Active Inference: The Predictive Brain

This post summarises my work with Durham University on the topic of Active Inference, a neuroscience-inspired framework that treats the brain as a prediction engine. The slide deck is embedded below and puts software spin on Active Inference, and it is the best place to explore the full visuals and examples.

The Brain as a Prediction Machine

Active Inference begins with a simple but powerful idea: the brain is not a passive receiver of information. It is a proactive model builder that constantly predicts what it expects to sense next.

The goal of every living system is to minimize surprise.
Surprise is the difference between what the brain expects and what the senses actually deliver.
When a prediction is wrong, the brain registers a prediction error and must correct it.

This is similar to a live machine learning model that is continuously updated against a streaming dataset.

Why the Brain Is So Efficient

One of the deck’s first points is how energy-efficient the human brain is:

The brain runs on roughly 20 watts of power, a tiny fraction of daily caloric intake.
Modern AI systems like large language models need megawatts and huge data centers.
Active Inference explains this by showing how the brain avoids recomputing reality from scratch.

Instead, the brain relies on strong priors—internal expectations that act like a memory cache. It only spends energy when a prediction error occurs.

Priors, Prediction Error, and Model Updates

A useful analogy from the slides is "tigers vs. kettles":

When you enter a kitchen, your brain expects a kettle, not a tiger.
If there is a tiger, reality violently clashes with your prior.
That creates a large prediction error.
The brain then updates its internal model using a high-energy response, such as fight-or-flight.

This demonstrates how the brain constantly balances stability against surprise.

Fixing Errors: Perception vs Action

When predictions fail, the brain has two options:

Perception: Update the internal model to match reality.
Action: Change the world so it matches the brain’s prediction.

The slide deck uses a neat software analogy:

Perception is like git pull—fetch the latest state from the world and update your local model.
Action is like git push—force the world to match your internal expectation.

Both mechanisms are part of the same predictive process.

Precision Weighting: Tuning the Brain

Not all prediction errors are equally important. The brain must decide whether a mismatch is meaningful or just noise.

Dopamine and other neuromodulators are thought to control this precision.
Precision acts like a learning rate or attention weight in machine learning.
If precision is high, the brain treats the error as important and updates the model quickly.
If precision is low, the brain treats the error as noise and ignores it.

Healthy vs. Aberrant Models

The deck contrasts two modes of brain inference:

A healthy model balances priors and sensory evidence.
An aberrant model gives too much weight to noisy inputs.

When precision weighting breaks down, the brain can invent bizarre explanations for noise, which resembles the kinds of symptoms seen in schizophrenia.

Everyday Examples

The slides include compelling real-world examples:

The silent heartbeat: You normally do not feel your heartbeat because your brain predicts it perfectly. The prediction error is zero, so the sensation is suppressed.
The placebo effect: Belief can produce real physiological change. A strong prior that a pill will help can drive the body to release its own opioids or dopamine.
Phantom limbs: After amputation, the brain still expects the missing limb. The mismatch between prediction and sensation can be interpreted as pain, and therapies like mirror-box therapy help resolve the error.

Strengths, Challenges, and Competing Theories

Active Inference is praised for:

Unifying perception, learning, and action under one principle
Being biologically plausible for how brain circuits may operate
Encouraging efficient, sample-efficient learning by seeking out informative surprises

At the same time, the model faces real challenges:

It is mathematically complex and can feel difficult to apply.
It risks overfitting because powerful Bayesian models can explain many outcomes.
Scaling it to modern AI systems remains expensive and computationally demanding.

The slides also compare Active Inference to other paradigms like Global Workspace Theory and reinforcement learning.

Summary

Active Inference is a compelling way to think about intelligence as prediction, surprise minimization, and energy efficiency. It is especially useful for understanding how brains can stay stable most of the time while still reacting strongly to important errors.

Want the full narrative? Browse the embedded slide deck below and follow the controls to step through each slide.