Modeling inter-trial dynamics in visual search: developing a hierarchical predictive-coding framework of response decisions in a variety of search tasks

该项目的目的是开发一个原则性的定量框架，描述视觉搜索中的动态加权过程。特别是，我们计划对不同类型的试验间效应进行建模，最重要的是：在单目标搜索场景中，重复与目标定义（特征和）维度的切换的影响。在过去的几十年里，这些效应已经得到了很好的定性描述，并导致了“维度加权”的概念（现在已经成为基于显著性的视觉搜索的一个公认的组成部分，包括J. M. Wolfe的引导搜索模型）。尽管试验间动态有时仍被视为搜索RT变化的“次要”来源，但它实际上占了响应时间（RT）变化的很大一部分，即使是在最简单的，据说纯粹是刺激驱动的单例特征“弹出”搜索中（例如，Found & m<s:1> ller, 1996），并且在单例连接搜索中占更大的比例（例如，Weidner等人，2002）-可以说比自上而下，基于模板的更有影响力)对搜索性能的影响(例如，Kristjánsson et al., 2002)。然而，虽然很重要并且很好地描述了这一点，但我们仍然缺乏加权动力学的原则性计算说明。在这里，我们建议通过将新的贝叶斯型感知决策和权重更新模型与RT分布的生成模型相结合来开发这样的帐户。更详细地说，由于单例特征搜索主要由目标显著性驱动，我们建议使用LATER（“遍历率阈值线性方法”）模型来建模RT分布。在此基础上，我们将开发一个感知模型，用于考虑先验知识对目标检测的影响，（基于特征和维度的）试验间效应，以及维度加权机制。此外，我们将重新审视冗余弹出信号的协同处理的概念，并检查其与试验间效应的关系。在第二个项目阶段，我们计划将我们的感知响应框架推广到一个单一的搜索范式，而不仅仅局限于弹出式搜索，以及寻找权重动态的大脑相关性。

The aim of the project is to develop a principled, quantitative framework describing dynamic weighting processes in visual search. In particular, we plan to model different types of inter-trial effects, foremost: the effects of repetitions vs. switches of the target-defining (features and) dimensions in singleton-target search scenarios. These effects have been well characterized qualitatively over the past decades and led to the notion of 'dimension weighting' (which has now become an accepted component of saliency-based accounts of visual search, including J. M. Wolfe's Guided Search model). Although inter-trial dynamics is sometimes still seen as only a 'minor' source of search RT variation, it actually accounts for a large portion of response time (RT) variability even in the most simple, supposedly purely stimulus-driven singleton feature 'pop-out' search (e.g., Found & Müller, 1996), and an even larger portion in singleton conjunction search (e.g., Weidner et al., 2002) - and is arguably more influential than top-down, template-based) influences on search performance (e.g., Kristjánsson et al., 2002). However, while clearly important and well characterized, we are still lacking a principled, computational account of the dynamics of weighting. Here, we propose to develop such an account by combining a new, Bayesian-type perceptual decision and weight updating model with a generative model of RT distributions. In more detail, as singleton feature search is driven largely by target saliency, we propose to use the LATER ('Linear Approach to Threshold with Ergodic Rate') model for modeling the RT distributions. On top of this, we will develop a perceptual model accounting for the influences of prior knowledge on target detection, (feature- and dimension-based) inter-trial effects, as well as dimension-weighting mechanisms. In addition, we will revisit the notion of coactive processing of redundant pop-out signals and examine its relation to inter-trial effects. In a second project phase, we plan to generalize our perceptual-response framework to a singleton search paradigm not limited to the pop-out search, as well as looking for brain correlates of the weighting dynamics.