Mechanisms for causal and non-causal predictive learning

人类学习的一个关键特征是利用过去来预测未来可能发生的事情。例如，人们依靠过去的经验来预测雷声之后往往会有倾盆大雨。但有不同的方法来了解事件是如何相关的。例如，有些事件是同时发生的（打雷和下雨），而另一些则是因果关系（云层和下雨）。人们还可以了解到长期的关系，例如，下雨会导致更多的花，从而导致更多的昆虫，即使下雨可能不会直接影响昆虫的数量。该项目将研究不同的大脑区域如何支持这些不同类型的学习能力，并将填补理解大脑如何响应经验变化的重要空白。因此，这项研究将大大推进我们对大脑如何支持学习和记忆的理解。由于预测和因果关系是许多高级认知活动的核心，包括推理，语言，决策，因此拟议的研究将产生广泛的科学影响。更广泛地说，这些发现可能会对教育产生重要影响，因为它阐明了我们是如何学习的，也对人工智能产生了重要影响，因为因果推理是人工智能的一个主要前沿领域。通过我们的外联活动，还将对教育和社会产生更多影响。一组对记忆很重要的神经区域显示出与学习相关的变化，这些变化代表了预测关系。然而，目前尚不清楚预测性记忆是如何形成的，或者它们反映了观察到的经验。这个项目将调查哪些“核心预测记忆”区域支持因果或非因果预测关系的学习。实验目标将集中在因果学习的三个主要原则：对混淆的敏感性，时间特异性和结构的表征。如果这些区域根据理论神经科学的经典理论进行学习，计算模型将对这些区域应该如何对证据做出具体的预测。一个成熟的功能磁共振成像测量关系强度将测试这些模型的预测在核心记忆区。第一个目标将测试核心记忆区域是否对混淆敏感，或者反映简单的同现。第二个目标是测试核心记忆区域是否具有时间特异性。这将解决是否以及哪些区域学习时间上精确的因果关系，而不是更适合规划的累积预测关系。第三个目标是测试核心记忆区域是否代表外显结构。总的来说，这些发现将加深对预测记忆区域如何协同工作以支持高阶认知的理解。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

A key feature of human learning is use of the past to predict what is likely in the future. For example, people rely on past experience to predict that thunder is often followed by a downpour. But there are different ways to learn how events are related. For example, some events co-occur (thunder and rain), while others are causally related (cloud cover and rain). People can also learn long-range relationships—for example, that rain will lead to more flowers, which leads to more insects, even though rain may not directly affect insects population. This project will investigate how different brain areas support these different kinds of learning abilities, and will fill important gaps in understanding exactly how the brain changes in response to experience. As such, the proposed research will significantly advance our understanding of how the brain supports learning and memory. Because predictive and causal relationships are at the heart of many high-level cognitive activities including reasoning, language, decision making, the proposed research will have a broad scientific impact. More broadly, the findings may have important implications for education by elucidating how we learn, and for artificial intelligence, in which causal reasoning is a major frontier. Additional impacts on education and society will also be enabled through our outreach activities. A set of neural areas, important for memory, show learning-related changes that represent predictive relations. However, it is unknown how predictive memories form, or what they reflect about observed experience. This project will investigate which “core predictive memory” areas support learning of causal or non-causal predictive relations. Experimental aims will focus on three major principles of causal learning: sensitivity to confounds, temporal specificity, and representation of structure. Computational models will make specific predictions about how these areas should respond to evidence if they learn according to classic theories from theoretical neuroscience. A well-established fMRI measure of relational strength will then test these model predictions in core memory areas. The first aim will test whether core memory areas are sensitive to confounds, or reflect simple co-occurrence. The second aim will test whether core memory areas are temporally specific. This will resolve whether and which areas learn temporally precise, causal relations as opposed to cumulative predictive relations better suited for planning. The third aim will test whether core memory areas represent explicit structure. Overall, these findings will sharpen and deepen understanding of how predictive memory areas work together to support higher order cognition.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.