Your brain is not a passive receiver of information. Instead, it actively constructs a model of the world and constantly makes predictions about what sensory input it expects to receive next. Think of it like a scientist forming a hypothesis. The sensory data that follows is the experiment, and the brain checks if the results match its hypothesis.

When there's a mismatch, a "prediction error" occurs. This error signal is the crucial piece of information. The brain doesn't waste resources processing the entire sensory stream in detail. It focuses on the difference between what it predicted and what actually happened. This is far more efficient. If you predicted the floor would be flat and it is, there's not much to learn. But if you step into an unexpected pothole, that's a surprise worth noting.

We can formalize this. Let's call the incoming sensory data $y$. The brain's prediction, which comes from its internal or generative model, is represented by $\mu$. The generative model, often written as $g(\vartheta)$, is the brain's internal simulation of how causes in the world (represented by parameters $\vartheta$) produce sensory effects. The prediction $\mu$ is the expected outcome of this model.

The error, $e$, is the residual signal. It’s what's left over after the brain's best guess has been subtracted from reality. This is what gets passed up the cortical hierarchy for further processing. The goal is always to minimize this error over time, either by updating the model or by acting on the world to make the prediction come true.

This process isn't flat; it's organized in a hierarchy. Think of the brain's cortex as having different levels. Higher levels make abstract predictions, like "I'm seeing a face." This top-down prediction ($\mu$) is sent to a lower level. The lower level compares this prediction to the actual sensory input ($y$) it's receiving, like patterns of light and shadow.

The calculated error signal is then sent back up to the higher level, which uses this information to update its own model. In this hierarchical system, what one level considers "sensory input" is actually the prediction from the level above it. Only the very lowest level of the hierarchy deals with raw sensory data from the outside world.

To make this more precise, the error that gets passed upward is often weighted by its expected precision, or reliability. This precision-weighted prediction error, denoted by $\xi$, is what drives learning. If the sensory input is noisy or unreliable (low precision), the brain shouldn't overreact to the resulting error. Conversely, a high-precision error signal means the model needs a significant update.

When an error occurs, the brain has two main strategies to resolve it, which we can think of as 'EXPL' (Exploration) and 'PRED' (Prediction).

Both strategies work to minimize future prediction error. The brain's choice between them depends on context and the precision of the sensory signals and prior beliefs.

This entire process has a plausible neural implementation. Top-down predictions are thought to be carried by deep pyramidal cells in the cortex. These predictions are sent to superficial pyramidal cells in lower cortical areas. These superficial cells compare the prediction with the bottom-up sensory input and compute the error. This error signal is then passed up to the superficial layers of higher cortical areas to update the model. Deep cells send predictions down; superficial cells send errors up.

This elegant dance of prediction and correction allows the brain to efficiently process a complex world, focusing its resources only on what is new, surprising, and informative.

Ready to test your knowledge of prediction error?