AbstractExpectations and prior knowledge strongly affect and even shape our visual perception. Specifically, valid expectations speed up perceptual decisions, and determine what we see in a noisy stimulus. Bayesian models have been remarkably successful to capture the behavioral effects of expectation. On the other hand several more mechanistic neural models have also been put forward, which will be referred as “predictive computation models” here. Both Bayesian and predictive computation models treat perception as a probabilistic inference process, and combine prior information and sensory input. Despite the well-established effects of expectation on recognition or decision-making, its effects on low-level visual processing, and the computational mechanisms underlying those effects remain elusive. Here we investigate how expectations affect early visual processing at the threshold level. Specifically, we measured temporal thresholds (shortest duration of presentation to achieve a certain success level) for detecting the spatial location of an intact image, which could be either a house or a face image. Task-irrelevant cues provided prior information, thus forming an expectation, about the category of the upcoming intact image. The validity of the cue was set to 100, 75 and 50% in different experimental sessions. In a separate session the cue was neutral and provided no information about the category of the upcoming intact image. Our behavioral results showed that valid expectations do not reduce temporal thresholds, rather violation of expectation increases the thresholds specifically when the expectation validity is high. Next, we implemented a recursive Bayesian model, in which the prior is first set using the validity of the specific experimental condition, but in subsequent iterations it is updated using the posterior of the previous iteration. Simulations using the model showed that the observed increase of the temporal thresholds in the unexpected trials is not due to a change in the internal parameters of the system (e.g. decision threshold or internal uncertainty). Rather, further processing is required for a successful detection when the expectation and actual input disagree. These results reveal some surprising behavioral effects of expectation at the threshold level, and show that a simple parsimonious computational model can successfully predict those effects.
When expectations are not met: unraveling the computational mechanisms underlying the effect of expectation on perceptual thresholds