The Louder Surprise

The brain predicts. Before visual input reaches conscious processing, cortical circuits generate expectations about what the eyes should see. When reality differs from prediction, the brain signals the mismatch — a prediction error. The dominant theory: these error signals encode the difference between expected and actual input. The brain computes a residual — what the world is minus what the model predicted — and sends that residual up the hierarchy for processing.

Furutachi and Hofer at UCL found no evidence for this mechanism. Using two-photon calcium imaging in mice navigating a virtual reality corridor with unexpected images on the walls, they recorded individual neurons in the primary visual cortex during surprise events.

The brain does not compute a difference signal. It turns up the volume on the neurons that already prefer the unexpected input.

Two neural populations cooperate. VIP-expressing inhibitory interneurons in V1 act as a switchboard — they detect the mismatch condition (expectation violated). Pulvinar neurons in the thalamus provide the amplification — when VIP neurons switch on, pulvinar neurons boost cortical activity; when VIP neurons are off, pulvinar neurons suppress it. The result: neurons tuned to the unexpected stimulus fire more strongly, while neurons tuned to the expected stimulus remain at baseline.

This is not subtraction. It is selective amplification. The error signal is a consequence of boosting, not a computation in its own right. The brain doesn't calculate what's different. It makes what's different louder.

The distinction matters for artificial neural networks that implement prediction error explicitly — computing residuals and propagating them as dedicated signals. The biological system uses a simpler strategy: notice the mismatch, amplify the relevant detectors, let the louder signal do the work.