is there a Bayesian Cabernet, which is what my brain REALLY wanted that to read.

    • TK-421 says:

      I misread it as “Bayesian cabinets”, and was briefly surprised that there was a head of state somewhere who was mathematically fluent enough to have their ministers properly calibrate their confidence intervals.

  2. Phlinn says:

    After reading this, I did a quick search on ‘autism LSD’ and apparently it’s been successfully used as a treatment, although I believe the controlled studies are a bit lacking. For some reason, despite not being the point of the post, those sections caught my attention and suggested a reason LSD would be effective.

    Upon further thought, I now wonder if fibromyalgia might be a special case of the handshake failure you described for autism.

  3. Siah Sargus says:

    How do normal people deal with the tags RIGHT ON THEIR NECK all of the time? Why would you want that?

    Also is this related to the whole “your tongue is in your mouth” sort of deal never doing anything for me?

    • UltraRedSpectrum says:

      I’m pretty sure autism is a broad spectrum of interconnected screwups on top of the failure to handshake. ‘Cause yeah, those tags are friggen’ painful.

    • Toggle says:

      In order to consciously feel the tag in my shirt, I had to physically reach around and wiggle it some. That sensation is deeply invisible to me.

    • Scott Alexander says:

      Yeah, I hate tags and cut them off all of my shirts. Somebody told me it was an autistic thing but it just seems like common sense.

      • SlightlyLessHairyApe says:

        But how will you know if a shirt has to be washed in warm or cold water and whether it should be with like colors?!!

        [ Only 50% joking. My partner thinks I’m crazy for following the instructions on the tags. ]

        • Gazeboist says:

          Does anything not say “with like colors”? I’ve never seen an example of such.

          The worst clothes, of course, are the ones that just have a tenperature.

        • Dahlen says:

          [ Only 50% joking. My partner thinks I’m crazy for following the instructions on the tags. ]

          Preach. I’ve taken to doing all of my own laundry (yay leveling up in adulthood) because of the OCD horror that whoever is doing the laundry in my household irons everything as though it were linen (on the very hottest temperature). Whereas all I own is to be ironed on the lowest temperature. How many times do I need to explain this?!

        • Berna says:

          Bed sheets and towels get washed at 95 °C, the rest gets washed together at 30 °C. If that messes anything up, it gets thrown out.

        • Shieldfoss says:

          These are the Rules, as taught by my mother, as taught by her mother.

          The rules can be deviated from once you have experience. Some of the common deviations are described at the bottom of the post with a footnote.

          The rules all have reasons. Some reasons are complicated and will get skipped. Others get marked with a footnote.

          The purpose of these rules are: Avoid unhygienic situations, avoid damaging clothes by exposing them to too-high temperatures, avoid mis-colored clothes by mixtures that are more likely to stain or by temperatures that are too high for the dyes.

          For all cloth: Wash inside-out with zippers/buttons etc. closed.[0]

          For all cloth: Wash before use at 30°C.[1]

          Grouping by colors:

          Clothes come in 4 colors, detergents in 4+1. Clothes are “Light, Dark, Red, Color,” and the +1 detergent is Bleaching White. Cloth is considered “light” if it is white, or if it could be mistaken for white in bad lighting or by somebody with a limited color vocabulary. Dark clothes same rule except “black” instead of “white.”

          Clothes that have more than 50% non-light color are not light.[2]

          Clothes that have more than 10% red are are red.[3]

          First time wash: Alone, or with extremely similar colors, even if it doesn’t warn you to wash only with similar colors.

          Second time and on: How much you split your wash up or join it together depends on how great a hurry you’re in and what detergents you have.

          At the very basic, get two separate detergents – “white” and “color.”

          Whites are to be washed separately. Dark can be grouped with either “Red” or with “Color,” and does not need to be considered a separate color until later.[4]

          Caution: Bleaching White detergent is only for actually-white clothes, others should be washed with non-bleaching white or (if you accidentally only purchased bleaching white) with color detergent.

          Once you have Color and Non-bleaching White, supply up with Dark, Red and Bleaching White detergents as fits your wardrobe.[5]

          Grouping by temperature:

          Underwear: At least 60°C, no matter the color, material or what the label says.[6]

          Linen: 60°C.

          Very dity linen: You can go all the way up to 90°C if white, otherwise wash at 60°C but for longer.[7]

          Cotton: 30°C if non-white, otherwise 60°C if white.

          Very dirty cotton: You can go all the way up to 60°C

          Synthetics: 30°C

          Very dirty synthetics: 30°C, but for longer.

          Wool: 30°C at most, see if your washing machine has an actual wool program, unless you are washing wool underwear in which case 60°C. Do not centrifuge wool.

          Silk: I am unfamiliar with silk. I will instead link to this:

          and ask you to pay attention to the fact that this is not hygienic. Do not purchase silk underwear.

          In order of durability, Linen > cotton > synthetics > wool > silk. Some wools can endure washing machines, others cannot. I generally treat all wools like they’re synthetics, from the principle that I was never going to hand wash them anyway so worst case I cannot wear them because they’re ruined, same situation as before when I couldn’t wear them because they were dirty. This has so far ruined exactly zero wool garments for me but you should be aware of the possibility. Wool that you cannot endure losing (gift from your dead grandmother or w/e) you should hand wash using special wool soap.

          [0] Washing machines exert wear on your clothes. You turn them inside-out so the wear is on the side that isn’t visible. You close buttons/zippers so they don’t get caught on each other and exert even more wear, and so they don’t knock into the cylinder and make noise.

          [1] The first wash is to get rid of excess dye, chemicals used in cothing production, pesticides used while growing the materials etc. It shouldn’t be necessary but unfortunately it often is.

          [2] The reasoning is that almost any colors can stain white, so don’t throw them into your whites wash. This is one of the rules that some experience can tell you when to ignore – in particular, if you have e.g. a white shirt with blue stripes, if the blue stripes aren’t miscoloring the rest of the white shirt, they probably will also not miscolor your other whites.

          [3] The reasoning is that red can stain many other colors, so don’t throw them into your other-colors wash. Same deviation as above – if the red accent on a partially red garment isn’t staining the rest of the garment, it probably also will not stain your other garments. Just be very careful mixing reds with whites.

          [4] Red+Dark is allowed because red staining does not show on dark clothes. Dark+Color is allowed because Dark and Color are washed exactly the same unless you have dark detergent or a dark program on your machine.

          [5] Bleaching white is basically for white linen that has a tendency to acquire a yellowish sheen even when clean, but can be used for other completely-white clothes.

          [6] You want it at 60 degrees to kill bacteria. If you have synthetic, silk or wool underwear, the label may warn you that this will ruin your underwear. This does not mean you wash synthetic/wool underwear differently, it means you should buy cotton underwear.

          [7] High temperatures can be murder on colors

          • Deiseach says:

            Agree with this, although I am not so fastidious 🙂

            You should always wash a new garment, simply to see if the dye will run – some materials do, and you don’t want to find this out when your wash all comes out a uniform pale blue when it didn’t go in that way.

            Whites/mostly white – all in together and at a higher temperature because nothing gets whites clean like a boil wash – except the synthetic materials of today can’t handle that heat so as hot as you can safely go, and never mind that “our detergent gets it just as clean at a lower temperature” racket. Whites that are whites and/or cotton, linen, non-synthetics – add bleach (there are ‘gentler’ laundry bleaches now) for extra hygiene/cleaning power. Especially if you’re doing a wash after someone’s been ill and you’re washing bed linen, towels, etc.

            Colours/mostly coloured – all in together again and usually at a lower temperature. May need to sub-sort these out based on what material they are and lighter/darker colours. Be careful with reds – they do tend to run and when they do you may as well dump all the things they stained because it is never coming out.

            Things that say “hand wash only” – you probably can get away with putting these on to wash at a low temperature. I’ve been lucky so far with this because who wants to stand over a sink full of sudsy water scrubbing clothes with your own actual hands like a cavewoman? 🙂

            Exceptions to above: silk and wool. Wool will shrink if it is not washed according to directions and it may be a pain in the backside to have it dripping everywhere when you can’t tumble dry it but that’s better than taking it out of the washing machine or dryer and finding it is now two or three sizes smaller.

            I hate ironing so I don’t do it but yes, do not use “one temperature for everything”.

        • pheltz says:

          It’s very simple. Wash everything cold. Virtually everything these days has something in it that is harmed by excessive heat, even if it’s only the stitching. Don’t worry about mixing colors. Modern clothing doesn’t really run.

          • The Nybbler says:

            If modern clothing doesn’t run, why do new blue jeans stain my hands?

          • Bassicallyboss says:

            It’s my experience that colors often run when new, but pretty much never does after 2 washes. So be careful to separate colors when washing new garments, but otherwise don’t worry.

          • pheltz says:

            Blue jeans are an exception; they’re done a little differently, since people expect them to fade, distress, etc.

          • caryatis says:

            I do this too. Who wants to spend hours on laundry?

        • caryatis says:

          You also would forget all the sizes.

      • Tibor says:

        Actually, I noticed they’ve started sewing them into the seam at the side of the shirts/T-shirts, at least the ones I buy (but I don’t do that on purpose even though I hate the tags at the neck as well). Sometimes the information is printed on the T-shirt itself, avoiding the tag completely.

        • Phil says:

          that’s sort of interesting, I find the side tags extremely annoying and cut them off, but have never really been bothered by the neck tags

          probably because the side tags disturb my priors (and get noticed), while I have a lifetime of priors for neck tags (and don’t notice them)

      • Aaron Woodside says:

        I never feel tags. In fact, while reading this I tried to see if I could. I focused really hard, noticing all the different shirt sensations, then slowly tried to narrow my focus to the neck where the tag would be. After about 5 minutes I started feeling the tag, but only by deliberately moving my neck back and forth…

        then i reached up to re-position the tag to see if it would make a difference, and discovered I’m wearing a tag-less shirt!

      • Bugmaster says:

        If I leave the tags on, they scratch my skin and leave red patches of inflammation / skin irritation if I wear the shirt for a long time. I’ve found that, with most tags, cutting off the corners is sufficient to solve the problem. I’m not entirely sure how autism would explain that.

    • Baughn says:

      The things are bloody annoying, autistic spectrum or not. I can ignore them most of the time, but sometimes they’re made of some scratchy material that actually gives me a welt.

      I’ve found that, for this case, scissors work well.

    • Vaniver says:

      I think I only notice the tag if it somehow flips up and so is rubbing the part of my neck above the shirtline. Otherwise, yeah, it’s invisible.

      When in my apartment, I typically strip down to just my boxers because it’s slightly more comfortable / it reduces the tiny cognitive load associated with clothes (typically the part that’s not resting correctly, rather than the shift of perception with movement). This works better if you live alone, but is still probably worth considering if you don’t. (This is especially worthwhile if you live in a hot place, since you can then save money on AC.)

      • Virbie says:

        +1, I can’t say I really understand when people are fully clothed at home (especially with shoes!). I do this all the time and every roommate I’ve ever had has apparently just silently accepted it, though having your shirt off is an acceptable compromise if your roommates have a problem for some reason.

    • TK-421 says:

      How do normal people deal with the tags RIGHT ON THEIR NECK all of the time?


    • pku says:

      I always rip those and, and assumed everyone else did too. Are you saying some people leave those in? How do they live like that?!

    • I seem to be a normal person in this regard. I just don’t feel clothing tags. It’s not a matter of getting used to them.

      I appreciate manufacturers who print sizes/washing instructions on the insides of clothing, but it’s because it presumably makes life easier for a lot of other people, it doesn’t do me any good. It’s probably cheaper for the manufacturers.

      Damned if I can figure out how people can enjoy loud music or the taste of beer.

      • Kyle Strand says:

        You need good beer first. Also, different people prefer different beers. I have a friend who didn’t like beer at all until he tried some really strong sours. Meanwhile, I didn’t think I liked beer until I tried a Belgian Tripel, and now my favorite beer is probably coffee stouts.

    • Deiseach says:

      How do normal people deal with the tags RIGHT ON THEIR NECK all of the time?

      I rip the tags off, I have to, else I’ll go crazy with the ITCHY SCRATCHY IRRITATING RUBBING AGAINST MY SKIN SENSATION. Dunno if I fit into category “normal person” 🙂

    • Jill says:

      Some people have reactions to certain fabrics, sort of like an allergy, although I don’t know if it is ever technically that. In that case, it would be a problem mainly when the tag is made of a certain material. But other people’s skins are overall more sensitive than others to things like clothing tags, without the person necessarily being autistic.

      Sometimes a sort of over-sensitivity is a gift. I suppose wine tasters and other people who taste various fine foods for a living probably have such a gift. Maybe some musicians have certain kinds of what might be termed over-sensitivity to sound.

      I guess if the over-sensitivity works for you, or at least doesn’t impair you, then good. If it does get in your way, then you and/or others think you have a disease. But some of this depends on your particular environment, line of work etc. and the sorts of sensory stimulation that are common there.

      • Cadie says:

        I rarely notice tags, but occasionally if they’re made with very scratchy threads then I will. I had a pair of shirts that had some sort of weird metallic thread that felt like bits of plastic, and I ended up cutting the tags off. They kept scratching me and getting on my nerves. No idea why the company thought using stiff plastic-like threads on a shirt tag was a good idea, especially for a children’s shirt. Mine were extra-large, which is big enough for many adults, but the smaller sizes are identical except smaller and with a different size printed on the same type of tag, and clearly those are meant for actual children. Who tend to be more tag-sensitive than adults even if they don’t have any particular sensory issues. I don’t understand that design decision at all.

      • Besserwisser says:

        I hate knitwear and can’t comprehend how anyone can wear that, much less find it comfortable which I’ve often see them described as. The most common reason for not wearing the sweater made by grandma is because of how it looks, which would be a very secondary concern to me.

        • Deiseach says:

          I think traditionally people wouldn’t have worn wool next to their bare skin as the reason for big woolly jumpers/sweaters was it is darn cold and you’d be wearing layers like a shirt and undergarments under the sweater anyway.

          • A Kind Of Dog-Octopus says:

            As someone who just spent three days in the same pair of wool hose in high temperatures it does feel scratchy at first but is easily ignored after a while – or maybe it just wears down? Either way, it’s fine. And one thing we’ve mostly forgotten as a modern society is that wool is great in heat, too – it wicks away and evaporate the sweat.

            So yeah, people did wear wool on the skin frequently. From hose to hats to kilts to big scratchy blankets. And it’s great. Though wool coats were more often worn over body linens.

          • Tracy W says:

            Icebreaker merino wool is made to be worn next to the skin. I’ve worn it for days tramping (hiking) without issue.

    • HeelBearCub says:

      Well, ?most? shirts seem to have gone the tagless route, where they just print the label right onto the fabric. So, it seems like it’s common enough to find shirt tags annoying. Mostly for me this only happened if the tag wasn’t lying flat, for whatever reason.

    • I hate the tags so much that I have, generalizing from my own experience, considered them compelling evidence of civilization incompetence.

    • Zakharov says:

      I hated clothing tags when I was a kid, but don’t notice them anymore.

    • Jai says:

      While we’re on the subject – the t-shirts they gave out at EAG 2015 (not sure about 2016) had no tags and were incredibly tolerable, more than pretty much every other shirt I’ve ever found. I’ve since looked up the exact shirt and purchased enough to wear every day:

    • I mostly just don’t notice tags – in the sense of “Oh yeah, I suppose some of my shirts have those, I can remember looking at them and seeing them, they’re useful for washing instructions I guess” – to the point where last time I noticed one it was because I had the shirt on the wrong way around, and it was scratching my front, where there isn’t supposed to be a tag (which suggests I’m just filtering it as opposed to actually not feeling anything).

      (The worst bit is, this was already halfway through the day – even with the tag in the wrong place, it’d taken me that long to notice. In my defense, it’d been a day.)

      That said, I have never understood how people tolerated turtlenecks. Do they really not notice being choked?

      • I think the skin on people’s backs is less sensitive than the skin on the front, but a fast search didn’t turn up anything convenient on the subject.

        I’m another one who’s not crazy about turtlenecks, but I think my dislike of hats (distracting thing pressing on head!) is unusual. I’d much rather wear a scarf or a hood.

      • Tibor says:

        I once wore a T-shirt inside out the whole day without noticing until my then-girlfriend met me at home in the evening and asked me why I was wearing it that way. In my defence, the T-shirt looked kind of good that way too (except for the tag, which it has on the side in the seam).

    • Shion Arita says:

      I’m not autistic and I cut those damn tags off all the time too. It is freaking annoying.

    • Rachael says:

      Is “tags” a US thing? I (UK) call the bits of fabric with washing instructions “labels”, and the bits of cardboard with the price “tags”. Presumably everyone cuts the latter off.

      As a data point, my 6yo who has sensory issues and is being assessed for autism doesn’t mind labels, but my 2yo who seems neurotypical gets me to cut them out.

      • Americans use “tags” to refer to “labels,” and either “tags” or (more commonly) “price tags” to refer to “tags.” Knowing the difference between “tags” and “tags” is purely contextual.

    • Peter says:

      I’ve got an Asperger’s diagnosis, and I don’t have the thing with the tags. In a way this isn’t surprising – people really seem to vary in terms of which traits they have and which ones they don’t.

    • Berna says:

      I’m more or less ‘normal’, but those things bug me no end, I always cut them off.

    • Andrew says:

      Does no one else wear undershirts? I don’t notice the tags on my button-ups or polos because there’s a good layer of cotton between me and it to spread out the contact!

    • LPSP says:

      Honestly it’s extremely rare for me to ever notice a tag. I’m more likely to be bothered by a pinching in the armpits of a poorly tailored/seamed shirt.

    • postpoptart says:

      It can’t be too uncommon. There’s a Hanes/Michael Jordan commercial highlighting the discomfort a tag causes (they’re in a movie theater, Jordan rips it off the guy throws it in the guy’s soda, everyone laughs).

    • Am I the only person who *like* tags? Well, some tags. The majority of tags are unremarkable, it’s just extra fabric. But some tags fall within a special window of “a bit coarse, a bit rigid, but not too much”, and when I find one of those, I just mindlessly fidget with them over and over all day long, rub my fingers on them again and again to feel the texture. This gets to the point where I often chose what clothes to wear based on the knowledge of how “nice” their tag is. I will even sometimes interrupt my going somewhere to fidget with the tags of clothes left hanging to dry.

    • alaska3636 says:

      100% tagless t-shirts. Boom. I almost always where a (tagless) t-shirt under a button-down too – unless it is very hot.

    • IamtheTarpitz says:

      Almost never notice labels at the neck; can’t stand labels at the side. Rip them out because I’m too lazy and disorganised to find scissors.

  4. Adam says:

    “Greatest uncertainty” is a perfect way of putting it. I feel like I can almost never tell what a human really thinks or even what they intend with spoken language much of the time. Conversely, it’s why I like animals so much. I feel like it’s really easy to interpret the facial expressions and vocalizations of other mammals and they don’t use idioms, metaphors, or any other type of figurative language.

    • Deiseach says:

      LRF think that autistic people dislike social interaction because it’s “the greatest uncertainty”; other people are the hardest-to-predict things we encounter.

      What about simply not liking people? I’m asking for… a friend. Yeah. My… friend… says “I don’t know if you like me or not, I don’t care if you like me or not, we have had our requisite interaction for the intended purpose, now please bugger off betake yourself about your business and leave me alone”. That’s not so much “I cannot interpret what is going on and the confusion alarms and distresses me”, it’s “Yeah yeah yeah, got it, now why are you still hanging around annoying me?”

      • Jill says:

        Interesting. Still there may be the question of WHY one does not like people. Or maybe why one does not like casual loosely structured non-goal-oriented communications? Of course, many people in the modern world are this way. Perhaps most Americans are this way nowadays, being cyborgs, attached more to our electronic devices than to other humans. Maybe we’re all becoming autistic, or something like it.

      • Adam says:

        That’s different than me, and it’s obvious from reading your comments that you’re very different from me in general. I don’t dislike people at all. I’m mostly indifferent to other people as long as they aren’t trying to talk to me too much, or they can as long as they don’t expect me to listen or respond.

      • Tom Crispin says:

        There’s a t-shirt for that. *grin*

        I get lots of grins when I wear it

      • Stanley says:

        Perhaps your friend is schizoid.

  5. Toggle says:

    Also, once psychotics develop a delusion, that delusion usually sticks around. It might be that a stray word in a newspaper makes someone think that the FBI is after them, but once they think the FBI is after them, they fit everything into this new paradigm – for example, they might think their psychiatrist is an FBI agent sent to poison them. This sounds a lot like a new, very strong prior! Their doctor presumably isn’t doing much that seems FBI-agent-ish, but because they’re working off a narrative of the FBI coming to get them, they fit everything, including their doctor, into that story. But if psychosis is a case of attenuated priors, why should that be?

    Actually, this seems pretty straightforward to me. The “FBI is watching me” prior is reinforced every time it’s confirmed, or else this wouldn’t be a Bayesian process in the first place. For a healthy person, the first time you hear a lion roar, it’s probably the neighbors watching a movie. Once enough additional sense data also fits in the previously unlikely lion hypothesis, it’s probably a lion. And in (this model of) the brain, ‘strong prior’ probably just means something like ‘more likely to be the model to which bottom-up sense data will snap, given multiple self-consistent interpretations’. So lion-consistent sensory data sets up a feedback loop, where a few lion-consistent experiences make you more likely to interpret marginal noise as lions.

    If your brain is acting weird such that too much sense data constantly ‘snaps’ to something, then the rich models get richer in an unfortunate way; any model capable of halfway-plausibly explaining this incoming junk is going to rapidly become more and more ‘likely’. Since the FBI, space aliens, and so on are all-powerful, a broad spectrum of random noise is explicable in those terms, and so these types of word-models tend to become omnipresent in brains that rapidly ‘confirm’ all random noise in terms of one hypothesis or another. It’s an extreme edge-state of Bayesian updating when p(x) can only go up.

  6. moridinamael says:

    On the level of neurotransmitters and Bayes, I wonder if there’s a distinction between anxiety disorders and delusional paranoia. Both would seem to indicate a tendency to be overly-aggressive in assigning salience to potential threats. Google tells me that with paranoia, you actively believe there is a threat out there, whereas with anxiety, you only strongly believe that something bad will happen in the future, but to me this seems like splitting hairs. Paranoia and anxiety both seem really similar, and both seem based in a tendency to let perceived risk influence priors.

    Like, on some level you’re doing:

    risk = (probability) * (consequence if true)

    and if you’re paranoid/anxious then you’re either overestimating (consequence if true) or (probability). Or from a straight Bayes perspective, which term is being corrupted:

    p(“FBI is after me” | “therapist smiled cryptically”) = p(“therapist smiled cryptically” | “FBI is after me”) * p(“FBI is after me”) / p(“therapist smiled cryptically”)

    I have found it impossible to even communicating to people with anxiety the idea of priors, and the fact that p(“FBI is after me”) should initially be really low, and furthermore that assigning an especially high value to p(“therapist smiled cryptically” | “FBI is after me”) is not reasonable, and finally that p(“therapist smiled cryptically”) is potentially pretty high, and that all of this suggests that p(“FBI is after me” | “therapist smiled cryptically”) should be really low. But I’m not sure what link in that chain is the primary one that is breaking down.

    • Peter says:

      (EDIT: Erm, that was meant to be in reply to the OP, not to your post)

      On psychosis: there’s the “jumping-to-conclusions” (JTC) bias which shows up a lot: there’s a classic “beads task”. To quote some random paper: “Typically, 2 jars of beads are presented to participants, one containing substantially more pink than green beads and the other containing the reverse. One by one, beads are taken from a single hidden jar and presented to the participants who are required to guess from which jar the experimenter is taking beads.” Overconfidence: “Participants with delusions tend to make firm decisions as to the identity of the jar much sooner than controls, occasionally after the very first bead is presented.”

      I think there are at least two senses of “overconfidence” here. Take the case of some guy who is interested in some girl and his friend says “go on, ask her out”, and the guy says “no, she’ll laugh in my face”. The guy may (or may not) be overconfident in his predictions, but he’s definitely not overconfident in his attractiveness, and doesn’t give the general “overconfident” vibe.

      Another paper from a trawl – it looks like people with Social Anxiety Disorder don’t display the JTC bias for the generic beads task, in fact they take longer to draw conclusions than controls do, but when you replace the beads with bits of information in some social scenario, out comes the JTC bias. As if “hot” thoughts and “cold” thoughts go through different reasoning pathways.

    • Garrett says:

      I suspect that paranoia is something like a consistent erroneous belief that *somebody* is out to get you. You just don’t know who. So you’re suspicious of everybody.

      In contrast, I see anxiety showing up in two different forms, both of which are due to sensory overload. You have a person who doesn’t believe that *somebody* is out to get them – they live a normal life. But suddenly somebody punches them! No. Wait. They were just sticking their hand out to shake hands – it’s the new guy from sales. And then somebody shoots at you! No, wait, that was popcorn in the microwave. And it smells delicious. That is, things which are similar to actual signs of impending danger get ramped up and treated as actual signs of impending danger.

      This can also manifest itself with social cues. Because that woman yawning is clearly signifying her lack of disinterest, your beta-ness, your lack of desirability, and general lack of value as a human being. No – she stayed up late watching Matlock re-runs, spent an hour in the gym and hasn’t had her coffee yet. Except that now the whole office is shunning you, completely ignoring your presence. Oh, wait, the new Star Trek trailer is out and everybody is watching it.

      • JZig says:

        As someone who is highly anxious but not paranoid at all, this makes sense to me. It can be hard for the brain to correctly modulate “strong” impulses, so you get that shot of adrenaline from the handshake even though it doesn’t make logical sense. I’m constantly being assaulted by completely inconsequential things that mean nothing.

        But the reason I’m not paranoid is that I can look back at the stimuli afterwards and explain why they’re happening, they reinforce my prior belief of “I am an anxious person who has flawed perceptions” instead of my prior belief that “I am a consistent, wholly rational person so someone must be out to get me”.

        Paranoia seems like what happens when the wrong prior beliefs get fed into, not just having stimuli you can’t handle. PTSD is a lot closer to anxiety than paranoia is in my mind.

      • Equinimity says:

        (Disclaimer on this, I know nothing about neuroscience and could have been completely mislead by the people who’ve told me this. I’ve just had a serious anxiety/panic disorder for a long time so this is relevant to my interests.)

        The sensory overload anxiety fits with both with my own experiences and with some research I’ve heard about. In the late 90’s, someone was researching the amygdalaes response to imagery, and tried putting a small group of anxiety sufferers under the MRI as well. One function of the amygdala appears to be to trigger a response to potentially threatening stimuli, such as sudden noises or threatening looking objects. Anxiety sufferers had a much lower threshold of response than normal people, basically their amygdala was leaking non-threatening input through and triggering unnecessary threat responses higher up.
        Last year I heard about about another bit of research, where someone found a pathway from the frontal lobes to the amygdala, which appeared to damp the amygdalaes response, and in most of the anxiety/panic sufferers they looked at, this pathway appeared to be faulty. So if that’s correct, my amygdala is stuck in threatening environment mode, triggering responses to all sort of harmless stimuli, and my frontal lobes can’t throttle it back when I’m safe.

        It matches my experience with my anxiety, as an example, I shy and start at things other people don’t even notice, rather like a horse. With that constantly repeated threat response, it’s very easy to learn that something harmless is a threat to me, and damned hard to get that out of my head once it’s there. The sudden false positives give my beliefs a sharp push in one direction, and reason is much less effective at pushing them back the other way.

  7. rsaarelm says:

    So, next question: Are there some kind of learnable biofeedback, meditation or self-hypnosis methods to change your AMPA or NMDA signaling intensity?

    • moridinamael says:

      People do report permanent or extraordinarily long lasting effects from single doses of psychotropics. Meditation can also lead to hallucinations and “permanently” altered cognition. I suspect the answer is “yes, we’ll understand this in fifty years.”

  8. baconbacon says:

    How does onset fit into this explanation? Autism makes sense to me, it first presents early in brain development, but schizophrenia? 10 year olds can do some pretty good pattern matching/expectations managing/social interaction tasks, why doesn’t schizophrenia present earlier at least some of the time?

  9. Murphy says:

    Whenever I see the pattern of someone pointing to all the confirmations I get an itchy feeling that screams “confirmation bias”

    The biggest problem with this model that I can think of in a few seconds:there are people with co-morbid Autism and Schizophrenia.

    So what? they’ve got both thin and wide intervals?

    Why would Schizophrenics have “breaks” in their late teens and 20’s and start manifesting symptoms while people with autism just view it as how they grew up?

    It feels like the authors rationalizing to fit the hypothesis so SSC: anyone able to think of ways to falsify the hypothesis or other examples which don’t fit?

    • Daniel Kokotajlo says:


    • FullMeta_Rationalist says:

      Why would autism and schizophrenia conflict? I don’t think schizophrenia was described as a confidence interval thing. The way I interpreted schizophrenia was “apophenia”. And the way I interpreted autism was “pedantry”.

      The model predicts the existence of signals A and B, and that modifying A or B should reliably cause subjects to behave {schizophrenic, autistic, tripping}. It seems like AMPA receptors govern signal A, and NMDA receptors govern signal B. So e.g. if dumping glutamate into AMPA receptors induces schizophrenia, this vindicates model. If such fails to induce schizophrenia, this falsifies the model. In sum, “if we understand it we can control it”.

      What Scott’s doing is categorizing the space of Bayesian failure-modes, sorting actual diagnoses and drugs into the categories, and noticing “the diagnoses and drugs in category X are related to AMPA, what a coincidence. Maybe the causal chain is (drugs -> AMPA -> signal A -> schizophrenia).”

      I dunno about the “age 20” thing, but I don’t think it’s relevant to the model unless it’s related to AMPA receptors (or something analogous to signal A).

    • Ilya Shpitser says:

      Falsifying things is not the Bayesian way!

      As I said on fb, Scott’s thesis should be the other way around: Bayes theorem can be used to represent what the brain is doing, not vice versa. Map/territory.

      • Mr Mind says:

        Falsifying things is not the Bayesian way!

        That’s very true! But something similar is the Bayesian way: come up with data that is very probable in this model and very improbable in every other models… it will give your prior the strongest boost. We need a good name for this!
        Trueifying? Exemplifying? Championifying?

        • Ilya Shpitser says:

          Sorry, confused by this. Picking models based on how probable data is given a particular model is maximum likelihood inference (e.g. not the Bayesian way). Bayesians do not maximize data probability, Bayesians just compute posteriors.

    • met syn an says:


  10. James Babcock says:

    > “But then how come most psychotics don’t seem to be manic – that is, why aren’t they overconfident in anything except their delusions?”

    Because the “confidence” in mania is not of the same type as the confidence discussed here; it’s a consequence of altering the tempo of thoughts in such a way that doubts don’t have time to register properly.

    • Jill says:

      Good point. There are numerous effects of speeding thoughts and that is likely one of them. For a lot of things manics can’t or don’t do, the explanation why is that they can’t slow down their thinking and/or behavior enough to do them. Certain tasks have to be done slowly, or they can’t be done at all.

      One of the benefits of meditation is that, if you can do it, you can learn to do some new and useful things with your mind that have to be done slowly– abilities where the pace of modern life prevents most people from ever learning to do them.

  11. Neanderthal From Mordor says:

    Why do we enjoy consuming psychoactive substances anyway? Giving yourself a bit of temporary schizophrenia doesn’t sound like fun.

    • I don’t even know why people like optical illusions. Or magic tricks.

      • LPSP says:

        There’s a challenging in finding the illusion, or a lesson to be learnt in understanding how one picture have can several perceptions.

    • Old Lamps says:

      It’s a nice place to visit, but you wouldn’t want to live there.

    • drethelin says:

      Framing it as temporary schizophrenia makes it sound bad because schizophrenia is a word we associate with bad things. Try instead to think of it as conscious dreaming, euphoric visions, or making the world feel like it makes sense.

    • Jill says:

      And then there’s of course the possibility that psychoactive substances are actually not very much like temporary schizophrenia at all. To really know, we’d need to talk to someone who’s experienced both.

      • Riothamus says:

        And who is lucid enough to explain things in terms an unstoned neurotypical could understand.

        One hard hat-trick to pull off, seems to me.

    • Jill says:

      Perhaps we know so little at this time about what is actually going on with either psychoactive substances OR schizophrenia, that they may have numerous important differences that we are not yet aware of.

    • Bugmaster says:

      I might have mentioned that before, but I know at least two people who feel exactly the same way about movies and fiction literature.

    • wysinwyg says:

      After reading the OP, I suspect I was unconsciously self-medicating for mild ASD.

  12. Timothy Johnson says:

    “Also, once psychotics develop a delusion, that delusion usually sticks around.”

    I think this happens because people (and not just psychotics) are not perfect Bayesians. We usually reason for only two for three steps.

    You can measure this by asking a group of people to pick numbers from 1 – 100, where the winner is the person closest to 2/3 of the average. The winning guess is usually around 20:

    As a result, any time someone develops a narrative, it’s very hard for them to retrace and reconsider the events that caused them to believe in that narrative. They likely have forgotten much of the initial evidence, and are simply left with a strong impression that their narrative is the complete truth.

  13. PSJ says:

    Dopaminergic reinforcement learning =/= Bayes!!!

    Dopamine is generally believed to code a point estimate of error (because how can magnitude of dopamine release code for a more complex distribution) which means thinking of it as a “confidence interval” is going to be misleading at best. See one discussion of RL and Bayes in comparison here:

    I guess this is a bit of a nitpick, but I think that over-Bayesing things will lead to worse understanding of both Bayes and those things.

    • Eli says:

      It’s actually a matter of open dispute in neuroscience whether the human brain runs on some combination of model-free and model-based reinforcement-learning controllers, a Bayesian reinforcement learning algorithm designed to reuse the models produced by causal learning, or nothing but Fristonian active inference at various scales/levels of hierarchy.

      The sheer amount of work done in machine learning and theoretical neuroscience has given us an incredible wealth of untested theories which now require better experimental apparatus to be tested.

  14. illocution says:

    Andy Clark’s recent Surfing Uncertainty also has a number of similar linkages between consciousness, prediction, embodiment, and Bayes. He also goes into much more depth on the Bayesian linkages to schizophrenia and other types of delusions and “biases”, as well as doing a very good job of explaining how optical illusion surprisal may be related to direct Bayesian modelling in the brain.

    If you liked this post, you may find it quite interesting.

  15. Cerastes says:

    Quick thought on Amphetamines – could the blurry vision be due to non-neurological causes? Specifically, if the amphetamines increase blood pressure, that can lead to blurry vision for purely physiological reasons related to the amount of fluid in the vitreous humor. So the computational side may still work as hypothesized, but the quirks of the human body cause the physical sensor to mess up.

    • Jill says:

      “the quirks of the human body cause the physical sensor to mess up.”

      I think this is true in numerous situations like this.

    • Who wouldn't want to be anonymous says:

      That was my first thought. My second was what if they just don’t have very good eye sight and the blah blah blah makes them more accurately perceive their poor vision.

  16. Moorlock says:

    Re: “the entire world collapses into a single narrative, a divine grand plan in which everything makes sense and you understand all of it”

    FWIW, nitrous oxide intoxication has this effect in spades, and you might try to squeeze it into your model somewhere. I’m not qualified to say anything about how N2O works, but the google stumbled me on to this statement: “N2O inhibits NMDA receptors in the brain whilst simultaneously encouraging the stimulation of the parasympathetic GABA receptors.”

    See William James’s brief and amusing notes on the subject here:

    • Brin says:

      FWIW, nitrous oxide intoxication has this effect in spades, and you might try to squeeze it into your model somewhere.

      …hmm, I wonder if that does tie in. See, I’ve noticed that the stories I’ve heard from people describing what being on nitrous was like for them (take with salt, plural of anecdote is not data, etc) seem to fall into two basic categories, and AFAICT (again, salt) the determining factor is autistic vs non-autistic.

      Non-autistics describe experiences much more like the archetypal hallucinogen: their stories are similar to the stories of people who took LSD and the like, including that sensation of understanding how things work.

      Autistics, on the other hand, remain lucid while on nitrous oxide. While they do hallucinate, they can tell which of their perceptions are hallucinations and which are real. They don’t babble aloud about wild theories, and they tend not to babble about wild theories in their heads either.

      (Judging from my childhood experiences with dental anesthetic, I respond to nitrous in the autistic manner, and I’m glad of this. Being high can be fun–I found the loss of proprioception to result in a pleasant feeling of incorporeality*, and it was interesting to poke at the drug’s effects and figure out what was offline and what was altered and how it was altered–but I don’t want to be non-lucid.)

      *”Incorporeality” is the word I used at the time, and I remembering feeling rather patronised when the pediatric dental assistant asked after turning the gas off if I was still feeling “floaty”, although in hindsight I suppose it’s not too bad a word. (Also, before you ask, I’m not dysphoric. It’s a nice body, but being a disembodied spirit can be nice too.)

  17. Pattern-matching Failure says:

    > (Teh fcat taht you can siltl raed wrods wtih all the itroneir ltretrs rgraneanrd is ahonter empxlae of top-dwon pssirocneg mkinag nsioy btotom-up dtaa sanp itno pacle)

    This might be the hardest such sentence I have ever read. Because of the earlier mention of the word cat, I first misinterpreted “Teh fcat” as “The cat” rather than “The fact”. It also took a while to figure out that “itroneir” meant “interior”, and I only got that “pssirocneg” meant “processing” after I read the entire article and went back to it, at which point I had become used to the word “processing” coming after “top-down”.

    • Subbak says:

      I had to use an anagram finder for “itroneir”. I was expecting “inner” or “internal” and somehow I couldn’t connect the dots to “interior”.

      Also, I will point out that “The” -> “Teh” does not match the rule of only reordering interior letters.

      • FullMeta_Rationalist says:

        My local newspaper includes an anagram puzzle. It usually takes me between 30 and 60 minutes to complete. The words are always short and quotidian. I wonder what par is.

  18. Steven says:

    I think one (or several, there are many) of the traditional models of enlightenment seems really really close to just being schizophrenia but with positive affect instead of negative.

    • Jill says:

      Interesting. I’d love to talk to someone who’s experienced both, or read an article by them. Maybe someone who had a schizophrenia-like state as a reaction to some prescription drug, and then experienced “enlightenment” later and could describe the differences, if any.

    • JohnReese says:

      Well, many mystical works also paint a picture not unlike the florid symptomatology of schizophrenia. Methinks the quest for enlightenment comes from , at least in part, a need for extreme data compression. To unpack that, if we are constantly matching our models of the world to sensory data, and develop more and more general world-models as we age and accumulate “experience” (first-hand, second-hand and “wisdom” which is merely someone else’s verbal abstraction of robust predictions from their generative model), some of us (well, many of us) get this notion into our heads that there must be an ultimate generative model, one that affords us maximal data compression, such that its attainment would be a personal singularity of sorts. It is possible that there is no ground truth to give such an ultimate model adequate confirmation, and I personally have never encountered any school of enlightenment-thought that did not invite oneself to indulge in confirmation bias like nobody’s business (to all those who are likely to bring up the case of Buddhism and its many schools…well, if you have noble truths for your priors, you are going to be susceptible to confirmation bias along the way given how the brain handles propositions about the world…a true seeker of truth would probably hold only tentative truths in mind, and go MaxEntropy for most beliefs ;)).
      Perhaps the reason the “madness” of revelation or enlightenment feels good, is due to the fact that one has a sense of having attained something , usually through effort/will/practice, as opposed to the universe being out to get you?? It is perhaps no coincidence that tribal societies with shamanic traditions had people who would “guide” younger members going through what we would term acute mental illness, and these shamans themselves seem to be people who, either by insight or through guidance, turned illness into enlightenment? How you react to this thread would depend on your priors of course; I digress.
      This is all very interesting, and I am very glad to see Scott’s take on the use of Bayesian approaches to understand brain function as well as dysfunction.
      On the technical side of things, I do not think there is a great gulf between model-based reinforcement learning, hierarchical predictive coding, or the variational schemes proposed by Friston..likely the last named is a high-level parsing of what goes on in the brain (which is likely a bit of active inference to construct models of world-states in conjunction with reward-driven policy-learning and optimisation).

  19. Jill says:

    Some of our thinking that X is like Y– e.g. the psychedelic experience is like schizophrenia, or being sensitive to a tag on my shirt is like experiencing autism– may be due to our current beginning state of understanding in neurotransmitters, neurophysiology etc. Fish “flying through the air” when they jump above the water, may seem very much like birds flying, before you understand well what fish and bird bodies are actually doing at those times. The same may be true of the insides of the bodies of fish, birds, and humans.

    Gotta do more research. And also interview people who have had both experiences you are comparing, and who perhaps can describe the similarities and differences in their own particular experiences. Also, high functioning autistics and high functioning schizophrenics should be interviewed more often than they are, to find out what experiences they have. Even low functioning schizophrenics sometimes make sense when speaking, so they should be interviewed too, perhaps using simpler language. Also we need more research and interviews on use of psychedelics.

    That would be preliminary research, of course. But very important. Because a lot of research fails because it never thinks to ask the best questions, because many researchers only read journal articles about research, and thus have limited first hand experience with the phenomena they are researching. Perhaps there need to be a few extroverted researchers here and there, who would enjoy doing interviews with people who have autism, schizophrenia, fibromyalgia etc., asking about their experiences.

    There is not anywhere near enough interview type preliminary research out there– perhaps because there are not many extroverted researchers out there. Maybe scientists could hire some of the investigative journalists who are unemployed, now that there is hardly any news any more in this day and age, but only entertainment. They could do some of the interviews, if not enough extroverted researchers could be found.

  20. Too Loud, Too Bright, Too Fast, Too Tight: What to Do If You Are Sensory Defensive in an Overstimulating World has it that difficulty (sometimes at the level of being debilitating) with sensory experiences that most people can handle isn’t specifically a result of autism, though I think it’s more common for autistic people.

    Sensory problems are frequently misdiagnosed as autism, depression, or neuroticism, and possibly some other things I don’t remember.

    This is the only book I know of where the title makes some people feel better.

  21. Marvy says:

    You never did mention what LRF stands for. (Also: stimming. Which is such a rare word that my spell checker flags it. But at least I can look that up.)

    • Callum G says:

      LRF refers to the aberrant precision account of autism study. It’s the first initial of the last names of each author involved.

  22. Oleg S. says:

    1. I wonder, how depression fits into this picture, with SSRIs (=increased serotonin signaling) and ketamine (NDMA antagonist + AMPA agonist via its metabolite) being used as drugs. Is depression caused by too much top-down procession, like anti-schizophrenia?

    2. From the “narrow confidence interval” hypothesis, autism should be treated by amphetamine or other dopamine positive modulators. Which is far from what is prescribed, I presume.

    3. Prepulse inhibition by ketamine finally makes sense to me (decreased top-down processing -> inability to predict next pulse from prepulse -> constant surprize); but prepulse inhibition by dopamine agonists stops making any sense.

  23. Gary says:

    I’m reading “Philsophy Between the Lines” and I’m realizing that some of my paranoid fears during social interaction aren’t necessarily unfounded. Esoteric speech apparently is pretty common in most cultures. How does one know for sure he isn’t being gas lighted by a crew of organized gang stalkers or if he is schizophrenic?

    • Enkidum says:

      “How does one know for sure he isn’t being gas lighted by a crew of organized gang stalkers or if he is schizophrenic?”

      Because it is very unlikely that anyone cares enough about you (or me, or any other random person) to bother going to the effort to gaslight you in this way?

  24. Besserwisser says:

    Wait wait wait wait, there’s a psychological test where you’re shown a photo and if you’re perception coincides with the reality of the picture, you’re the crazy one? I can’t describe to you how deeply this concerns me. Granted, this is similar to that psychopath (sociopath?) test about the funeral and killing parents where the “correct” answer is also a diagnosis. But still.

    Weirdly enough, this reminds me of Nietzsche and his idea of a super-human who is above conventional morals. I didn’t give much about this idea, but now I’m much more sympathetic towards it since so much what we consider wrong seems to be an accurate representation of reality.

    • Riothamus says:

      I often consider that what makes a person insane is not what their perceptions actually are, but whether they recognize the relationship between their perceptions and other people’s.

      It isn’t seeing the convex or the concave face that makes a person mad, but not understanding that other people see it differently.

    • Enkidum says:

      “Wait wait wait wait, there’s a psychological test where you’re shown a photo and if you’re perception coincides with the reality of the picture, you’re the crazy one?”

      No. This isn’t a test of whether you’re crazy. It’s just that people with certain mental illnesses are more likely to perceive what is actually in front of them, in this one specific case. It’s certainly not true that they are more likely to perceive what is actually in front of them in general, quite the opposite.

      The brain is required to develop a representation of the world in realtime from incredibly imperfect evidence. In order to do this, we have to use shortcuts, generally known in perceptual psychology as heuristics. Things like “assume that light is shining from above, because that’s where the sun is,” or “assume that seats are roughly knee high, because otherwise they’d be useless.” But in the rare circumstances where some clever bastard shines light from underneath, or builds a giant chair, our heuristics can lead us astray. They work most of the time, because otherwise we’d keep walking off cliffs or into walls. But under certain specific cases, they don’t.

      Here’s a really simple example. Have you ever looked at a picture of something in front of something else? I don’t know, let’s say a picture of a house with a tree in front of it. You very clearly see the spatial relationship, right? The tree is in front of the house. But you’re totally wrong! There is no depth at all, it’s a flat picture. The artist is taking advantage of certain heuristics and fooling you into seeing depth where there is none. If it wasn’t for your silly brain making these kind of assumptions, you’d be able to see the picture “more accurately”, as a bunch of blobs of paint with no spatial relationships other than in 2D. (And I believe there are certain mental illnesses for which this is one of the symptoms, but I can’t remember off the top of my head.) And if you didn’t make these assumptions, then seeing actual depth in the real world would be a lot harder.

      Think of perception as something like a (very) informed guess about the world. Informed guesses can sometimes be wrong. If you know the shortcuts that people are using to make these guesses, you can seriously mess with them, which is what all illusions are (including pictures). But if you don’t make these guesses, you’re stuck in limbo, unable to do anything for lack of perfect information.

  25. Anonymous says:

    >Amphetamine users report increased visual acuity (weirdly, they also report blurred vision sometimes; I don’t understand exactly what’s going on here).

    Pretty sure this is due to increased blood pressure, but could also be from dry eyes.

    >But people don’t generally think of mania as primarily dopaminergic. Why not?)

    They don’t? I’m not a doctor but I’ve always associated them. Maybe because of the stimulant connection?

  26. Puerto Rico Rolf says:

    This is an interesting field of research. There are different version of these Bayesian brain ideas circulating in the (theoretical) neurosciences with Friston’s free-energy (FE) principle being the most notorious. It basically assumes that this mechanism of an internal (Bayesian) world model that is permanently updated in case of prediction errors is basically implemented in one way or another everywhere in the brain and thus is able to explain virtually every brain function on every level. An interesting facet of FE is how it not only tries to explain how the brain manages to make sense of the world (perception) but also how we act in the world by unifying both problems under the common goal of minimizing prediction errors (or minimizing surprise). The perception side of this has been described above: Everything is fine as long as the sensory input matches its prediction. In case of an prediction error, the (top-down) model needs to be updated. But actions on the other hand are thought to follow the same goal: If there is an prediction error, instead of updating my model, I can also act to change my environment back to a predictable state. For instance, I may remove the tag from my shirt to restore sensory order or I may put my toys in the right order or I may chose a life partner that has exactly the good and bad qualities I grew up to expect life partners to have, or I vote for the presidential candidate that turns the world into the dangerous place I internally expect it to be anyway, or whatever…

    The crazy thing about FE is how there is no limit to the phenomena that people have tried to explain as instances of FE minimization. It actually starts with the origins of life itself. Maybe this gives you a taste of the grandiosity involved:

    The free-energy principle is a simple postulate with complicated implications. It says that any adaptive change in the brain will minimize free-energy. This minimisation could be over evolutionary time (during natural selection) or milliseconds (during perceptual synthesis). In fact, the principle applies to any biological system that resists a tendency to disorder; from single-cell organisms to social networks. The free-energy principle is an attempt to explain the structure and function of the brain, starting from the very fact that we exist […]. If one looks at the brain as implementing this scheme […], nearly every aspect of its anatomy and physiology starts to make sense.

    One criticism against FE is that it is much too general, making it not falsifiable and thus not qualifying as a proper scientific theory.

    But still, it’s an interesting line of thought, I think. One example I particularly like is how mindfulness meditation was described as a limit case in the FE framework: You limit yourself to just observing, not acting, and by that enforcing that all prediction errors have to be resolved completely internally, i.e., by updating your world model. Which somehow gives the Vipassana aim of “seeing the world as it really is” an interesting new twist.

    If you want to read more, I think this blog post is a quite good start for this whole topic:

    • Eli says:

      One criticism against FE is that it is much too general, making it not falsifiable and thus not qualifying as a proper scientific theory.

      I think the interesting question about the free-energy principle is: what actually counts as free-energy (that is, as a prediction error?)? That is, what in particular is being predicted? Friston always attempts to rebut the Dark Room Objection, to give a specific example, by pointing out that Dark Rooms might have a low H(s) (unsurprising sense-data), but may have a very high H(s|m) (surprisal of the sense-data given the particular model m constituted by some agent, like a real human being). Of course, since he doesn’t claim to measure P(s|m) or H(s|m), this means that via the particular expectancies he claims real people have, he’s sneaking in his own pre-theoretic preconceptions and real-world experiences regarding how people actually are, rather than showing that his model precisely necessitates how people actually are.

      I would say that free-energy minimization does for real, embodied, physically-implementable agents something like what Solomonoff Induction does for Cartesian theories of AI: it provides a deliberately over-general, but nonetheless informative, framework for considering an entire class of agents. We can talk about free-energy minimizers as a mathematical idealization of biological brains – a testable one, even – but the theory needs a lot of filling-in to have the empirical content we really care about.

      (For instance, Jacob Hohwy’s and Andy Clark’s books on the predictive processing hypothesis actually disagree quite a lot on perceptual directness versus indirectness and the embodied-ness of cognition. I’m not entirely sure the debates continue to make sense in light of the facts about perception, cognition, and action being necessarily statistical functionalities, but I’m also not sure that philosophers-of-mind actually accept any such necessity.

      There’s also the question of whether the brain is only predictive, even in action, or if it also possesses one-or-more reinforcement learning systems so that evolution can train action with a clearer, more precise supervision signal.)

      That said, it is a testable theory – via behavioral experiments, neuroanatomical and electrophysiological studies, and computer simulations – and it has been successful in many of the precise predictions it does manage to make.

  27. flyfly says:

    Doesn’t this make an awful case _against_ relying on Bayes Theorem as a logic tool?

    You’d want your logic tools to not have the same problems as your brain potentially has.

    And then you abuse Bayes just a little, and get your priors a little wrong and get the idea that the Multiverse is real. Which sounds an awful lot like the kind of hallucinations you’d hear from drug addicts.

    • Eli says:

      Well yes, actually! And that’s entirely correct: without restricting yourself to the class of models we generally call “scientific”, Bayesian reasoning works atrociously for extrapolating beyond the properties of the dataset to ontic statements about the world.

      It works just atrociously enough to actually work, if you get what I mean.

      • Mr Mind says:

        I don’t get it. What are the models for which Bayesian reasoning work atrociously?

        • Eli says:

          They work “atrociously” in the sense that they output whatever parameters/solutions are optimal for the evidence given and the prior assumptions, rather than necessarily recovering the ground-truth within some bounded amount of samples or computation (as classical statistics likes to do).

    • FullMeta_Rationalist says:

      I think it’s just a sign we have poor priors. When I look at the night sky, I feel terrified. Because I’m reminded that the universe is UNFATHOMABLY HUMONGOUS and my entire life experience is an infinitesimally minute fraction of it.

      For example. Consider that we only see three primary colors. Supposedly, a Mantis Shrimp can see twelve. Twelve. Think of the rainbows we’re missing! Also, SSC was the first time I’d heard of SJW’s. I’ve never met one in person. They could be fictional for all I know. I will leave it as an exercise to the reader to construct further examples.

      Sure, we probably have reasonable priors for things we care about and experience on a daily basis, like “stoves are dangerous… probably”. But when we philosophize about multiverses and Time Cubes, of course we get weird answers. We have no priors for them. I.e. how many multiverses have you visited?

      • Eli says:

        We have priors about a lot of things, but:

        1) They’re often very imprecise priors. This is a big point that the Bayesian Brain theorists have for us: if your mental models are so imprecise that their probability densities overlap heavily in the space of stimuli, then a large measure of stimuli can’t be predicted better than random, so you will always have a high prediction error. We would colloquially say in that circumstance that you “Don’t know anything” about it.

        2) The hierarchical Bayesian map still is not the territory. We really probably have at least two or three “maps”: a perceptual map, a goal map, and an action map. The perceptual map involves fewer variables, and especially doesn’t involve actions, so it by necessity oversimplifies the world (just to avoid wasting information given to the senses). For example, one way that I can tell an object in front of me from a clever trick done with mirrors is to try to grasp it with my hand: if I can grasp it, it’s real, but if my hand goes through, it’s a trick with mirrors. If I can’t reach out my hand, but can instead only engage visual saccades (eye movements), then both “trick with mirrors” and “real object” hypotheses have to get collapsed into “real object”, which absorbs all the probability mass.

        3) Very distal causes (causes “far-away” from my observations in time and space) don’t put much precise information into my observations, so they’re hard to detect. The particular arrangement of fine-tuned physical constants during the Big Bang might have exactly allowed for me, and not someone else, to be born, but the other causal factors between me and those constants are so noisy that I can’t really detect that quiet and subtle a signal.

        A major, major point in the literature is that many illusions cannot, in principle be resolved via perception alone, but can be resolved (the hypotheses differentiated in the evidence they yield) by very simple reality-testing actions. This mirrors philosophy-of-science in theoretical neuroscience and vice-versa: it’s hard to justify believing in things that cannot, in principle, be acted-upon.

    • Enkidum says:

      “You’d want your logic tools to not have the same problems as your brain potentially has.”

      Well, the basic problems are limited processing time and imperfect evidence. Using Bayesian reasoning is essentially a way to deal with that. Neither problem is ever going to be solved, assuming no God. So what’s the alternative?

      • wysinwyg says:

        Divination has satisfied this need for thousands of years now, but I suspect you wouldn’t deem it acceptable.

        • Enkidum says:

          What need is that? Genuinely not sure what you mean, although you’re probably right that whatever you think divination does, I don’t.

  28. Callum G says:

    It would be interesting to examine eating disorders in this aberrant precision context. There seems to be some sort of link between autism and eating disorders, at least in terms of social profile and behavioral rigidity, so there could be similarities in the cognitive handshake failure.

    Eyeballing the studies, it would make sense that people with anorexia nervosa also suffer from the same narrow confidence intervals for sensory data as people with autism do, although a lot more topic specific. Hyper-awareness of food, weight and related things could be a predictive failure similar to an ever present scratchy shirt. People with anorexia have been found to be overly reactive to environmental stimulus and a small study shows that misophonia, the hatred of sound, is more prevalent in people with anorexia than the general population.

    I’m not a doctor, and there’s a likelihood that I’m missing something here. How about the self-esteem effects? If topic specific predictive failures are possible then why food and not… cars? I still think it’s interesting though. Any thoughts?

    • Deiseach says:

      I think what we’re learning is that “Holy crap, brains are complicated (and so are the rest of the bits’n’bobs)”.

      It does make me rather more sceptical of the “soon we’ll have great scans of brains and know precisely how they work and then we can recreate human minds in silicon forms” school of thought.

  29. Jill says:

    New article on people tripping on ayahuasca, in the New Yorker. Apparently becoming increasingly popular.

    How ayahuasca, an ancient Amazonian hallucinogenic brew, became the latest trend in Brooklyn and Silicon Valley.

  30. Anaxagoras says:

    Regarding the theory about autism as handshake failures — if this theory is correct, wouldn’t we expect tricks like “the the” to work less well on people with autism?

    Also, the standard meditation thing of being aware of all your surroundings, all the sensations you would normally ignore — that sounds sort of similar as well.

    • fermatastheorem says:

      wouldn’t we expect tricks like “the the” to work less well on people with autism?

      That matches my experience and observation.

      re: meditation – this post by Benjamin Hoffman seems to report some interesting effects on sensory processing after a meditation retreat.

      • Anaxagoras says:

        As someone somewhere on the spectrum, I don’t find myself particularly resistant to falling for the “the the” trick, nor similar things reliant on the brain glossing over incongruities. For what it’s worth, I’ve not got much in the way of sensory issues, and I don’t much mind typos in other people’s writing, but I am bothered by them when they appear in my own.

    • Scott Alexander says:

      There are certain sensory tests where people with autism perform differently. I’m not sure if the very simple examples of top-down processing I gave are among them. I might test this on the next SSC Survey.

      • E.P. says:

        Missed opportunity, I fully expected: “I’m not sure if the the very simple examples…”

      • DefectiveAlgorithm says:

        Single counterexample (read into it what you will): I’m diagnosed with asperger’s syndrome, and was affected so strongly by the ‘the the’ thing that I had to use ctrl-f to find the three times you’d done it, as even when I was consciously searching for the duplicated ‘the’s I kept filtering them out (though I did manage to find two of them before I got annoyed and just did a search).

        EDIT: Though it occurs to me that you might be referring to more severe cases of autism as opposed to the high functioning variety.

        • AnonEmous says:

          Same. I think it’s because, being overwhelmed by bottom data, I over-rely on top down models. Then again, I’m really really good with top down models.

  31. B says:

    “[…]an autist wearing uncomfortable clothes might be constantly focusing on their itchiness rather than on whatever she’s trying to do at the time, but she’s not going to start thinking they’re a sign from God.”

    Correction: she’s not neccessarily going to start thinking that they’re a sign from God.

    That could be a quick and dirty test of the model. Is schizophrenia more common among autists than among neurotypicals?

    • Besserwisser says:

      As far as I’ve heard, other psychological conditions are more common among autists. I don’t know if schizophrenia is especially common, compared to other things among autists, other crazy people or normies.

    • Scott Alexander says:

      It’s really hard to test for schizophrenia among autistic people because all of these conditions are based on vague collections of symptoms, and a lot of the symptoms of autism are similar to the symptoms of schizophrenia. Also, all of these conditions are related to damage to the brain, and people whose brains have been damaged in some ways seem more likely to be damaged in others.

  32. FullMeta_Rationalist says:


    I’ve been diagnosed with mild Asperger’s (I guess the DSM calls it Autism these days). And none of the explanations I’ve read so far have hit the mark exactly. But the Aberrant Precision narrative is the first one that feels generally correct.

    My most defining personality trait is that I feel an intrinsic need to understand every conceptual detail. Otherwise I have difficulty getting shit done. Also, I have a low-key yet perpetual fear that I’m participating in something akin to Mao’s Great Leap Forward without even realizing. Which often leads to paralysis unless I can prove to my satisfaction that a course of action is airtight.

    Also, I can’t stand interruptions that divert my attention. Internally, I describe my mind as having extraordinary inertia. It takes immense effort to get going. And once it’s chugging along, it takes immense effort to stop or change direction. So when an interruption demands my attention, I feel frustrated that I have to toss away all the momentum I’ve built up. Emotionally, it feels like I’d spent the past hour building a house of cards and a draft knocked it down.

    The interruption tick may not seem relevant at first. But it makes more sense now, given the Aberrant Precision framework. Because my Modus Operandi is to tap the Flow. Which is when I get into this “laser focus mode” and forget my immediate environment. And I feel sorta annoyed when something unexpectedly pulls me out.

    I dislike when others change plans without notice. I usually consider myself a go with the flow kind of guy. But if a set plan has already been negotiated, and said plan is changed but nobody tells me, consider my jimmies rustled.

    Yes I stim. Only a little bit. Usually fingers, toes, or a foreleg. I’m doing it even now. I think neuro-typicals often do this too though. So I don’t put a lot of significance in it.

    Yes I need blankets when I sleep. Even if I feel hot. I feel awkward otherwise. Never thought of the weights trick. Imma try that. +1

    Yes I avoid social-interaction because it’s a hassle. Supposedly I’m not that bad at it when I try. But it’s rarely worth the effort imo. Hell isn’t other people. Hell is the Other.

    No, shirt tags don’t really bother me.

    Also, my first musical love was the House genre. I’m a sucker for Four on the Floor beats. Which is about as repetitive as it gets. (c.f. Daft Punk: Super Heroes; Around the World)

    • LPSP says:

      I’d definitely concur with the inertia description. I used to relate myself as being akin to a train (cue much groaning) in a world of RC cars – difficult to start and difficult to stop, and not too keen on travelling along a dodgy set of rails. And once I’m halfway along a journey, having expent time and fuel and organisation to put the trip into motion, the idea of stopping and returning is akin to… leaving a wonderful dish of food to rot outside for a day, or slashing months of built up interest/loyalty membership for something stupid and petty.

      When I’m responsible for the upset, I feel terrible and resolve deeply to account for it and never let it buck the applecart again. When another is responsible, I feel a great sense of punitive indignation, which is probably the only form of social difficulty I experience – not that it’s always a problem, but some people don’t like a bombastic chewing-out.

    • wysinwyg says:

      I’m not diagnosed with anything, and I probably don’t feel any of this quite as prevalently as you seem to based on your descriptions, but this all sounds very similar to my perceptions about myself. I also felt the aberrant precision idea seems to make a lot of sense applied to my ASD-type behaviors and inclinations.

    • Tracy W says:

      but if a set plan has already been negotiated, and said plan is changed but nobody tells me, consider my jimmies rustled

      Ah, is there anyone who likes this one?

      • Deiseach says:

        Ah, is there anyone who likes this one?

        Yeah, but it’s the difference between “Ah crap, this is mildly inconvenient and I’m a bit annoyed but I’ll adjust” and “YOU CHANGED THE PLAN WELL NOW I’M NOT GOING”.

        I tend towards the second rather than the first – I hate changes like that sprung on me. (Not officially diagnosed as anything so I can’t say “As someone on the autism scale, I find that …”)

        • FullMeta_Rationalist says:

          imo the diagnosis is a joke anyway. The symptoms aren’t a sign of autism, the symptoms are the autism. Just like with depression, fevers, etc.

          Your head feels hot. Do you have a fever? “Uh, idk man. I haven’t been diagnosed.”

          I get the vague feeling that the diagnosis (like insanity) is actually a legal artifact. Scott’s posts reinforce this.

          • LPSP says:

            Pretty strongly concur. I’ve used the analogy of hyperactivity vs ADHD to describe the relationship between autistic behaviours/symptoms and autism itself; anyone can behave hyperactively, but some are more hyperactive than others and we give them a label. Same for behaving in a sensitive-exacting/pedantic manner.

  33. Anthus says:

    ah, something I feel like chiming in on.

    My central point is that there will still need to be some layer of meaning outside of or on top of statements or assertions; some nature of truth or falseness, and that the nature of communication about true and false things is not exactly a simple matter of probabilities.

    Consider a person contemplating reality and God who thinks about their creative ability to push thoughts into the stream of their perception. There are a range of reactions, from some solipsistic powerless maybe, to “I’m co-creating the universe” (what I consider ‘healthy’) to “I MUST LITERALLY BE GOD” depending on how much force you’re willing to exert on that stream.

    I didn’t go down that last route myself, thankfully (or quite likely I wouldn’t be here), but I think I understand the principle: if you imagine updating certain critical priors at the same time you’re exalting in your personal creativity you end up doing what in my personal lexicon is a ‘hard write’ to the priors… and it makes revising those priors somewhat difficult.

    If you’re God how could you be altered further?

    It’s a hideous loop, and your unit goes in the scrap heap.

    > Probably this will end the same way most things in psychiatry end – hopelessly bogged down in complexity.

    Which might be a point about authoritative sources and our reluctance to commit to updating our societal priors. Let’s take a journey with our poor would-be Jesus as ‘He’ tries to become an authoritative source. He submits updates to other people’s priors that fail re-uptake. Eventually this person acts in some manner that gets the attention of… well, something, which decides this person should be Not In Public.

    And they go Away.

    How do updates get moved to what ‘everyone’ thinks?

    So back to my central point about meaning or truth or falsehood or whichever. Bayesian mechanics and knowledge of it is not the same as knowledge of your personal authority, or when your communication with a person of authority is threatening to conflict with yours, or in particularwhy it is in conflict. You can imagine constructing a bunch of Bayesian statements within or about the conflict but that doesn’t indicate understanding of the conflict: it just means something like… you know that a table can support weight without knowing why people build tables.

  34. TomA says:

    Do autistic people ask to be treated or cured? Or does society (acting through it’s institutions) proactively intervene? When does deviation from the “norm” become psychosis that warrants intrusion? I’m not asking about the extreme cases with acute peril for self or others, but deviation in the first or second sigma from the mean.

    In a sense, I’m asking if abnormality in the “handshaking” mechanism is a feature or a bug? Are these individuals malfunctioning or hyper-functioning? What is the ideal; maximum functionality or maximum social integration?

    • FullMeta_Rationalist says:

      Personally, I’d prefer to be left to my own devices. It’s NT’s who are weird as far as I’m concerned. I get the impression other aspies prefer likewise, unless they have a specific problem.

      The question “feature or bug” reduces to “which is more fit, X or Y”. It depends on the environment. Floating Points impart greater precision than Ints, but they cost more bits. So there’s a tradeoff. In other words, Algernon’s Law.

    • LPSP says:

      I think you’re asking many very different questions at once and assuming the answers will resemble each other. To stick to the first one, I’m sure any sort of autistic prone to reflection would be interested in something that let them function as smoothly in society as neurotypicals seem to do – I am assuming anyone autistic enough to recieve a diagnosis must incur some friction in order for this to be the case – but wouldn’t want to change their personality any more than anyone would.

    • Bugmaster says:

      I’m not autistic (as far as I know, shirt tags notwithstanding, terms and conditions may apply), but if someone offered me the chance to install a broadband communication system in my brain, I’d take it. Of course, I’d check the risks involved before doing so; but if the risk was acceptable — yeah, sign me up. Especially if the vast majority of people already had the upgrade installed at birth, and I was one of the unlucky few who had not.

  35. JZig says:

    > But if psychosis is a case of attenuated priors, why should that be?

    This paragraph seems to assume that psychotics are always in a state of attenuated priors, but does that actually match evidence? I don’t have much personal experience with psychotics, but my impression is that they tend to alternate between periods of lucidity and delusions. If there’s some sort of feedback system running with the NMDA where it fluctuates heavily over time, that could lead to a situation where some priors are over-emphasized while others are incorrectly ignored. In theory it could be a mood-based thing and might be testable by checking levels over time. Or, maybe there’s something where specific contextual thoughts and environments make you more prone to incorrectly prioritizing stimuli. Does anyone know if it’s possible to measure receptor activity over time in the way you would need to see if things are fluctuating weirdly?

    From my own personal experience, I definitely have trouble with modulating stimuli and will fluctuate wildly between oversensitivity and undersensitivity without being able to control it, but I don’t really have many non-perceptual symptoms of autism and actively dislike repetition. The LRF hypothesis matches my own personal experiences well, it often feels to me like my error correction is off. As for social situations, I have trouble with them because the stimuli of looking someone in the eye is REALLY REALLY OVERWHELMING. I have no problem with reading social cues in a neutral environment, but as soon as I realize that I’m looking directly at someone else’s eyes I shut down my signal processing to stop myself from breaking down (which leads to the undersensitivity). When I’m drunk my sensitivity seems to be lowered enough that I can be perfectly normal in social situations.

  36. Snahgle says:

    I went looking for the base rates of blindness and schizophrenia, and found this 2013 paper on the hypothesis by Silverstein, Wang and Rochet (Front. Psychol., 03 April 2013, Excerpt from the third paragraph:

    The conclusion that there are no C/E blind people with schizophrenia is based on a small number of studies that involved relatively small samples. Clearly, this argument would be strengthened by larger, population-based studies. This is because, as a simple calculation demonstrates, a case of congenital blindness and schizophrenia would be extremely rare even if there was no protective effect of blindness: if schizophrenia occurs at a rate of 0.72% in the population (McGrath et al., 2008) and congenital blindness occurs at an estimated rate of 0.03% in people born in the 1970s and 1980s (based on Robinson et al., 1987), then the joint probability of a person having both conditions, if the two are independent, would be 0.02% or 2 out of every 10,000.

    I just wanted to point out that they multiplied (0.0072 x 0.0003) and got (0.0002), which is off by two orders of magnitude – it should be (2.16e-6) or 2 out of every million.

    Carrying on with their argument, they did successfully multiply (0.72% x 0.03% x Population of the US) to estimate that there should be about six hundred people in with both congenital blindness and schizophrenia, so maybe their math works out in the end. (The obvious error certainly doesn’t inspire confidence, though.)

    Does anyone here have a sense of what fraction of people with schizophrenia actually seek treatment, and then get documented in a way that they would be noticed in this kind of a study? Not being familiar with the field, I have no idea whether it’s 10% and we should have found 60 people with these traits by now, or 0.01% and we would be lucky to have found one.

  37. Zippy says:

    “congentally” is missing an i, should be “congenitally”.
    “the back fo the chair” should be “the back of the chair”.

  38. Good Burning Plastic says:

    The third image might look meaningless until you realize it’s a cow’s head;

    Nah, I saw the cow straight away. A much better example is this

    • LPSP says:

      Now that is truly special. Genuinely didn’t even notice the brown body, I was so focussed on identifying what I thought was a mouse or small bird.

    • wysinwyg says:

      I saw it right away, but it’s probably because I had read the URL before following the link. The cow took a second even though I’d seen that particular image before.

  39. Abobybous says:

    if only there were AMPA/NMDA stabilizer add-on pills sold with drugs..

  40. Seb says:

    Why the white noise machine?

    • Enkidum says:

      Because if there is no signal, any aberration in any direction becomes normal and hence less worrying? (Best guess, anyways).

    • Bill Stewart says:

      I’ve been dealing with tinnitus the last few years – age-related high-frequency hearing loss means I’m losing most sound above 8kHz, and apparently the brain’s hearing system decides there ought to be something happening up there, so even though it’s not getting any signals from the sensory hardware down below, it fills in a high-pitched noise level anyway. It’s not constant (enough background noise means I don’t notice it), but I spend most of my time in quiet environments so it’s usually there. Blargh.

  41. sconzey says:

    Thanks for this Scott, thought-provoking as usual.

    Your comment about the weighted blankets made me think about swaddling.

    I’ve got two kids and as babies they loved being swaddled when sleeping. One of the things about babies is that they can very easily get ‘over stimulated’ and upset– I guess in your hypothesis above, this would be because the top-down model is bad/sparse, so almost everything is upsetting/surprising/exciting.

    The conventional explanation for why babies like swaddling is that it reminds them of the womb, but I wonder if it might be to do with either autistic-like super stimulus, or to stop babies startling themselves awake (they will do this) because they’ve not yet properly calibrated their proprioception.

    • LPSP says:

      Strikes me as highly plausible. “Reminds of the womb” is in no way exclusive either – assuming babies nearing birth have some perception and cognition, and that the need of little babies are not too different, the instinct to be safely esconced in soft and friendly material would have changed little.

  42. Enkidum says:

    “I think the paper might be claiming that “attention” just means a localized narrowing of confidence intervals in a certain direction; for example, if I pay attention to the feeling of my shirt on my skin, then I can feel every little fold and micromovement. This seems like an important point with a lot of implications.”

    I haven’t read the paper yet, but I work on attention (well, bits of it, anyways), and I think most people in the field would agree that it’s more than that. Let’s say you’re trying to attend to everything with a certain feature, let’s say with a red surface. In this case you’ll get (I’m hand-waving away literally all the details here, and grossly simplifying stuff that I do not fully understand, but I think I’m being more or less accurate):

    – lowered thresholds for red-detectors (so faster reaction times, among other things)
    – higher precision of feed-forward red signals (I think this is your point)
    – greater amplitude of FF red signals
    – attenuated amplitude of FF non-red signals, particularly ones that originate from areas close to the red object
    – etc.

    So you’re not wrong about the localized narrowing of confidence intervals, but there’s more to it than that. (And the specific effects are going to differ widely depending on where in the processing chain you are, and what type of attention you’re talking about, etc.)

    “I notice that I am confused about the relative role of NMDA and AMPA in producing hallucinations and delusions. CFF say that enhanced NMDA signaling results in hallucinations as the brain tries to add excess order to experience and “overfits” the visual data. Fine. So maybe you get a tiny bit of visual noise and think you’re seeing the Devil. But shouldn’t NMDA and top-down processing also be the system that tells you there is a high prior against the Devil being in any particular visual region?”

    Again, a hand-wavy answer, but “top-down processing” isn’t a single system. The types of top-down processing involved in deciding that this bit of random noise is probably the devil are very different from the types that “know” that the devil is not a very likely thing to be there. The first is doing some form of pattern matching with stuff from your memory, the second is more like a logical, sequential, language-based form of reasoning. I think, anyways. (I have absolutely no certainty about this, but it seems right to me anyways.)

  43. ekonomi says:

    Interesting article about bayzian and top-down models predict what we’re going to see, bottom-up models.

  44. Howard Treesong says:

    Many modern t-shirts are tagless, and they’re the only kind I’ll now buy. I’m solidly in the hate-the-tags camp. Dress shirts are much worse: insofar as I can tell, they almost always have multiple tags: one branding tag at the rear center of the collar that is sewn down on all four sides, and then a smaller size tag that flaps around. I cut all of them out; especially after sending shirts out to be laundered, the tags are quite irritating to my skin for some reason.

    I’m otherwise quite a normal person, or so I’m told.

  45. LPSP says:

    10/10 post, a great consolidating update on the work you did as Yvain on LW.

  46. Latetotheparty says:

    Anecdote here:

    I’ve never thought that I was autistic or anything, but ever since I was a kid I have engaged in what I understand now to be similar to “stimming” behaviors when I am overwhelmed–jumping up and down and shaking my hands up and down when I am excited, and sometimes pulling my hair when I am anxious. I still have a tendency to do these things even as an adult, although I can (mostly) control it. Occasionally, though, I’ll still find myself pumping my hands up and down when I am immersed in playing a really good computer game, which thankfully my wife finds endearing.

    I guess I could find the LRF hypothesis plausible. If I am playing a really good computer game, there can be moments where the novelty–the surprise or prediction error–is almost too much to bear, and I have to pause and let out the excitement through flapping my hands. It’s almost as if my brain is telling me, “This is too surprising! Get back to something more familiar!”

    Another anecdote: I have tripped quite a bit (25+ times) on an assortment of psychedelics, and I have had experiences ranging from serene bliss to mile-a-minute word and thought salad.

    In general, I would say that psychedelics actually weaken the top-down processing. The way I have always tried to explain the experience is that it is like looking at things through the eyes of your 10-year-old self, albeit while maintaining all of the conscious knowledge of your adult self at lower doses. (At higher doses, even the conscious knowledge becomes difficult to conceptually access and disentangle).

    For example, at lower doses I can look at a leaf and know that it is “just” a leaf, but there is a distinct feeling that I am not obligated to reduce it down to just “merely” being a leaf. The top-level categorizing is weakened. I am suddenly open to categorizing it in new ways, or not categorizing it at all. At stronger doses, there is a feeling that my priors have been temporarily wiped clean. (This is why I think mental re-programming is actually quite possible on psychedelics–for good AND ill. If your mental priors tell you that other people are untrustworthy, for example, you can have a strong trip with someone trustworthy and permanently alter both your view and feeling about other people being mostly trustworthy by default. Or, if you trip with someone untrustworthy, you could permanently override your old prior that people were trustworthy (painstakingly built up since childhood) with a new prior that people are untrustworthy. Etc. I feel like psychedelics put you in an “imprinting” state similar to that of small ducklings that will “imprint” onto the nearest sympathetic organism. Powerful stuff.

    If someone wanted to start a cult, I can see how psychedelics would be the ultimate tool for gaining devoted followers. That said, psychedelics also give heightened bottom-up sensory awareness, so if the cult leader was consciously misleading his followers, then I think the followers would sense something off, and would feel a revulsion from a kind of uncanny-valley of a person trying to seem trustworthy without really being trustworthy. It would have to take either an expert actor/con-artist or a true believer who believed his own delusions that he really was the messiah.

    In my experience, the best trips are when you do not feel compelled to reduce the heightened bottom-up sensory inputs to top-level schemas at all. If you allow yourself the luxury of merely observing everything without categorizing it or trying to make sense of it, then that is the most enjoyable. You’ll have no idea what’s going on, and you won’t be able to make decisions, but if you were smart enough to put yourself into a situation where you were sure you wouldn’t have to make decisions, and if you hold onto that one thought, the rest of the top-level judgment and decision-making can be dispensed with for the duration of the trip. You can just revel in, “So wonder. Much green. Wow.” And though I am comparing it to doge ironically here, the experience really is pleasurable. It’s like discovering EVERYTHING for the first time again. People often talk about how they wish they could make themselves forget their favorite movie or video game just so that they could rediscover it again and have as much fun as they did the first time around. Tripping is like that, but where life and the huge, wondrous, expansive universe as a whole is the object of attention, rather than some narrow movie or game. It’s pretty sweet. It’s like the Planetmind in SMAC being hooked up to the Planetary Datalinks all at once. Obligatory SMAC link here:

    But if you are in a bad situation where you feel like you suddenly have to make big-picture, strategic decisions, you are going to have to build up new priors on the spot from scratch, and they are probably not going to be accurate ones. Your top-level schema-making is already weakened. Trying to exercise it in that weakened state is like trying to lift weights with a broken arm. You are going to strain it. If you think you NEED an explanation for the swaying of green and brown in your field of view, and your link to the prior of “these are trees swaying in the wind” that you built up since childhood is attenuated due to the weakened top-level schema-making, then you are going to grasp at other priors, and “this is an alien being with antennae trying to communicate with me through sign language” might seem like a perfectly reasonable alternative on this leveled playing field. (I have in fact thought while tripping at times that trees were gesturing to me or other people, for example). This is where the word-salad, thought-disorder trips come from.

    So, in conclusion, with psychedelics I would say that the occasional delusional experiences do not come from heightened top-down process, but rather the opposite. Delusions are the result of the top-down processes being weakened and the poor tripper’s mind feeling compelled to flail about and grasp onto whatever weak sauce schemas the tripper can conjure up at that point. This is also why I feel like tripping is easier if you have a higher IQ. It can, to a certain extent, compensate for the temporary weakness of your top-down processing, if needed.

    Is it not also true that schizophrenics often have general cognitive deficits? Maybe their top-down processing is actually under-active, and they feel compelled to latch onto whatever sparse, feeble explanations are offered up rather than walk around in a haze of not having an explanation at all for anything and everything going on in reality….

    • alaska3636 says:

      +1 for accurate portrayal of the joy of a good trip.

      The “oneness” feeling many people report seems more likely a removal of entrenched categories. As you put it, it allows a temporary reframing of things that may have assumed a mundane presence in your life, like a leaf.

      The key to a good trip is planning a good environment. No big decisions. Safe atmosphere. Trust-worthy people.

      I tripped recently on my bachelor party. I wanted to see if reframing my relationship would give me insight. It didn’t; but, that is sort of affirming in its own way.

  47. SantocultO says:

    (this might be a good time to remember that no congentally blind people ever develop schizophrenia and no one knows why)

    The congenitally blind people can imagine *

  48. John says:

    This all strikes me as perfectly sensible – have a corroborating piece of anecdotal evidence. As an autistic person, some people strike me as “annoyingly neurotypical”, or “more neurotypical than normal”; their behavior can generally be rephrased as “slightly schizophrenic”. This behavior includes ascribing motives to people with high confidence based on what seem to me to be fairly innocuous minor details. They also treat me like an idiot for failing to assemble observations into the same narrative that they do, when their case often seems very tenuous to me. When I try to overcome my autistic symptoms, I tend to imitate such schizophrenic symptoms myself, and thereby overshoot an ideal mental state – in fact, I’m not even sure it’s possible to approach an ideal mental state by canceling out autistic symptoms with schizophrenic symptoms; two errors don’t necessarily cancel each other out even if they’re errors of opposite types.

    • Paul C says:

      You’ve just identified a major problem I have with such areas as modern philosophy, politics, social justice, and popular views on economics. At times, it seems as though a single scrap of information is delusionally extrapolated into something utterly bizarre or hostile by people I otherwise know to be sane, intelligent, and trustworthy.

      Literature criticism/appreciation has this in spades. I remember in high school and college writing papers that I was convinced were a few steps away from utter nonsense, only to be given high praise. The lesson I took home is that in many of those fields, no one is truly concerned that your conclusion is based on weak evidence, merely that your argument is self-consistent.

  49. US says:

    “Autistic people get the same bottom-up data: an eye-twitch here, a weird half-smile there – but it never snaps into recognizable models; it just stays weird uninterpretable clues.”

    I wasn’t sure if I should comment on this because I’m not sure how important this really is and how it relates to the rest of the stuff in the post, but I thought I should note that the implicit assumption that autistics get/have access to the same bottom-up data as non-autistics is in some contexts questionable. Though it may depend to some extent on what is meant by bottom-up data and what denotes ‘access’. At least I think one should be careful about what this means. I tend to deliberately avoid eye-contact when interacting with people unless there’s some important reason not to (and I’m not very good at that stuff even when I assume it’s important), because I dislike it and because it distracts me. So social signals available to other people may not be available to me during a social interaction. When I consider it to be particularly important to understand the verbal content of the social interaction in which I find myself I also tend to deliberately/strategically disregard/avoid focusing on non-verbal communication during the interaction, because it’s difficult for me to handle both verbal and non-verbal signals in real time, so it’s not just eye contact; in some contexts, e.g. when I’m talking about research I’ve recently read or similar, I tend to avoid looking at the people with whom I’m talking because it’s easier for me to recall the material if I’m not distracted by social signals related to body language.

    I’m on the mild end of the autism spectrum, and I’m not sure how common this sort of strategic approach to dealing with social input is, but I could easily imagine it being somewhat common for relatively high-functioning individuals. In some contexts I assume that high-functioning individuals may know about//be aware of potentially important sources of information (eye contact, posture, etc.) but even so they may sometimes choose to disregard them because they’re too cognitively expensive to deal with during real time social interactions, whereas lower-functioning individuals may not even know that this sort of potentially relevant information is something they’re missing out on.

    • US says:

      On a different note, “LRF think that autistic people dislike social interaction because it’s “the greatest uncertainty”; other people are the hardest-to-predict things we encounter.”

      That’s definitely part of it, I’ve stated something along these lines myself in the past during conversations with non-autistics. Drunk people are even worse than your usual neurotypical, incidentally; those people are completely impossible to model, so I tend to stay far away from environments where you might encounter people who are under the influence of alcohol. Another part of it is that most normal people don’t really like autistics, don’t like to interact with them I mean. Autistics get socially rejected a lot by non-autistics. Although they’re not as good at social learning as non-autistics, many of them are not completely useless at this stuff; getting rejected a lot teaches them over time that most people aren’t very nice to them, and so they may tend to avoid people whom they don’t know/already consider to be loyal (usually family). Or they get anxious around people they don’t know, knowing there’s a high likelihood that they may get rejected (anxiety-related problems are common among autistics), leading to sub-optimal social behaviours that increase the probability of rejection and/or being taken advantage of (first, before getting rejected).

  50. Great write up. You might be interested in my research project modeling psychedelics in a Bayesian framework.
    I’ve found some evidence that the serotonin system might be regulating the diffusion of our ‘top down’ predictions.
    Deciding if the bottom up information will try to snap into ‘wide’ categories (trees, animales) or narrow more detailed ones (oak, pine… cat,dog). I believe psychedelics cause the predictions to become defused and over detailed leading to higher prediction error (less bottom up data can snap in to any prediction), this causes compensatory mechanisms that can lead to hallucinations.
    the work hasn’t yet been published but you can see a lecture from a conference and the manuscript on my web site.
    I think many of the phenomenon in the autistic and schizophrenic spectrum don’t take into account over compensation that happens in most biological systems that move away from homeostasis. for instance when you discuss the difference between autism and schizophrenia. if schizophrenia does starts from increased bottom up noise some specific top down mechanisms might go into over drive which is what will cause the paranoid delusions.
    personally, especially with autism, I think many different underlying causes can cause the wide verity of behavioral patterns categorized under the DSM and that bunching all these things together is actually hindering the advancement of this science (we have to wide of a top down prediction 🙂
    I also don’t think social behavior is ‘less predictable’ I just think the brain area’s that deal with this information are evolutionary and developmentally the latest to be activated in the brain. so it’s these area’s that might not develop correctly even if the underlying problem is over sensitivity to low level sensory input.

  51. alaska3636 says:

    I had a close friend have a schizophrenic break in his late-20s. His delusional pattern-matching went back to 2nd grade. He posted on FB semi-recently that he had a solution to the energy problems of Iran. It is unlikely that this is true but it is endearingly specific. I miss the guy. FWIW He (and I) would always have been considered eccentric although not necessarily on the autism-spectrum.

  52. KingOfNothing says:

    Reading this article while being super stoned.
    Stopped at the picture to make sure how I process it. Couldn’t perceive it in any other way than convex no matter how hard I try.
    What kind of bayesian update does that imply on the credibility of this text?

    • The Nybbler says:

      I find that image REALLY resistant to proper perception. If I block off all but the chin, I can see it as concave. But even a bit of the mustache is enough to force it convex. Or the hair, or the forehead. Flipping it over does not help. And this is clean and sober.

      On the other hand, I can see briefly a Necker cube as 2-dimensional, though it gives me a headache to do so.

    • creative username #1138 says:


      Read the article before and made sure to look at the the picture again when I was really stoned. No difference.

  53. edsorow says:

    I’ve become almost completely anhedonic after suffering some illness very similar to CFS and fibromyalgia. I have noticed that colors are not as bright, music sounds more dull, sensations like tags have stopped bothering me, food tastes more bland, and in general none of my senses and emotions are quite as vivid as they used to be. I no longer feel as curious, or feel awe. Even my dreams are not as vivid as they were. I also no longer feel anxiety. This is separate from depression, since I do not exhibit depressive behaviors. There are no large, sudden, and temporary improvements.

    Marijuana improves my mood slightly for 15-30 min, and then I stop feeling emotion altogether, and become withdrawn. Alcohol just makes my head spin, and LSD makes me psychotic. Bupropion eliminates nearly all my symptoms, but does not help much with the anhedonia. Adderall helps a bit, but mainly just gives me insane focus. Exercise helps with fatigue and motivation (without emotion), but I don’t get endorphin rushes.

    Before my illness I was on the opposite end of the spectrum. Every sensation I felt was magnified. Bright lights, loud noises, tags on shirts, and folds in socks were unbearable. This magnification of my senses just seemed to increase into my teen years, which eventually translated to chronic pain from slight sensations. I could get high just by imagining smoking, and Marijuana felt how I’ve heard heroin described. It was pure euphoria and psychedelic effects. Every drug seemed to affect me more than it should back then.

    Unrelated to the main post, but this has affected by ability to form relationships, since I come off as robotic and uncaring and it’s harder to be spontaneous. I also get less out of relationships.

    • JZig says:

      I went through this about 2 years ago, due to a combination of severe back pain, over stress at work, and multiple social failures. I also have a history of being over stimulated in the exact ways you mention. My first time using Marijuana was ridiculously strong. I had a very robotic relationship about a year ago where I knew I SHOULD have been excited to be with her and she was excited to be with ME, but no matter how hard I tried I just couldn’t summon emotion one way or the other. The relationship ended and I barely felt anything. I decided to stop trying to force relationships until I could feel more.

      I’m still dealing with the anhedonia, but it’s started to get better over the last 6 months and although it’s nothing like it used to be, I feel like I am feeling like 25% of what I “should” instead of 0%. It seems to me like my brain is finally starting to loosen up after 2 years of being shut down, but it doesn’t seem like anything in particular triggered my improvement. So, I’m not saying it will get better for you, but know that it’s possible, and brains change over time in confusing ways.

      • edsorow says:

        I’ve come to terms with it. I hope the anhedonia improves, but to be honest it seems like it’s only getting worse. Whatever, there’s more to life than happiness.

    • Bill Stewart says:

      I’ve had a few months of that a while back. I have restless leg problems that were interfering with my sleep, and the newer drugs for it are a dopamine agonist (similar to Parkinson’s drugs, but at a much lower dose.) Their job is to keep your legs from feeling like they have to twitch and keep you awake; the other drugs for the condition (benzos or opiates) deal with it by keeping you from waking up when your legs twitch.

      I didn’t get full-scale mania, and I wasn’t flying off to Vegas and gambling, but I did notice I was having a good time and buying a lot of crap on the Internet and doing more things than I usually got around to. And then I went to the dentist and got my usual dose of nitrous oxide, but instead of leaving me relaxed and sleepy, I was instantly very dizzy, serious nausea, and spent about three days hardly able to balance enough to get out of bed. Apparently nitrous gives you a burst of dopamine that hung around abnormally long because the drugs were blocking reuptake. Yuk.

      So for a few months after dropping that drug, I was fairly seriously anhedonic, and it took a while before I felt like myself again.

  54. Tom Passin says:

    This post talks as if there are two main kinds of processing going on, sensory and the top-down, Bayesian things. I think there must be at least three, two of which are being conflated here. There’s sensory input – which isn’t simple raw data since these “lower” levels do some organizing of their own. There’s some kind of pattern-matching, which is the top-down. And then there’s some kind of sanity checking that goes on. Maybe that’s really the “Bayesian” part.

    Sanity checking is not pattern matching. With three major systems in play instead of two, and with presumably different chemicals affecting each differently, no wonder Scott is finding discrepancies and pieces that don’t quite add up.

    It will be interesting to work out the probable effects of messing up the third system too. There are a lot of mix-and-match possibilities!

  55. Mark Lu says:

    Would this “snapping” process also describe how machine learning algorithms recognize things? Can they be described in bayesian terms? Maybe the brain is more like them?

  56. If the brain does processing decently, it’s approximating correct inference. (And despite Hoffman’s arguments about evolutionary fitness versus reality, we still expect some significant degree of correlation between reality and perception.)

    So… if it’s NOT approximating bayesian reasoning, that would be really surprising. It would be like finding that your brain’s apparatus for adding numbers wasn’t approximating normal arithmetic – sure, it’s not going to be exactly the same, and it’s optimized in various ways using approximations and memorization, but it better correspond somewhat, since it gets the right result most of the time.

  57. Captain Ford says:

    The model of top-down reasoning meeting bottom-up reasoning seems really insightful and really useful. It does describe autism really well, and clearly a number of other things. But it’s also a very simple model, and the brain is really complicated.

    As for why the brain treats two salient events different, probably the whole handshake thing is part of a low-effort system that scans for anomalies, and once an anomaly has been found the brain interprets the anomaly and figures out what to do with it, which isn’t described or predicted by this model. The model only describes the low-level scanning for info, and doesn’t describe any higher-level processes.

    Also the observations that AMPA and NMDA correlate with bottom-up and top-down respectively is a really coarse correlation. It’s like observing levels of ants and termites in a house, and noting that couches fall apart more often when there’s higher levels of termites, but not when there’s higher levels of ants, but without knowing termites eat wood and couches are often — but not always — made of wood.

  58. eb says:

    I’m interested in theories about how this could be mapped to the drug effects state space:

  59. Brian Donohue says:

    Very interesting. Since the top-down system enforces priors, it seems to me that, as we age, it comes to dominate, e.g. “nothing surprises me anymore” etc.

    On this view, maybe marijuana is a means of keeping an open mind about stuff and a greater ability to be surprised.

  60. Paul C says:

    I think this post has explained (at least in part) why I don’t listen to music. For me, it’s frequently overstimulating or distracting, and an utterly horrific experience when it’s loud. What’s more, to appreciate music I’d have to concentrate on listening to pick out the lyrics and patterns, which is a modestly mentally demanding task without being rewarding.

    Growing up, this utterly baffled my teachers and fellow students. On many occasions, we were told to bring in music we liked for projects… And I had politely explain that “No, really, I don’t listen to music at home.”

  61. Amphetamine: "increased visual acuity" vs. "blurred vision" says:

    One of my (rare) daylight amphetamine experiences i compare with “walking through a forest at night with a flashlight on”: Once you can accept to perceive only a spot in a much bigger picture, you see much clearer (within that spot) than without the flashlight. Otherwise “everything” appears blurred, because that terrible spot of light distracts you from seeing the whole picture.

  62. "dopaminergic mania" says:

    I perceive the effect of amphetamines as highly opportunistic: The mania doesn’t target the common sense (“the common mania”) but is very sensitive in projecting suppressed discrepancies onto “otherwise” manic individuals. Thus the “dopaminergic mania” doesn’t attract attention, but even strengthens the common sense/mania by “kicking downwards”.

  63. Some Dude says:

    Ok, so, does this mean that these receptors can be tweaked or optimized by ingesting glutamate-rich foods or supplements?

    Could it be argued that both NMDA and AMPA signalling could be increased resulting an an overall cognitive boost?

    What, if any, are the implications of this sensory processing system for abstract forms of thought like mathematics?

  64. the5chord says:

    so a being a 9/11 truther or an Obama birther is an example of bottom up or top down thinking?

  65. Agmatine says:

    Possibly related data point. I have taken memantine for a few years to help treat my bipolar II (currently 14mg xr but ranging from 7-40mg), which is an NMDA antagonist (which supposedly only blocks abberant amounts of glutamate) and I cannot have even a low grade psychedelic experience at even very high doses of psilocybin (6 grams dried shrooms) or 2cb. Given my concerns about LSD worsening my mental illness I have not tried doses higher than 60mg but even that generated minimal if not zero perceptual changes.

    Ketamine as one might expect does work quite well though I am not particularly interested in doses high enough to k hole. Salvia is works but is too intense for me to use any more frequently than once every few years.

  66. Mark says:

    I’m not convinced of the top-down/prediction/handshake aspect of this model. Seems overly complex.

    I’ve always conceptualized the process of perception as wholly bottom up – a process of converting raw sensory input (pixels, vibrations etc) into increasingly higher levels of abstraction. In the book “On Intelligence”, the author suggests there are ~7 levels of abstraction. Dark pixel -> line -> curved line -> circular edge -> cup -> coffee shop -> breakfast meeting -> i’m a lawyer -> does life have meaning?

    At each jump in abstraction, there will be many symbols (I use this to mean a concept at any level of abstraction) that are candidates. The choice depends upon:
    – How good the match is
    – How “strong” that symbol is (used a lot/used recently) – hence priming, and sticky delusions, seeing a cow when told to think about a cow
    – If that symbol was created (and therefore linked) to the emotional state we are current experiencing, it’s more likely to be chosen – hence devils, bad trips, paranoia

    Surprise is an unprimed good match.

    I think of hallucinogens as increasing the “enthusiasm” of this abstraction process. And I’d guess autism is below-normal enthusiasm, and schizophrenia above-normal.

  67. this makes me think about the documentary I started last night – The Brain That Changes Itself