Interactive demo · perception as inference

The Windowless Room

Imagine a being inside a sealed room. There are no windows. It never sees the outside world directly.

All it receives are patterns on the walls: knocks, scrapes, vibrations, pressure changes, faint pulses of sound.

The question is not: what is given directly?
The question is: what hidden cause could have produced these sensory effects?

outside world hidden from direct access

🕴️Something?

🌲Something?

🦊Something?

💨Something?

scrape… knock… pause… scrape…

Step 1 · The inverse problem

The cause is hidden

The being does not receive “the world itself.” It receives sensory consequences of the world.

A single pattern — scrape, knock, pause, scrape — could be caused by many different things. The same sensory evidence is compatible with more than one possible world.

Effect

The sound pattern arriving at the wall.

Hidden cause

The thing outside that produced the sound.

Inference

The process of moving from sensory effect to likely cause.

several causes could fit the same sensation

🕴️Burglar
at the wall

🌲Branch
tapping

🦊Animal
moving

💨Wind
and loose wood

Which cause best explains this?

Step 2 · Bayesian inference

Prior × likelihood

Bayesian inference gives a simple way to describe how the being updates its belief.

Posteriorwhat I now believe

∝

Priorwhat I expected

×

Likelihoodfit to the sound

Likelihood asks: how well would this cause explain the sound? Prior asks: how likely was this cause before this sound arrived? Posterior is the updated result.

the same signal is interpreted through a model

Posterior ∝ Prior × Likelihood

Step 3 · Interactive demo

Same sound, different prior

posterior updates live

hidden causes outside the room

🕴️Burglar
at the wall

🌲Branch
tapping

🦊Animal
moving

💨Wind
and loose wood

What is causing this pattern?

Urban night: the being has learned that burglaries are possible here. The same scrape-and-knock pattern is now interpreted against a more suspicious prior model of the world.

Adjust the priors

change the mind before the sound

Notice that the likelihoods stay fixed. The sensory sound is the same. What changes is the prior model of the world, and therefore the posterior interpretation.

Step 4 · What changed?

The sound did not change

In the demo, the likelihoods stayed the same. The sensory pattern was equally unchanged.

What changed was the being’s prior model of the world. In an urban night context, the burglar hypothesis becomes more plausible. In a forest cabin, branches, animals, and wind become more plausible.

This is why perception is not simply a copy of sensory input. It is an informed interpretation of ambiguous evidence.

same evidence · different interpretation

🏙️Urban prior

🌲Forest prior

The signal is fixed. The posterior changes.

Step 5 · Examples from experience

You already know this

Bayesian perception can sound abstract, but the basic idea is familiar from ordinary life.

Waiting for a text

You hear a vibration and think it is your phone. If you are expecting an important message, that interpretation becomes more likely.

Footsteps at night

The same creak may feel harmless in daylight but alarming at 2 a.m. Your prior context changes the posterior interpretation.

Seeing a face in shadow

A vague shape in dim light may become a person, a coat, or a tree depending on what you expect to find there.

Hearing your name

In a noisy room, a sound fragment may seem like your name because that hypothesis matters to you and fits the evidence well enough.

perception in daily life

📱phone
vibration

👣night
footsteps

🌘shape
in shadow

🗣️your
name

The mind must infer what is there.

Conclusion · From sensation to world

Perception is world-inference

The being in the room stands for any perceptual system. It is not given the hidden causes of its sensations directly. It must infer them.

Inverse problem

Sensory effects underdetermine their causes. More than one world could explain the same input.

Bayesian solution

The system combines prior expectations with current sensory evidence to form a posterior interpretation.

This prepares the way for active inference: the being can do more than passively infer. It can act, sample, move, test, and reduce uncertainty.

next: action and markov blankets

What can I do to find out more?