Bayesian modelling of hierarchical prior and its updating mechanisms: contextual modulation by cognitive load and sequential dependence

项目D1的目的是揭示多模态环境下动态上下文校准的机制，并建立一个描述上下文校准中先验更新机制的通用贝叶斯框架。特别是，我们计划关注震级估计中的集中趋势效应。集中趋势效应（也称为范围效应或回归效应），在文献中有很好的记录（例如，Helson, 1963；回顾，见Shi， Church, & Meck, 2013），是指由对所呈现的刺激抽样分布的先验知识产生的偏差。虽然在各种类型的感觉估计中发现了集中趋势效应，但目前还没有就上下文校准实际涉及的统计信息的范围达成共识。这个问题在多感官刺激中变得特别突出，因为有多个可用的统计数据（先验数据）并不总是相互一致的。在这里，我们计划建立一个通用的计算框架来研究大脑在上下文校准中是使用多模态特定的先验还是一个模态先验，以及大脑如何解决先验之间的不一致性，以及如何在量级估计中考虑动作先验。此外，我们将进一步开发试验计算贝叶斯模型（例如，卡尔曼滤波，粒子滤波，分层模型），以实现更好的动态先验更新和上下文校准预测。

The aim of project D1 is to uncover the mechanism underlying dynamic contextual calibration in multimodal environments, and to develop a general Bayesian framework describing the prior updating mechanisms in contextual calibration. In particular, we plan to focus on central tendency effects in magnitude estimation. The central tendency effect (also known as the range or regression effect), which has been well documented in the literature (e.g., Helson, 1963; for a review, see Shi, Church, & Meck, 2013), refers to a bias engendered by prior knowledge of the sampled distribution of stimuli presented. Although central tendency effects have been found in various types of sensory estimation, there is at present no consensus on what ranges of statistical information are actually involved in contextual calibration. This issue becomes particularly prominent for multisensory stimuli, given that multiple statistics (priors) are available which are not always consistent with each other. Here, we plan to establish a general computational framework to examine whether the brain uses multiple modality-specific priors or an amodal prior in contextual calibration, and how the brain resolves inconsistencies among priors, as well as how action priors are taken into consideration in magnitude estimation. In addition, we will further develop trial-wise computational Bayesian models (e.g., Kalman filter, particle filters, hierarchical models) to achieve a better prediction of dynamic prior updating and contextual calibration.