Prediction, uncertainty and model evidence

The Free Energy Principle

Free energy is a way of measuring how well a model explains sensory data, while also penalising overly flexible or implausible explanations.

1. From prediction error to free energy

Predictive coding focuses on reducing prediction error. The Free Energy Principle generalises this idea: organisms act and perceive in ways that keep their sensory states predictable under their own model of the world.

A simple teaching version is:

$F \approx \text{prediction error} + \text{complexity penalty}$

The first term rewards accurate explanations. The second term prevents the system from explaining everything by making beliefs arbitrarily complicated.

2. Interactive free energy landscape

The dot is the current belief. The target is the sensory input. Moving the belief closer to the target reduces error. But strong prior confidence creates a cost for moving too far from what the model expected.

Current belief -0.80
Free energy 0.00
Interpretation The belief is pulled between sensory evidence and prior expectation.

3. Accuracy and complexity

In variational formulations, free energy is often described as a balance between accuracy and complexity:

$F = \text{complexity} - \text{accuracy}$

Accuracy means the model predicts the data well. Complexity means the posterior beliefs have moved away from the prior. A good explanation should fit the data, but not by inventing an unnecessarily complicated story.

Accuracy

How well does the model predict the observed data?

Complexity

How far did beliefs move from the prior?

Precision

How strongly should errors influence belief updating?

Evidence

How plausible is this model as an explanation of the data?

4. Perception and action

There are two broad ways to reduce free energy. Perception changes beliefs to better explain sensory input. Action changes sensory input so it better matches predictions.

Perception

Update beliefs so predictions better match the world.

beliefs → predictions → lower error

Action

Change the world, or sample the world, so sensations better match predictions.

actions → sensations → lower error