The Experience Machine

1 - Unboxing The Prediction Machine

Hallucinations, from a predictive processing perspective, are not the same as our usual trippin’ balls perspective. Strong predictions (EG - feeling our phone vibrate when we’re under stress, even if it’s not in our pocket) can create these sensations.

Our brain, as part of it’s normal function, is always creating predictions, and this is what we subjectively experience. We don’t objectively consume sensory input.

Our previous predictions and experiences are always at play, and they’re unavoidable.

We can’t just think ourselves happier, but we do have some flexibility in shaping our interpretation of our sensory input.

The Smart Camera Model of Seeing

Our older models of processing involve a feedforward (top down) model - which assumed our brains took sensory input and gradually got more data out of it. This was unaware of evidence of “downward (and sideways) connectivity.” (Page 9)

Flipping the Flow

Instead of brains being cameras, it appears they may do the opposite, as it’s more energy efficient, maintaining a prediction model. It takes sensory input, checks it against existing miniature models, and generates prediction error signals when it doesn’t line up. Winston’s work in computational neuroscience goes by predictive processing, hierarchical predictive coding, and active inference interchangeably.

Bad Radios and Controlled Hallucinations

Hermann von Hemholtz argued that we generate controlled hallucinations based on previous experiences, conserving energy use in the brain. (The implication here is, this may take part in how we can make predictions about noisy data - like being able to recognize a familiar song mangled by poor FM reception.)

The Frugal Brain

Linear predictive coding is more of a mathematical concept that compresses data by storing only the deltas. (Like in video compression algorithms where frames are compared, and we only need to store the differences.) They suggest our brains work similarly - we just know a lot more.

Human brains seem to benefit from intelligent prediction strategies of just that kind… thanks to the use of multiple “levels of processing.” … simple predictions are nested under less simple, more abstract ones… prediction errors are formed and pushed upward through the system.

Brains reserve bandwidth for prediction errors, as it’s more efficient. Our experiences are hallucinatory.

(This section also talks a bit about prediction hierarchy without explicitly mentioning it.)

The Power of Prediction

A few examples of prediction:

They show the 12 13 14 ABC image, highlighting that unconscious (masked) biases at work.
The Hollow-Face illusion which prevents us from seeing the concave representation.
A Mooney image, ala https://www.behaviouralbydesign.com/post/neuroscience-of-strange-and-beautiful-experiences. Predictions are easily permanently altered.
Sine Wave Speech - an auditory example, similar to above.
Think about Brainstorm or Green Needle while watching this. https://www.youtube.com/watch?v=1okD66RmktA - this easily flips between one or the other.
Suggesting you can hear a familiar song in white noise will often cause you to hallucinate it faintly.
The Ponzo Illusion
The dress that I’m tired of hearing about. (But they do offer the explanation that we will often see the dress in a certain color depending on our masked assumptions on lighting.)
Experientially, if we looked at a tree, and we were not expecting to see a robin at the top of it, we’d likely initially not notice the robin, be surprised when we saw it, and then we’d see it clearly.

2 - Psychiatry and Neurology: Closing the Gap

Much in the same way we do not experience the direct stream of data from or photoreceptors (we see approximate concepts,) we do not directly experience information from our nociceptors. The of thought division between mental and physical, as well as psychological and neurological, is misleading.

While nociceptive pain describes sensory input, and neuropathic pain describes damage to pain reception systems, a third category of pain is nociplastic pain exists - pain without biological pathology.

Beyond Tissue Damage

Expectations both in and out of our awareness may shape our experience of pain. They cite an example of an fMRI study that showed religious images had an effect on how individuals subjectively rated their pain. I uh, think there’s some obvious flaws in this study with attribution.

When shown religious images, religious subjects rated a sharp pain as less intense than atheists shown the same image.

p37 - Andy Clark - The Experience Machine

If you did that to me I’d just straight up be pissed off and my pain would likely spike out of spite.

A more interesting study investigated precision in the prediction of pain processing. Researchers gave one group accurate information about the heat they’d soon be experiencing (low, medium, high.) In the second group, they asked them to expect an unknown level of intensity. When the subjects knew it would be high intensity, they rated it as extra painful, and low intensity experiences were underexperienced. The over-experience and under-experience effects disappeared in the ‘unknown’ group.

In summary of the rest of the section, we do not need to be consciously aware of these predictions, they still have an effect.

Placebo and Nocebo Effects

They highlight the effects of placebo, hypnosis, ‘impure placebos’ (Irving Kirsch would call these active placebos, basically where there’s side effects that make the placebo seem stronger,) and nocebo.

Self-Confirming Cycles of Pain

After conditioning subjects to associate a visual temperature cue to either high or low amounts of heat. This effected subjective ratings of pain, and was also observable under fMRI. The twist was that the heat pad was always set to about 48c.

Functional Disorders

Functional disorders are those with no observable physiological cause. Predictive processing may give theoretical insight to some of these.

Disordered Attention

In predictive processing theory, attention is precision variation. (Since I’m a hypnosis nerd, this might be more evidence that we should stop hammering on subjects to focus harder during a session.) This is adjustable - for example, when searching for a needle in a haystack. Or looking for your keys on a pile of familiar crap on your desk.

Precision estimation is the brain’s way of telling itself where, and by how much, to place its bets.

- P51, The Experience Machine

An explanation for some functional disorders could be “unwilled misallocations of precision as self fulfilling prophecies.” (P51, The Experience Machine)

Hoover’s Sign

A practical troubleshooting tool in functional paralysis. If you ask a normal person to lift one leg, they’ll exert downward force from the heel of the other. If you ask someone physiologically paralyzed to lift their non-responsive leg, the oppositional force will be missing. In a functional case, the oppositional force will still be there as the “good” leg is raised, revealing it’s not physiologically paralyzed.

Expectancy and Its Role in Chronic Pain

Predictive processing may not only be a component of, but an explanation for many types of chronic pain.

Altered Balances in Autism Spectrum Condition

There’s emerging research investigating if autism is related to overweighting incoming sensory input. (Instead of underweighting predictions.) Evidence against the underweighting hypothesis is present in Mooney images - NTs and neurospicy individuals both have no problems with them.

Enhanced Sensory Worlds

ASD individuals may receive more direct sensory input - making it difficult to do things like use interoception to accurately construct that they’re hungry.

The McGurk Effect

Not this guy.

A practical example of the McGurk effect is in ventriloquism. Voices are observed from the puppet, not the speaker. There’s an example of this phenomena here. NT’s are more likely to be influenced by visual input.

Altered Balances in Schizophrenia

Schizophrenia could be the byproduct of a highly-weighted incorrect prediction error signal.

Post-Traumatic Stress Disorder

Researchers presented two mildly angry looking faces, A and B. First, 30% of the time, after face A was presented, they delivered a small shock, with B never having a shock. After a while, they switched the shock to 30% on face B. They measured how much the subjects anticipated the shock by checking their electrodermal response (sweating.) The predictive processing model best fit the responses they saw. Individuals with PTSD overestimated, outside of their awareness, the likelihood of being shocked. Those that overweight this expectation may go on to be more likely to develop PTSD.

So Which Balances Are Best?

Some individuals, who hear voices but did not have a psychosis diagnosis, were more able to accurately predict obfuscated and muddled speech in the background. Paranoia may serve an advantage in dangerous environments.

Action as Self-Fulfilling Prediction

Interestingly, according to predictive processing, predictions of bodily sensation (the feelings of movement) cause action.

The brain predicts how things would look and feel if the action were being successfully performed, and by reducing errors relative to that prediction, the action or movement is brought about.

- Andy Clark, The Experience Machine, P70

This lines up very nicely into hypnosis. This chapter appears to be gold for motor suggestions.

Ideomotor Theory

William James in the mid 19th century came up with the idea that it’s the end representation of the effect (grabbing your cup of coffee off the desk) that causes physical movement, rather than controlling a bunch of micro-movements along the way. Interest in this perspective was renewed by the advent of cognitive psychology, and fits neatly into the predictive processing model. This doesn’t just occur with coffee, but we can see it in Ouija boards and Chevreul’s Pendulum.

Those predictions (of what we would see and feel as the right movements unfold) then act… as motor commands… predictive processing provides… how the “idea” of a successful action can be the very thing that brings that action about.

- Andy Clark, The Experience Machine, P73-P74

Here’s some more reading on the topic:

Seeing Seagulls

They provide a quick story explaining how predictions across cognition and our senses are interconnected. They heard seagulls outside, expected to see them, did not see them from the corner of their vision, got a prediction error, then turned to look, prediction errors being sent in a loop until they turned their head to look at the birds out their window.

One Wiring Diagram to Rule Them All

The bidirectional flow of information in our motor cortex suggests something pretty wild - in order to move, we need to ignore accurate information about our current position to generate the errors to resolve the system.

… When I want to move my hand… forcibly downgrading genuine sensory information associated with the currently immobile state of my arm, while upgrading its own prediction of the proprioceptive signature of the grasping motion… my brain needs to downplay some perfectly accurate information about my own bodily state.

- Andy Clark, The Experience Machine, P77

Precision weighting can be thought of as the predictive processing brand of attention. We reduce the amount of weighting we give to our current position, and increase the weight on the prediction, creating error correcting signals.