Paradoxical influences of prediction are resolved across time

Abstract

It is widely thought that our brains use expectations to optimise perception, yet theoretical accounts disagree about the form of this optimisation. Bayesian accounts propose that perception is biased towards expected events, rapidly generating more veridical experiences, whereas cancellation accounts argue that unexpected inputs are perceptually prioritised as they are more informative. A recent opposing process theory reconciles these views by proposing a temporal reversal in prioritisation: perceptual processing is pre-emptively biased towards what is expected, followed by later enhancement of only particularly surprising inputs that are informative for learning and model updating. Here, we tested this account using time-resolved decoding of EEG data, while participants observed avatar action outcomes that were either expected or unexpected based on their own movements. Expectation effects on the neural representations of these action outcomes indeed unfolded over time in such a manner, with higher decoding for task-relevant expected outcomes starting before stimulus onset, followed by a later post-stimulus advantage for unexpected outcomes. These findings support the opposing process theory, demonstrating a reversal in prioritisation that is not predicted by current Bayesian or cancellation accounts. By exerting distinct influences across time, expectations can thus render perception veridical, while still allowing for the accurate perception of particularly unexpected events that are important for updating our beliefs or ongoing courses of action. Significance statementHow we perceive the world is shaped by our expectations. Yet it remains unclear how the brain can use these expectations to make perception both more accurate and more informative, given these demands require opposite influences of expectation on processing - upweighting versus downweighting the expected, respectively. Here, we show that this apparent conflict is resolved across time. The expected is enhanced before a stimulus appears by pre-activating what we expect, whereas unexpected inputs are enhanced later. This temporal reversal allows expectations to support both demands, rapidly generating generally more accurate perceptual experiences, while also remaining sensitive to unexpected information that is critical for learning and adaptive behaviour.