One of the underlying themes in this course is that Bayesian analysis has substantial advantages over traditional statistical analysis, especially when the problems are complex. Now there's evidence for my intuitive argument that the process of Bayesian inference mimics the way we learn about the world.
Playing a fast-moving tennis ball is a complex task. Tracking the ball is hard and vision provides imperfect information about its bounce location. Based on visual input a player can estimate how likely different bounce locations are (red shading on cover). In theory, it's possible to improve on this estimate using information available on a longer time scale: not all locations are a priori equally probable. During a match there will be a probability distribution, the 'prior', of possible locations (green). Bayesian theory tells us that an optimal estimate of the location (contours) is obtained by combining the prior with the visual estimate. In an abstraction of this tennis task, Körding and Wolpert provide evidence that subjects learn the prior distribution and integrate it with visual input in a way consistent with a bayesian process.
Full text from Nature.
Leave a comment