Why Your Brain Guesses at the World
Perception isn't reception. Your brain builds a prediction of what's out there and corrects it only where the evidence insists, which means everything you see and hear and feel is a hypothesis the data haven't yet overturned.
Part 1 left the control model with a hole in it. A thermostat reads a number and takes it at face value. No biological system gets that deal. Your sensors are noisy, they lag, and they measure proxies rather than the thing being regulated.
This is the half of the model that fills the hole. It's also the half psychology currently believes, so most of what follows will be familiar to anyone who's read a computational psychiatry paper in the last decade.
How does the brain decide what you're seeing?
By inference, not by reception. The eye receives a flat pattern of light intensities, and that pattern is compatible with an infinite number of three-dimensional scenes. Something has to choose. The visual system combines the sensory evidence with what it already knows about how the world tends to be arranged, and the choosing happens beneath awareness. What reaches you is the conclusion, delivered as though it were the premise.
Try this now.
Close your left eye. Hold your right thumb out at arm's length and look at it steadily. Move the thumb slowly to the right, fifteen or twenty degrees, keeping your gaze fixed straight ahead. At some point your thumb disappears. It doesn't blur. It's gone. Move it a little further and it comes back.
That's the blind spot, the patch of retina where the optic nerve leaves the eye and there are no photoreceptors at all. Every eye has one, roughly the size of nine full moons laid side by side, and you have never noticed it.
The absence isn't the interesting part. What fills it is.
You don't perceive a black disc, or a gray smudge, or a hole. If your thumb vanishes against striped wallpaper, you see continuous stripes. The visual system never reports missing data. It supplies the most likely answer and hands that answer over with no marking to separate it from anything else you're seeing.
Aside. If you've read Plato's The Allegory of the Cave, this is kind of congruent with that, but enough intellectualizing here.
The congruence is closer than a loose analogy. The prisoners aren't making an error of reasoning. Their inference is correct given the evidence reaching them, and the evidence reaching them is a projection rather than the thing itself.
That's the predictive account of perception in a sentence, arrived at roughly twenty-four centuries early. What the modern version adds is the machinery: the shadows are the sensory input, the prior is what the prisoner already expects a shadow to mean, and precision is how much authority the shadow gets against that expectation. Plato's prisoner freed and turned toward the fire is a prior being violated hard enough to update, which is why it hurts and why he resists it.
Three more, quickly, because the pattern is everywhere once you look for it.
You're at a party, loud enough that you can't follow the conversation two feet away, and somebody across the room says your name, and you hear it perfectly. The signal carrying your name was no louder and no cleaner than the noise that swallowed everything else. What made it audible is that your name is a hypothesis your auditory system holds ready at all times, so a fragment of degraded evidence is enough to trigger it. The corollary comes free: you sometimes hear your name when nobody said it.
You've walked down a staircase in the dark, reached the bottom, and taken one more step that wasn't there. The jolt is disproportionate and unmistakable, and nothing hit you. Your motor system issued a prediction, floor, at this height, in one hundred milliseconds, and the prediction was violated. The lurch you felt was the size of the mismatch.
And then there's the dress. In 2015 a photograph of a striped garment divided the internet into people who saw white and gold and people who saw blue and black, each group finding the other's report literally unbelievable.
Lafer-Sousa, Hermann and Conway measured the split and located its source: the image is genuinely ambiguous about the color of the light falling on the dress, and observers resolve that ambiguity with an implicit assumption about the illumination (Lafer-Sousa et al., 2015). Assume warm indoor light and the brain discounts yellow, leaving blue and black. Assume cool daylight and it discounts blue, leaving white and gold.
Same photons, same retinas, different assumption, different color. Not different opinions about the color. Different color, seen.
Hermann von Helmholtz named this in the 1860s and nobody has improved on his term: unconscious inference.
I'll flag something here, because it runs under the rest of this. I spent years as an engineer before I ever sat with a client, and the finding that took me longest to accept is that people don't reason their way to a position and then develop feelings about it. They decide emotionally and then manufacture a logical-sounding case for what they already chose. A Vulcan couldn't run a real business, because rationalization runs in front of logic in every room I've worked in, including the ones full of very smart people.
What Helmholtz described is that same order of operations one level down, in the wiring. The guess goes first. The evidence gets a vote, not a veto.
What is prediction error?
The part of the signal nobody anticipated. The traffic runs opposite to what most people were taught: higher regions send predictions downward, a running account of what the level below should be receiving. Lower regions compare that against what actually arrived and send only the difference back up. Successful prediction is silence. What ascends is the residual, and the residual is what gets learned from.
Rao and Ballard gave this its influential neural formulation in 1999. They built a hierarchical model of visual cortex where feedback carried predictions and feedforward carried residual error, trained it on natural images, and found it reproduced a set of otherwise puzzling findings (Rao & Ballard, 1999).
The efficiency argument is easy to feel. Most of what strikes your senses in any given second is exactly what struck them the second before. Transmitting only the departures from expectation compresses the traffic enormously, and whatever does get passed along is automatically the part worth attending to.
Clinicians already know a version of this, though probably not under this name.
In 1997, Schultz, Dayan and Montague recorded from midbrain dopamine neurons in monkeys learning to associate a cue with a juice reward, and found a pattern that had confused the field (Schultz et al., 1997). Early in training the neurons fired to the juice. After learning they stopped firing to the juice and fired to the cue instead. And if the cue appeared and the juice didn't arrive, the neurons went silent at exactly the moment the juice was due, a dip below baseline, precisely timed.
What dopamine tracks is the difference between the reward received and the reward expected. An unexpected reward produces a burst. A fully predicted reward produces nothing, because nothing was learned. An expected reward that fails to arrive produces a negative signal.
This matters here for one reason. Most clinicians already accept the reward prediction error account, because it underpins how we talk about reinforcement, addiction, anhedonia and behavioral activation. That account is the same idea applied to value instead of sensation. If you already believe the brain learns about reward by tracking the gap between expectation and outcome, you've already accepted the architecture.
What is precision, and why does it do the clinical work?
Precision is how much the system trusts a signal, assigned separately to what it expected and to what it sensed. What you end up believing is a weighted blend of the two, and precision sets the weights. Trust the prediction more and perception drifts toward expectation. Trust the evidence more and perception gets dragged toward the input. Identical evidence, different weighting, different experience, and neither person is being irrational.
Everything up to here can be stated without much subtlety. This one can't.
You're trying to identify a bird and someone hands you a photograph. If the photograph is sharp, every feather and the exact shade of the throat patch, then whatever you thought the bird was before hardly matters. The evidence is good and it should dominate.
If the photograph is a smeared gray blur taken through a rainy window, it should barely move you at all. You should mostly go on where you're standing, what season it is, and what birds are common there. The blurry photograph still counts as evidence. It counts as weak evidence, and the rational move is to weight it accordingly.
Four terms carry the whole machinery.
Prior. What the system expected before any evidence arrived.
Likelihood. How well the incoming evidence fits a given hypothesis.
Prediction error. The part of the input the prediction failed to cover.
Precision. How much the system trusts a signal, assigned separately to the prediction and to the evidence.
Two things follow immediately, and both should be uncomfortable.
Identical evidence produces different perceptions in different people, with neither of them being irrational, if they assign different precision to it.
And a system can generate a false percept in two entirely different ways: by holding its prior too confidently, or by treating its sensory evidence as too unreliable. Behaviorally these can look the same. That problem is deeper than it first appears, and Part 7 is where it gets its due.
The move that made precision indispensable was Feldman and Friston's proposal that attention just is the assignment of precision to sensory channels (Feldman & Friston, 2010). When you attend to something, on this account, you aren't shining a spotlight on it or moving a limited resource around. You're raising your estimate of how reliable that channel is, which raises the gain on its prediction errors.
Attention becomes a control parameter rather than a faculty. Something that felt like a psychological capacity turns into a setting on a signal-weighting mechanism, and settings can be miscalibrated in ways faculties cannot.
One caution before anyone runs with it. The obvious next question is what implements precision in the brain, and the honest answer is that the mapping is loose. Dopamine, acetylcholine and noradrenaline all get named. Each has many functions across many systems, the assignments differ between authors, and the evidence is largely indirect. When you read that a drug "increases sensory precision," read it as a model-fitting result and not as a measurement.
How does the body fit into this?
The same way the eye does. Your brain receives a continuous stream from inside the body, cardiac and respiratory and gastric and metabolic, and it infers what that stream means rather than reading it off. A racing heart is data. Whether it means fear, excitement, exertion, caffeine or illness is a conclusion, arrived at inferentially and delivered to you as a feeling instead of as a judgment.
There are two ways to reduce a mismatch between what you predicted and what you sensed. You can change the prediction, or you can change the world so the prediction comes true.
The second is action, and treating it that way is the step from predictive coding to active inference. On this account no separate motor system receives commands and executes them. Movement is the fulfillment of a proprioceptive prediction. The brain predicts the pattern of muscle and joint signals that would occur if the arm were already extended, and the reflex arcs resolve the resulting error the only way available to them, by moving the arm until the predicted sensations arrive.
Turn that machinery inward and it becomes clinically unavoidable. Seth's proposal is that emotional experience is interoceptive inference, your best hypothesis about the causes of your internal signals, generated by the same machinery that produces your best hypothesis about the causes of light on the retina (Seth, 2013).
It's WebMD implemented in hardware. One real signal in, the most confident available diagnosis out, and no confidence interval printed anywhere on it.
Barrett and Simmons made the anatomical case, arguing that the visceromotor cortices issue interoceptive predictions against which ascending signals are compared, rather than receiving interoceptive information and reacting to it (Barrett & Simmons, 2015). The direction of traffic is the substantive claim.
Peter Sterling pushed it further: homeostasis is the wrong model for how a body is actually governed (Sterling, 2012). Homeostasis is reactive, meaning a variable deviates, the deviation is detected, and a correction follows. Allostasis is predictive. The system anticipates demand and adjusts before the deviation occurs. Blood pressure rises in the seconds before you stand. Insulin releases at the sight and smell of food, not in response to the glucose. Cortisol climbs in the hour before waking.
The clinical implication isn't decorative. If regulation is predictive, then chronic dysregulation needs no broken sensor and no failed correction. It can consist entirely of a body being prepared, accurately and efficiently, for demands that aren't coming. A system still expecting a world it no longer lives in.
Where does the prediction model run out?
At motivation and at force. Predictive processing explains beautifully how an estimate gets built and says very little about why the organism cares what the answer is, or how hard it acts once it knows. It gives you belief without stakes. The control model had a sensor-sized gap and this one has a goal-sized gap, and they're the same gap viewed from opposite ends.
Which is the whole reason to put them together.
That's Part 3, and it's the only genuinely new thing in this series.
References
Barrett, L. F., & Simmons, W. K. (2015). Interoceptive predictions in the brain. Nature Reviews Neuroscience, 16(7), 419–429. https://doi.org/10.1038/nrn3950
Feldman, H., & Friston, K. J. (2010). Attention, uncertainty, and free-energy. Frontiers in Human Neuroscience, 4, 215. https://doi.org/10.3389/fnhum.2010.00215
Lafer-Sousa, R., Hermann, K. L., & Conway, B. R. (2015). Striking individual differences in color perception uncovered by 'the dress' photograph. Current Biology, 25(13), R545–R546. https://doi.org/10.1016/j.cub.2015.04.053
Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79–87. https://doi.org/10.1038/4580
Schultz, W., Dayan, P., & Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593–1599. https://doi.org/10.1126/science.275.5306.1593
Seth, A. K. (2013). Interoceptive inference, emotion, and the embodied self. Trends in Cognitive Sciences, 17(11), 565–573. https://doi.org/10.1016/j.tics.2013.09.007
Sterling, P. (2012). Allostasis: A model of predictive regulation. Physiology & Behavior, 106(1), 5–15. https://doi.org/10.1016/j.physbeh.2011.06.004
The full documents, free and ungated