Purchase this article with an account.
Yoshiyuki Sato, Konrad Kording; Learning of likelihoods for Bayesian computations. Journal of Vision 2013;13(9):750. doi: 10.1167/13.9.750.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Introduction: The statistical distributions of the quantity that has to be estimated (called prior), and currently available sensory information about that (called likelihood), are the basis of all Bayesian models. Performance of our sensory systems could also change over time and it also depends on the context such as the properties of the stimulus source. Our brain thus has to constantly estimate the reliability of our sensory systems. Although many studies have shown that priors are learned from experience, little is known about the learning of likelihood uncertainty. Here, we show that human subjects can learn the uncertainty of likelihood in a context dependent way (e.g. red is associated with more uncertainty than green) and correctly generalize this knowledge to the combination with new priors. Methods: We used a sensory-motor task in which subjects estimated a location ("hidden coin") from a splash that it caused. The prior of coin locations was displayed explicitly, and the width of the distribution of the splashes around the coin (likelihood) was cued by the splash color. To show the subjects actually learned the likelihood, we examined whether the learned (color cued) likelihood generalizes to a new prior. Results: We found that subjects relied more on visual splashes when the color indicated that the splashes were more reliable; the subjects learned to switch between different likelihoods depending on the context. When the prior is changed, the subjects adjusted how much they relied on the cue vs the new prior even for the cues associated with the color that was not used during the learning of the new prior. Conclusion: We showed that human subjects could learn the quality of sensory information (likelihood) in a context dependent way and combine the learned likelihood with different priors to make efficient estimations.
Meeting abstract presented at VSS 2013
This PDF is available to Subscribers Only