Abstract
Perception can be conceptualized as a process of unconscious inference, in which people combine sensory input with beliefs gleaned from past experience. Understanding the form, function, and neural representation of these beliefs, commonly called priors, is key to understanding how we perceive the world. To characterize priors, researchers can fit Bayesian ideal observer models to response data from perceptual experiments. The success of this strategy, however, is limited by the type and quantity of responses that can reasonably be measured and by the approach used for modeling the prior. Here, we systematically study these issues in simulated observers to understand the strengths and weaknesses in existing approaches, and to optimize future experiments for characterizing priors. We focus on a psychophysical experiment used to characterize priors about visual motion. We generated a set of observers with known priors that were all biased towards slow speeds, but varied in shape. We simulated a two-alternative forced choice task in which the observers determined the faster of two moving stimuli with variable speed and contrast. We then used two well-known approaches to recover an estimate of the prior from the simulated data. The first approach assumes that the prior is Gaussian with zero mean and unknown variance. This approach is computationally efficient, but makes strong assumptions about the shape of the prior, which limit the ability to fit other shapes. A second approach approximates the log of the prior as a piecewise linear function. While this approach is more flexible, we show that it can lead to shape estimates that are systematically biased. For both methods, we determined the requisite number of trials to reliably differentiate two priors with a given level of shape similarity. These results suggest experimental design and analysis improvements that can strengthen our inquiry into perceptual priors.