Tiesinga Paul, Womelsdorf Thilo
bioRxiv. 2025 Sep 3:2025.09.03.674042. doi: 10.1101/2025.09.03.674042.
Inferring the behavioral relevance of visual features is difficult in multidimensional environments as many features could be important. One solution could involve tracking the experience with multiple features and using attentional control to decide which subset of features to explore and chose. Here, we characterize this attentional control process with a model of parallel belief states and test it with a task requiring the learning and updating of attention to features with varying selection histories and motivational costs. We found that exploring and exploiting features was accounted for by a model that tracks the latent beliefs about the relevance of multiple features in parallel. These parallel belief states accounted for the fast learning of feature-based attention, for perseverative selection history effects for features that were previously relevant, and for enhanced learning performance when the motivational costs of making errors increased. Taken together, these results quantify how multiple parallel belief states guide exploration and exploitation of feature-based attention during learning. We suggest parallel belief states represent attentional priorities that are read out by a competitive attentional control process to explore and exploit those visual objects in multidimensional environments that are believed to be relevant.
During goal-directed behavior attention is allocated to features relevant for the behavioral goal. But in real-world settings with multidimensional objects, it is often unknown which features are maximally relevant and should be attentionally prioritized. We found a solution to this problem by quantifying the hidden beliefs about the relevance of multiple features in parallel. By tracking belief states about feature relevance we found that subjects consider multiple features in parallel during the learning of feature-based attention. These belief states correspond to attentional priorities and explained when attention is biased towards previously relevant but now irrelevant features, and when learning about relevant features is enhanced by motivational incentives. These findings quantify the parallel hidden beliefs that guide attention in complex environments.
在多维环境中推断视觉特征的行为相关性很困难，因为许多特征可能都很重要。一种解决方案可能涉及跟踪对多个特征的体验，并使用注意力控制来决定探索和选择哪些特征子集。在这里，我们用一个并行信念状态模型来描述这个注意力控制过程，并用一个需要学习和更新对具有不同选择历史和动机成本的特征的注意力的任务来测试它。我们发现，探索和利用特征可以由一个并行跟踪多个特征相关性的潜在信念的模型来解释。这些并行信念状态解释了基于特征的注意力的快速学习、先前相关特征的持续性选择历史效应，以及当犯错的动机成本增加时学习性能的提高。综上所述，这些结果量化了多个并行信念状态在学习过程中如何指导基于特征的注意力的探索和利用。我们认为并行信念状态代表了注意力优先级，由竞争性注意力控制过程读出，以探索和利用多维环境中被认为相关的视觉对象。
在目标导向行为中，注意力被分配到与行为目标相关的特征上。但在具有多维对象的现实世界环境中，通常不清楚哪些特征最相关，应该在注意力上被优先考虑。我们通过并行量化关于多个特征相关性的隐藏信念找到了这个问题的解决方案。通过跟踪关于特征相关性的信念状态，我们发现受试者在基于特征的注意力学习过程中并行考虑多个特征。这些信念状态对应于注意力优先级，并解释了何时注意力偏向先前相关但现在不相关的特征，以及何时动机激励增强了对相关特征的学习。这些发现量化了在复杂环境中指导注意力的并行隐藏信念。
